for peer review - unige€¦ · for peer review 3 introduction the burden of unrelieved pain is a...
TRANSCRIPT
For Peer ReviewPain management decisions in emergency hospitals are
predicted by brain activity during empathy and error monitoring
Journal: British Journal of Anaesthesia
Manuscript ID BJA-2018-01024-LC070.R2
Article Type: Laboratory Investigation
Date Submitted by the Author: 29-Dec-2018
Complete List of Authors: Corradi-Dell'Acqua, Corrado; University of Geneve, Department of PsychologyFoester, Maryline; Lausanne University Hospital, Emergency DepartmentSharvit, Gil; University of Geneve, Department of Fundamental NeuroscienceTrueb, Lionel; Lausanne University Hospital, Emergency DepartmentFoucault, Eliane; Lausanne University Hospital, Emergency DepartmentFournier, Yvan; Hopital Intercantonal De La Broye Site De Payerne, Emergency DepartmentVuilleumier, Patrik; University of Geneve, Department of Fundamental NeuroscienceHugli, Olivier; Lausanne University Hospital, Emergency Department
<a href=https://www.nlm.nih.gov/mesh/MBrowser.html
target=_new>Mesh keywords</a>:Neuroimaging, Decision Making, Pain Management
British Journal of Anaesthesia
For Peer Review
1
Pain management decisions in emergency hospitals are predicted by brain
activity during empathy and error monitoring
C. Corradi-Dell’Acqua1,2*, M. Foester3, G. Sharvit2,4,5, L. Trueb3, E. Foucault3, Y. Fournier6, P.
Vuilleumier2,4,5† & O. Hugli3†
1Theory of Pain Laboratory, Department of Psychology, Faculty of Psychology and Educational
Sciences (FPSE), University of Geneva, Geneva, Switzerland.2Geneva Neuroscience Center, University of Geneva, Geneva, Switzerland.3Emergency Department, University Hospital of Lausanne (UHL), Lausanne, Switzerland.4Laboratory for Behavioural Neurology and Imaging of Cognition, Department of Neuroscience,
University of Geneva, Switzerland.5Swiss Centre for Affective Sciences, University of Geneva, Geneva, Switzerland.6Emergency Department, Hôpital intercantonal de la Broye, Payerne, Switzerland.†These authors contributed equally.
*Correspondence should be addressed to: Corrado Corradi-Dell'Acqua, University of Geneva –
Campus Biotech, Ch. des Mines 9, CH-1211, Geneva, Switzerland. Tel: +41223790958. E-mail:
[email protected]; URL: http://www.unige.ch/fapse/toplab/
Running Title: Brain signatures of pain management decisions
Manuscript Information:Running Title length: 46 characters (including spaces)Summary word count: 247 wordsManuscript word count: 2999Number of Figures: 4Number of Tables: 1Number of References: 37
Page 1 of 26 British Journal of Anaesthesia
123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960
For Peer Review
2
Summary
Objective. Pain undertreatment, or oligoanalgesia, is frequent in the emergency department
(ED), with major medical, ethical, and financial implications. Across different hospitals,
healthcare providers have been reported to differ considerably in the ways in which they
recognize and manage pain, with some prescribing analgesics far less frequently than others.
However, factors that could explain this variability remain poorly understood. Here, we
employed neuroscience approaches for neural signal modelling to investigate whether
individual decisions in the ED could be explained in terms of brain patterns related to empathy,
risk-taking, and error monitoring.
Methods. For fifteen months, we monitored the pain management behaviour of ED nurses at
triage, and subsequently invited them to a neuroimaging study involving three well-established
tasks probing relevant cognitive and affective dimensions. Univariate and multivariate
regressions were used to predict pain management decisions from neural activity during these
tasks.
Results. We found that the brain signal recorded when empathizing with others predicted the
frequency with which nurses documented pain in their patients. In addition, neural activity
sensitive to errors and negative outcomes predicted the frequency with which nurses denied
analgesia by registering potential side effects.
Conclusions. These results highlight the multiple processes underlying pain management, and
suggest that the neural representations of others’ states and one’s errors play a key role in
individual treatment decisions. Neuroscience models of social cognition and decision-making
are a powerful tool to explain clinical behaviour and might be used to guide future educational
programs to improve pain management in ED.
MeSH Keywords
Pain Management; Neuroimaging; Decision Making
Page 2 of 26British Journal of Anaesthesia
123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960
For Peer Review
3
Introduction
The burden of unrelieved pain is a major unresolved public health problem, resulting in human
suffering and economic costs. Unlike other medical conditions, pain is difficult to quantify
objectively, and is mainly assessed using self-reports and indirect information about its intensity
and aetiology, including medical history, previous experience, etc. As such, pain is frequently
undertreated in hospitals (oligoanalgesia)1,2, an issue which is exacerbated by the fact that
healthcare providers vary widely in the willingness to prescribe analgesics, with only a fraction
of this variability explainable by simple demographic characteristics (gender, age or
professional experience)3–7.
In the last years, Emergency Departments (ED) worldwide have introduced
computerized protocols to guide nurses at diagnosing and managing pain. Although these
approaches improved the overall quality of pain management8–10, they did not counteract
oligoanalgesia, as ED nurses still underestimated and undertreated patients’ pain to a variable
degree11–14. This begs for the introduction of new approaches to better understand the
processes underlying individual pain management decisions, which could lead to appropriate
training procedures to reduce practice variation.
In the present study we exploited recent advances in cognitive and affective
neuroscience, which identified brain patterns related to personal affect and decision-making. In
particular, a network involving the insula, cingulate cortex, and postcentral gyrus, was
consistently implicated in empathizing with other people’s pain15,16. In addition, a partially-
overlapping network in the anterior cingulate, anterior insula, and lateral prefrontal cortex was
systematically associated with monitoring errors and negative outcomes from one’s
Page 3 of 26 British Journal of Anaesthesia
123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960
For Peer Review
4
choices17,18. This growing knowledge about brain functions provided an opportunity to
understand the processes underlying individual differences in pain management. In particular,
we hypothesized that brain patterns related to empathy might explain individual differences in
diagnosis, as healthcare providers who are less sensitive to others’ suffering might report less
the pain of their patients. Further, we predicted that brain patterns related to error-processing
might also influence decisions at the bedside, as individuals most concerned about their
performance might refrain from administering analgesics in fear their side effects.
Methods
Ethics Approval
The study was approved by the Ethical Commission of Canton Vaud (CER-VD N°95/13) and
conducted according to the declaration of Helsinki. Each participant signed an informed
consent form.
Nurse-Initiated Analgesia Protocol
This study took advantage of a nurse-initiated analgesia protocol implemented in 2013 in the
ED of the Lausanne University Hospital (Switzerland). The ED receives around 40,000 patients
annually, each of which is initially triaged through the Swiss Emergency Triage Scale19. Each
nurse certified at using the protocol was prompted by an electronic health record (EHR) to
report: (a) whether the patient was in pain (> 0 using a numeric rating scale ranging from 0 [no
pain] to 10 [the worst pain imaginable]); (b) whether there were contraindications to analgesia;
(c) whether the patient wished to receive analgesia; (d) whether an appropriate treatment
(paracetamol, ibuprofen, tramadol) should be selected (Figure 1A). Importantly, as protocol
Page 4 of 26British Journal of Anaesthesia
123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960
For Peer Review
5
data were recorded at triage, the assignment of patients to nurses was based exclusively on
personnel availability, without any preselection in terms of acuity/aetiology. Hence, the nurses’
identity was independent from the cases examined.
Pain Management Measures
We used the EHR to retrieve information about the pain management decisions of each
certified nurse for 15 months following the protocol implementation. Specifically, we focused
on data from eligible patients (> 16 years old, in pain for less than 3 months, without history of
drug/alcohol abuse, and no life-threatening condition) to estimate the following measures (see
Figure 1A for more details):
1. Treatment Application: proportion of decisions to deliver analgesia on triaged patients.
This index was then broken down into two sub-indexes:
2. Documentation Rate: the proportion of pain documentations on triaged patients.
3. Contraindication (CI) Rate: the proportion of CIs to analgesia documented in those
patients who were in pain.
Participants
Nine months after the protocol implementation, all certified nurses were invited to take part to
a survey probing for demographic information, work experience, and the anxiety from
uncertainty scale20. Subsequently, between 16-18 months after the protocol implementation, a
subgroup was invited to take part to a study involving functional Magnetic Resonance Imaging
(fMRI). This subgroup comprehended equal proportion of individuals from each tertile of the
Treatment Application distribution obtained from a preliminary analysis of protocol data (6
Page 5 of 26 British Journal of Anaesthesia
123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960
For Peer Review
6
months from the implementation). This selection ensured that the tested individuals would
represent a broad spectrum of protocol use.
Neuroimaging Intervention
The neuroimaging study involved the following three experimental paradigms (see
Supplements for more details).
1. Empathy for pain task15,21. Nurses saw pictures depicting hands in painful situations
(wounded, pierced by a syringe, etc.), and control stimuli involving hands without any
aversive feature. The task included 30 stimuli per condition, each presented for 2.5 sec
and followed by an inter-stimulus interval ranging between 2.5-4.1 sec. This task lasted
about 15 minutes.
2. Balloon Analog Risk Task (BART)22,23. Nurses had to adjust to risk in a gambling context,
by pressing a key repeatedly to inflate a virtual balloon as much as possible and stop just
before it exploded. If they stopped before the explosion, they received a virtual
monetary gain proportional to the volume of air pumped (win condition); however, they
received nothing if the balloon exploded (loss condition). The task involved 28 game
iterations, each leading to a potential win/loss. Every game comprehended up to 11
inflations, each remaining on the screen until a response was provided, and followed by
an inter-inflation interval ranging between 1.5-2.5 sec. Win/loss feedbacks lasted 2.5 sec
and were followed by an interval ranging between 2-4 sec. The task never exceeded 15
minutes.
3. Social Harm Avoidance Monitoring Experiment (SHAME)24. We implemented an error-
monitoring task involving similar stakes to clinical decision making, where one’s errors
Page 6 of 26British Journal of Anaesthesia
123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960
For Peer Review
7
may cause harm to another person (the patient). The nurse inside the scanner took
turns with a colleague outside (another nurse from the experimental group) in
performing a dot-counting task. Overall, there were 98 trials, organized in 14 blocks (7
per player) of 7 trials each. Every erroneous response had a 50% probability to cause a
painful stimulation to the arm of the nurse outside the scanner, and was signalled with
an ad hoc feedback for 5 sec, followed by an interval ranging between 2-9 sec. The
overall amount of correct/erroneous trials depended on participants’ proficiency in the
counting task, whose difficulty was adjusted on-line to avoid ceiling/floor effects. The
critical condition was when the nurse in the scanner caused pain to the one outside
(one’s painful errors). This was compared with a condition in which the same harmful
outcome was caused by the nurse outside to him/herself (others’ painful errors). The
task lasted 12 minutes.
Data Analysis
In the behavioural survey, we first assessed the dependency between the three pain
management measures through Pearson’s correlation coefficient. Subsequently, we assessed
how each of these three measures was related with age, gender, years of experience and
anxiety for uncertainty. Results are reported as significant under an α = 0.003 (Bonferroni-
corrected for 15 tests). Uncorrected effects (α = 0.05) associated with anxiety for uncertainty
scores are also reported, as one of the aims of the study was to investigate specifically how
error/uncertainty processing might affect different stages of pain management.
For the neuroimaging investigation, we first preprocessed functional data of each nurse
using SPM12 software (http://www.fil.ion.ucl.ac.uk/spm/) to account for head movements,
Page 7 of 26 British Journal of Anaesthesia
123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960
For Peer Review
8
geometric distortions by the magnetic field, and anatomical differences between subjects. The
preprocessed images were then fed to first-level General Linear Models (GLMs) testing, in each
task, for increased activity in the main condition of interest, and for the tailored control (see
previous studies15,21–24 and supplements for details). The activity maps estimated in each
individual GLM were then used for group-level analyses testing whether the condition of
interest in each task: (a) exhibited increased activity with respect to the control; (b) was linearly
modulated by nurses’ professional behaviour. Activations were reported if surviving correction
for multiple comparisons for the whole brain or for regions-of-interest masks. These masks
were obtained by reanalysing, under the same parameters used here, previous datasets
obtained by running the same three paradigms on lay individuals15,23,24 (see Supplements and
Tables S1-3 for more details).
In addition, we used Least Absolute Shrinkage and Selection Operator (LASSO)25–28 and
Random Forest (RF) regression29 to identify distributed patterns of activity that could predict
nurses’ professional behaviour. In particular, this analysis involved: (1) extracting the activity
associated with each event of interest from a priori masks (the same used for the univariate
analysis). (2) Feeding the extracted signal to the two algorithms for multivariate modelling. (3)
Testing the generalizability of the estimated models through cross-validation techniques: i.e.,
assessing whether a model tailored on a portion of subjects could predict the clinical behaviour
of the remaining (independent) subjects. (4) Obtaining an overall mean squared error (MSE) as
measure of prediction proficiency, which was then validated statistically through permutation
techniques (see Supplements).
Page 8 of 26British Journal of Anaesthesia
123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960
For Peer Review
9
Results
70 ED nurses responded to the survey, 33 of which agreed to take part to a subsequent
neuroimaging investigation (see Table 1 for details). Two nurses asked to discontinue the
neuroimaging session prematurely: hence, BART was completed by 32 participants, and SHAME
by 31.
Behavioural survey
When assessing the nurse-led analgesia protocol data, we found a large inter-individual
variability in Treatment Application (Figure 1B). This variability was related to both individual
Documentation Rate and CI rate: nurses that applied analgesia more frequently were more
inclined to document patients’ pain (r = 0.36, p = 0.002), and less likely to report
contraindications (r = -0.54, p < 0.001) (Figure 1C). None of these indexes were associated with
nurses’ age, years of experience (│r│ ≤ 0.17, n.s.) or gender (│t│ ≤ 0.99; except for potentially
larger Documentation Rate in males nurses t(30.31) = 2.15, p = 0.039, uncorrected). Interestingly,
nurses with higher scores on the anxiety from uncertainty scale showed higher CI rates (r =
0.29, p = 0.017 uncorrected; for the other indexes |r| ≤ 0.18, n.s.).
Neural responses to Others’ Pain
Subsequently, we engaged a subgroup of nurses in a fMRI task where they witnessed pictures
of injured hands. This task recruited a brain network classically associated with pain-processing
and empathy15,16,21, involving the posterior insula, postcentral gyrus, and midline cortical areas
(Figure 2A). No activation was observed in the anterior insula and middle cingulate cortex,
which are known to respond to others’ pain in lay individuals, but not in professional healthcare
providers30,31.
Page 9 of 26 British Journal of Anaesthesia
123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960
For Peer Review
10
We then tested whether these neural responses to others’ pain could predict nurses’
clinical behaviour. First, by using a univariate linear regression, we found a significant
relationship between the activity the right postcentral cortex and Documentation Rate, with
stronger neural response to injured hands in those who reported most frequently patients’ pain
in their daily work. We then tested whether clinical behaviour could be predicted from
distributed patterns of brain activity (rather than isolated regions) during this task. For this
purpose, we extracted the neural activity evoked by viewing injured hands from a predefined
network (see Methods), and fed it to two machine learning algorithms (LASSO and RF) to
predict clinical behaviour. Both algorithms revealed that empathy-related activity was a good
predictor of the documentation rate of individual nurses (Figure 2B). No significant effects
(neither univariate nor multivariate) were associated with the other two measures.
Neural responses to Negative Outcomes
We performed similar analyses for brain activity evoked when observing self-caused errors and
negative outcomes. When confronted with monetary losses (vs. gains) in the BART22,23, nurses
exhibited widespread activations in the middle cingulate cortex, anterior insula, and thalamus
(Figure 3A), a network often associated with the detection of errors17,18, and other salient
outcomes32,33. Univariate linear regression showed that the activity of several regions within
this network, including the insula and cingulate areas, were related to the documentation of
contraindications to analgesia. In addition, multivariate regression with LASSO and RF revealed
that distributed patterns of activity related to money loss was a reliable predictor of nurses’ CI
rate (Figure 3B).
Page 10 of 26British Journal of Anaesthesia
123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960
For Peer Review
11
Similarly, when observing harmful consequences of their own (vs. someone else) errors
in the SHAME24, nurses activated the anterior portion of the middle cingulate cortex. Moreover,
regression analysis showed that activity related to one’s painful errors was linearly coupled
with CI rate in both the middle cingulate cortex and left middle frontal gyrus. Thus, as found for
the BART, these areas were more strongly activated in those individuals who were more likely
to spot contraindications to analgesia. Finally, LASSO and RF regression confirmed that activity
patterns in the network activated by harmful errors were a reliable predictor of CI Rate (Figure
4). Data from neither BART nor SHAME were significantly associated with the other two clinical
measures.
Discussion
Healthcare providers appraise and treat pain very differently from one another3–7, resulting in
patients being more or less likely to receive analgesia according to the person who is in charge
of them. The demographic characteristics of healthcare providers explain only partially this
variability3, suggesting that other factors are at play. By using a battery of well-established
questionnaires20 and experimental paradigms from neuroscience15,21–24, we shed new light on
the mechanisms underlying these inter-individual differences. First, the likelihood of reporting
contraindications to analgesia in clinical practice can be explained by personal anxiety towards
uncertain outcomes (from the behavioural survey), as well as differences in brain responses to
negative feedbacks (neuroimaging investigation). Second, the frequency of documenting
patients’ pain can be explained by differences in brain patterns evoked by witnessing others’
injuries. Overall, our study underscores the role played by two main processes which exert
Page 11 of 26 British Journal of Anaesthesia
123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960
For Peer Review
12
opposite, but concurrent influences on the decision leading to the prescription of analgesia in
clinical practice.
Ideally, choices such as documenting a symptom, reporting contraindications or
prescribing treatment should be motivated exclusively by the clinical characteristics of patients.
Hence, no variability should be observed between ED nurses, as long as they all handle a similar
mix of cases, matched in aetiology and severity. Surprisingly however, nurses behave quite
differently from one another, ranging from those who prescribe analgesia to ~5% up to 20% of
patients (Figure 1B; see also3–7). Considering that patients’ assignment was independent of the
nurses’ identity, and that the clinical variables of interest were obtained by collapsing data from
all cases handled by each operator in 15 months (see Methods), it is unlikely that the observed
variability was influenced by the severity of patients examined. Instead, it is more plausible that
each nurse is characterized by a personal disposition/attitude towards pain management.
Previous studies have already categorized healthcare providers according to their attitudes
(more vs. less attentive to case severity5, more vs. less reliant on patients’ self-reports11),
without however shedding light on the processes that might contribute to this categorization.
Our study extends previous findings, not only by providing a working model according to which
pain management is driven by two clear dimensions, but also by associating these processes
with distinct brain networks.
Brain responses evoked by observing others’ pain have been thoroughly investigated in
neuroscience research, pointing to a major role of the insula, middle cingulate cortex, and
postcentral gyrus16. The most popular interpretation of these activations is that they reflect the
engagement of circuits implicated in first-hand nociception, which are then re-enacted
Page 12 of 26British Journal of Anaesthesia
123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960
For Peer Review
13
“empathetically” when pain is not felt on oneself but observed in others15,16. Critically,
however, these regions are not homogeneous in their function, but can be broadly classified
into two functionally-segregated networks, coding different aspects of the painful experience.
In particular, brain patterns in the anterior insula and middle cingulate cortex might not be
pain-specific, but generalize also to other aversive experiences such as arousing pictures15,
disgusting tastes, or monetary losses34. Hence, these regions could serve a domain-general
purpose involved in detecting events of high relevance for one’s survival32, including errors17,18
and risky decisions22,23, with painful or financial consequences for oneself and others33. In
contrast, the posterior insula and postcentral somatosensory cortex appear to process pain in a
more specific fashion, with little generalization to other forms of affect15,35. This might underlie
a sensory-specific component of the painful experience, which is re-enacted when witnessing
also others’ sufferance15,16. In our study, these functionally segregated networks were
associated with independent components of pain management, with the postcentral gyrus
predicting the frequency with which healthcare providers documented pain in patients, and the
middle cingulate cortex predicting the frequency with which they noted potential
contraindications.
Overall, our study offers a comprehensive model of pain management decisions in
which healthcare providers hold at least two distinct representations of their patient’s state.
First, there is the patient’s current pain, which is estimated through evaluation of diagnostic
signs as well as self-reports, but also influenced by doctors and nurses’ empathic skills. Second,
there is the patient’s prospective state, which is estimated by predicting the potential
consequences of analgesia and thus taps into one’s ability to make decisions under uncertainty
Page 13 of 26 British Journal of Anaesthesia
123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960
For Peer Review
14
and to learn from previous errors. Critically, although healthcare providers are deontologically
bound to relieve patients’ current pain with analgesia, they are equally bound to prevent
potential side-effects by withholding analgesia, a conflict which is resolved differently in each
individual, based on specific characteristics of the case, but also personal traits of empathy,
dispositions towards errors/uncertainty, etc. Training techniques already exist to modulate
empathy and compassion36, but also to help individuals reduce anxiety about potential errors37.
These could serve as a basis for future educational programs for doctors and nurses, to
promote a more efficient pain treatment and a more coherent level of care.
In this study, we exploited the rare opportunity to monitor pain management
behaviours of professional healthcare providers for 15 months, and relate them to brain activity
patterns in well-known tasks. The drawback of this approach lies in the difficulty of obtaining
independent cohorts (e.g., for assessing power or replicating effects), as other hospitals usually
do not record the same behavioural indexes. The application of rigorous cross-validation
techniques insured generalizability within the sample tested. However only future
implementations of the same pain management protocol in other EDs will allow extend our
findings to different countries and healthcare systems.
Funding
This study was supported by the Swiss National Science Foundation grant n. PP00O1_157424/1
(C.C.D.). The salary of one author (M.F.) was partially supported by an unrestricted grant from
the UPSA pain foundation and the French Society for Emergency Medicine (SFMU).
Page 14 of 26British Journal of Anaesthesia
123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960
For Peer Review
15
Authors’ contribution
C.C.D.: Study design, Data collection, Data analysis, Interpretation of Results, Manuscript drafting.M.F.: Study design, Subjects recruitment, Manuscript critical revision.G.S.: Data collection, Manuscript critical revision.L.T.: Data Analysis, Manuscript critical revision.E.F.: Subjects recruitment, Manuscript critical revision.Y.F.: Study design, Manuscript critical revision.P.V.: Study design, Interpretation of Results, Manuscript critical revision.O.H.: Study design, Interpretation of Results, Manuscript critical revision.
All authors approved the final version of the manuscript and agree to be accountable for all aspects of the work (thereby ensuring that questions related to the accuracy or integrity of any part of the work are appropriately investigated and resolved).
Declaration of Interest
None Declared.
Acknowledgments
We would like to thank Matthias Zunhammer for his advices in relation to the LASSO
method and Franca Davenport and Kimberly C. Doell for overseeing the quality of the English
text.
Appendix
De-identified data files and scripts for the multivariate analyses are available at Open
Science Framework: https://osf.io/2bved/.
Page 15 of 26 British Journal of Anaesthesia
123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960
For Peer Review
16
References
1. Rupp T, Delaney KA. Inadequate analgesia in emergency medicine. Annals of Emergency Medicine 2004; 43: 494–503
2. Wilson JE, Pendleton JM. Oligoanalgesia in the emergency department. The American Journal of Emergency Medicine 1989; 7: 620–3
3. Albrecht E, Taffe P, Yersin B, Schoettker P, Decosterd I, Hugli O. Undertreatment of acute pain (oligoanalgesia) and medical practice variation in prehospital analgesia of adult trauma patients: a 10 yr retrospective study. Br J Anaesth 2013; 110: 96–106
4. Heins JK, Heins A, Grammas M, Costello M, Huang K, Mishra S. Disparities in Analgesia and Opioid Prescribing Practices for Patients With Musculoskeletal Pain in the Emergency Department. Journal of Emergency Nursing 2006; 32: 219–24
5. Marquié L, Sorum PC, Mullet E. Emergency physicians’ pain judgments: cluster analyses on scenarios of acute abdominal pain. Qual Life Res 2007; 16: 1267–73
6. Green CR, Wheeler JRC, LaPorte F. Clinical decision making in pain management: Contributions of physician and patient characteristics to variations in practice. The Journal of Pain 2003; 4: 29–39
7. Barnett ML, Olenski AR, Jena AB. Opioid-Prescribing Patterns of Emergency Physicians and Risk of Long-Term Use. New England Journal of Medicine 2017; 376: 663–73
8. Kelly A-M. Nurse-managed analgesia for renal colicpain in the emergency department. Aust Health Review 2000; 23: 185–9
9. Muntlin Å, Carlsson M, Säfwenberg U, Gunningberg L. Outcomes of a nurse-initiated intravenous analgesic protocol for abdominal pain in an emergency department: A quasi-experimental study. International Journal of Nursing Studies 2011; 48: 13–23
10. Cabilan CJ, Boyde M. A systematic review of the impact of nurse-initiated medications in the emergency department. Australasian Emergency Nursing Journal 2017; 20: 53–62
11. Vuille M, Foerster M, Foucault E, Hugli O. Pain assessment by emergency nurses at triage in the emergency department: A qualitative study. J Clin Nurs 2018; 27: 669–76
12. Duignan M, Dunn V. Congruence of pain assessment between nurses and emergency department patients: a replication. Int Emerg Nurs 2008; 16: 23–8
13. Modanloo M, Sayed Fatemi N, Bastani F, Peyrovi H, Behnampour N, Hesam M. Comparison of pain assessment by patients and triage nurses. Iranian J Crit Care Nurs 2010; 4: 23–8
Page 16 of 26British Journal of Anaesthesia
123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960
For Peer Review
17
14. Puntillo K, Neighbor M, O’Neil N, Nixon R. Accuracy of emergency nurses in assessment of patients’ pain. Pain Manag Nurs 2003; 4: 171–5
15. Corradi-Dell’Acqua C, Hofstetter C, Vuilleumier P. Felt and Seen Pain Evoke the Same Local Patterns of Cortical Activity in Insular and Cingulate Cortex. J Neurosci 2011; 31: 17996–8006
16. Lamm C, Decety J, Singer T. Meta-analytic evidence for common and distinct neural networks associated with directly experienced pain and empathy for pain. Neuroimage 2011; 54: 2492–502
17. Klein TA, Endrass T, Kathmann N, Neumann J, von Cramon DY, Ullsperger M. Neural correlates of error awareness. Neuroimage 2007; 34: 1774–81
18. Taylor SF, Stern ER, Gehring WJ. Neural systems for error monitoring: recent findings and theoretical perspectives. Neuroscientist 2007; 13: 160–72
19. Rutschmann OT, Hugli OW, Marti C, et al. Reliability of the revised Swiss Emergency Triage Scale: a computer simulation study. Eur J Emerg Med 2018; 25: 264–9
20. Bovier PA, Perneger TV. Stress from uncertainty from graduation to retirement--a population-based study of Swiss physicians. J Gen Intern Med 2007; 22: 632–8
21. Qiao-Tasserit E, Corradi-Dell’Acqua C, Vuilleumier P. The good, the bad, and the suffering. Transient emotional episodes modulate the neural circuits of pain and empathy. Neuropsychologia 2018; 116: 99–116
22. Rao H, Korczykowski M, Pluta J, Hoang A, Detre JA. Neural correlates of voluntary and involuntary risk taking in the human brain: an fMRI Study of the Balloon Analog Risk Task (BART). Neuroimage 2008; 42: 902–10
23. Schonberg T, Fox CR, Mumford JA, Congdon E, Trepel C, Poldrack RA. Decreasing ventromedial prefrontal cortex activity during sequential risk-taking: an FMRI investigation of the balloon analog risk task. Front Neurosci 2012; 6: 80
24. Koban L, Corradi-Dell’Acqua C, Vuilleumier P. Integration of error agency and representation of others’ pain in the anterior insula. J Cogn Neurosci 2013; 25: 258–72
25. Wager TD, Atlas LY, Lindquist MA, Roy M, Woo C-W, Kross E. An fMRI-Based Neurologic Signature of Physical Pain. N Engl J Med 2013; 368: 1388–97
26. Chang LJ, Gianaros PJ, Manuck SB, Krishnan A, Wager TD. A Sensitive and Specific Neural Signature for Picture-Induced Negative Affect. PLoS Biol 2015; 13: e1002180
27. Krishnan A, Woo C-W, Chang LJ, et al. Somatic and vicarious pain are represented by dissociable multivariate brain patterns. eLife 2016; 5: e15166
Page 17 of 26 British Journal of Anaesthesia
123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960
For Peer Review
18
28. Zunhammer M, Geis S, Busch V, Eichhammer P, Greenlee MW. Pain modulation by intranasal oxytocin and emotional picture viewing — a randomized double-blind fMRI study. Scientific Reports 2016; 6: 31606
29. Breiman L. Random Forests. Machine Learning 2001; 45: 5–32
30. Cheng Y, Lin C-P, Liu H-L, et al. Expertise modulates the perception of pain in others. Curr Biol 2007; 17: 1708–13
31. Cheng Y, Chen C, Decety J. How Situational Context Impacts Empathic Responses and Brain Activation Patterns. Front Behav Neurosci 2017; 11
32. Uddin LQ. Salience processing and insular cortical function and dysfunction. Nature Reviews Neuroscience 2015; 16: 55–61
33. Singer T, Critchley HD, Preuschoff K. A common role of insula in feelings, empathy and uncertainty. Trends Cogn Sci 2009; 13: 334–40
34. Corradi-Dell’Acqua C, Tusche A, Vuilleumier P, Singer T. Cross-modal representations of first-hand and vicarious pain, disgust and fairness in insular and cingulate cortex. Nat Commun 2016; 7: 10904
35. Yarkoni T, Poldrack RA, Nichols TE, Van Essen DC, Wager TD. Large-scale automated synthesis of human functional neuroimaging data. Nat Methods 2011; 8: 665–70
36. Klimecki OM, Leiberg S, Ricard M, Singer T. Differential pattern of functional brain plasticity after compassion and empathy training. Soc Cogn Affect Neurosci 2014; 9: 873–9
37. Keith N, Frese M. Effectiveness of error management training: A meta-analysis. Journal of Applied Psychology 2008; 93: 59–69
Page 18 of 26British Journal of Anaesthesia
123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960
For Peer Review
19
Figure Legends
Figure 1. (A) Flowchart subsuming the key steps of the nurse-led protocol implemented in the
Emergency Department. Nurses were expected to follow and document this procedure for each
patient under their care. Data collected for each nurse over 15 months following the protocol
implementation were used to estimate three different scalars indexing their pain management
behaviour (Pain Documentation Rate, CI Rate, and Treatment Application). Each measure was
computed as the percentage among patients who passed a specific protocol step, as noted in
the flowchart. Full details in methods section. (B) Bar-graphs displaying between-nurse
variability in pain management behaviour. Each subplot represents one of the three scalars of
interest, whereas each bar represents one isolated nurse. Nurses’ identity is here coded with a
number ranging from 1 to 70 according to their percentage of Treatment Application value. (C)
Scatter plots describing the linear relation between the three measures. (D) Scatter plots
describing the linear relation between the Anxiety due to Uncertainty score and each of the
three behavioural measures of interest. Each plot shows a linear regression line (with a grey
area describing the 95% confidence interval), plus the Pearson correlation coefficient. The
significance of the correlation is highlighted as follows: ***p < 0.001, **p < 0.01, *p < 0.05.
Figure 2. Empathy for Pain (A) Whole brain map depicting regions implicated in processing
pictures of injured hands (Painful – Control Images). (B) Linear regression of Documentation
Rate. Surface rendering of a human brain highlighting suprathreshold coordinates in which
neural responses to Painful Images explained nurses’ Documentation rate in univariate linear
regression. Three subplots are also displayed. The left-low subplot describes the linear relation
Page 19 of 26 British Journal of Anaesthesia
123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960
For Peer Review
20
between Documentation Rate and the average parameter extracted by the right Postcentral
Gyrus (grey area refers to the 95% confidence interval). The remaining two subplots refer to
data from Multivariate Pattern Analysis (color-coded according to the machine-learning
algorithm used). On top, the overall proficiency of LASSO and RF classifiers for prediction of the
three clinical measures of interest is displayed. White circles refer to mean square error (MSE)
associated with out-of-subject predictions, superimposed with violin-plots of the permutation-
based null distribution of MSE. The right-low subplot describes the linear regression between
nurses’ Documentation rate and the value predicted by each of the two classifiers. PostC:
Postcentral Gyrus. PreC: Precentral Gyrus. SMG: Supramarginal Gyrus. IFG: Inferior Frontal
Gyrus. Ins: Insula. r: Pearson correlation coefficient. ***p < 0.001, *p < 0.05 associated with
standard parametric analysis (for linear regressions) and permutation-based analysis (for
MVPA).
Figure 3. BART (A) Whole brain map depicting regions implicated in Money Loss (Loss – Win).
(B) Linear regression of CI Rate. Surface rendering of a human brain highlighting suprathreshold
coordinates in which neural responses to Money Loss explained nurses’ CI rate in univariate
linear regression. Three subplots are also displayed. The left-low subplot describes the linear
relation between CI Rate and the average parameter extracted by the Middle Cingulate Cortex
(grey area refers to the 95% confidence interval). The remaining two subplots refer to data
from Multivariate Pattern Analysis (color-coded according to the machine-learning algorithm
used). On top, the overall proficiency of LASSO and RF classifiers for prediction of the three
clinical measures of interest. White circles refer to mean square error (MSE) associated with
out-of-subject predictions, superimposed with violin-plots of the permutation-based null
Page 20 of 26British Journal of Anaesthesia
123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960
For Peer Review
21
distribution of MSE. The right-low subplot describes the linear regression between nurses’ CI
rate and the value predicted by each of the two classifiers. MCC: Middle Cingulate Cortex. PreC:
Precentral Gyrus. Ins: Insula. OP: Parietal Operculum. r: Pearson correlation coefficient. ***p <
0.01, **p < 0.01, *p < 0.05 associated with standard parametric analysis (for linear regressions)
and permutation-based analysis (for MVPA).
Figure 4. SHAME (A) Whole brain map depicting regions implicated in painful outcomes of one’s
errors (One’s – Others’ Painful Errors). (B) Linear regression of CI Rate. Surface rendering of a
human brain highlighting suprathreshold coordinates in which neural responses to One’s
Painful errors explained nurses’ CI rate in univariate linear regression. Three subplots are also
displayed. The left-low subplot describes the linear relation between CI Rate and the average
parameter extracted by the anterior Middle Cingulate Cortex (grey area refers to the 95%
confidence interval). The remaining two subplots refer to data from Multivariate Pattern
Analysis (color-coded according to the machine-learning algorithm used). On the top the overall
proficiency of LASSO and RF classifiers for prediction of the three clinical measures of interest.
White circles refer to mean square error (MSE) associated with out-of-subject predictions,
superimposed with violin-plots of the permutation-based null distribution of MSE. The right-low
subplot describes the linear regression between nurses’ CI rate and the value predicted by each
of the two classifiers. aMCC: anterior Middle Cingulate Cortex. MFG: Middle Frontal Gyrus.
***p < 0.001, *p < 0.05 associated with standard parametric analysis (for linear regressions)
and permutation-based analysis (for MVPA).
Page 21 of 26 British Journal of Anaesthesia
123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960
For Peer Review
22
Tables
Table 1
Demographic information. Eligible ED nurses responding to the survey, and subsequently
subdivided into those who took part to the neuroimaging investigation, and those who did not.
Each of the three groups is described in terms of overall size, number of women (including
percentage value to the overall size), and median age, experience in ED and number of triages
per nurse in a time window of 15 months (bracket values refer to inter-quartile range). For each
of measures reported, the subgroup taking part to the neuroimaging investigation discloses
similar values to the group who did not.
SurveyNeuroimaging
Participants
Other
Participants
Population Size 70 33 37
Females 51 (73%) 22 (67%) 29 (78%)
Age [years] 33 [31, 38] 34 [31, 39] 33 [30, 37]
ED Experience [years]
6 [4, 9] 9 [4, 13] 6 [4, 8]
Triages per nurse 452 [273, 694] 480 [405, 694] 445 [210, 692]
Page 22 of 26British Journal of Anaesthesia
123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960
For Peer Review
(A) Flowchart subsuming the key steps of the nurse-led protocol implemented in the Emergency Department. Nurses were expected to follow and document this procedure for each patient under their care. Data collected for each nurse over 15 months following the protocol implementation were used to estimate three different scalars indexing their pain management behaviour (Pain Documentation Rate, CI Rate, and
Treatment Application). Each measure was computed as the percentage among patients who passed a specific protocol step, as noted in the flowchart. Full details in methods section. (B) Bar-graphs displaying between-nurse variability in pain management behaviour. Each subplot represents one of the three scalars of interest, whereas each bar represents one isolated nurse. Nurses’ identity is here coded with a number
ranging from 1 to 70 according to their percentage of Treatment Application value. (C) Scatter plots describing the linear relation between the three measures. (D) Scatter plots describing the linear relation
between the Anxiety due to Uncertainty score and each of the three behavioural measures of interest. Each plot shows a linear regression line (with a grey area describing the 95% confidence interval), plus the
Pearson correlation coefficient. The significance of the correlation is highlighted as follows: ***p < 0.001, **p < 0.01, *p < 0.05.
Page 23 of 26 British Journal of Anaesthesia
123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960
For Peer Review
Empathy for Pain (A) Whole brain map depicting regions implicated in processing pictures of injured hands (Painful – Control Images). (B) Linear regression of Documentation Rate. Surface rendering of a human
brain highlighting suprathreshold coordinates in which neural responses to Painful Images explained nurses’ Documentation rate in univariate linear regression. Three subplots are also displayed. The left-low subplot
describes the linear relation between Documentation Rate and the average parameter extracted by the right Postcentral Gyrus (grey area refers to the 95% confidence interval). The remaining two subplots refer to
data from Multivariate Pattern Analysis (color-coded according to the machine-learning algorithm used). On top, the overall proficiency of LASSO and RF classifiers for prediction of the three clinical measures of interest is displayed. White circles refer to mean square error (MSE) associated with out-of-subject
predictions, superimposed with violin-plots of the permutation-based null distribution of MSE. The right-low subplot describes the linear regression between nurses’ Documentation rate and the value predicted by each
of the two classifiers. PostC: Postcentral Gyrus. PreC: Precentral Gyrus. SMG: Supramarginal Gyrus. IFG: Inferior Frontal Gyrus. Ins: Insula. r: Pearson correlation coefficient. ***p < 0.001, *p < 0.05 associated
with standard parametric analysis (for linear regressions) and permutation-based analysis (for MVPA).
Page 25 of 26 British Journal of Anaesthesia
123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960
For Peer Review
BART (A) Whole brain map depicting regions implicated in Money Loss (Loss – Win). (B) Linear regression of CI Rate. Surface rendering of a human brain highlighting suprathreshold coordinates in which neural
responses to Money Loss explained nurses’ CI rate in univariate linear regression. Three subplots are also displayed. The left-low subplot describes the linear relation between CI Rate and the average parameter
extracted by the Middle Cingulate Cortex (grey area refers to the 95% confidence interval). The remaining two subplots refer to data from Multivariate Pattern Analysis (color-coded according to the machine-learning
algorithm used). On top, the overall proficiency of LASSO and RF classifiers for prediction of the three clinical measures of interest. White circles refer to mean square error (MSE) associated with out-of-subject predictions, superimposed with violin-plots of the permutation-based null distribution of MSE. The right-low subplot describes the linear regression between nurses’ CI rate and the value predicted by each of the two classifiers. MCC: Middle Cingulate Cortex. PreC: Precentral Gyrus. Ins: Insula. OP: Parietal Operculum. r: Pearson correlation coefficient. ***p < 0.01, **p < 0.01, *p < 0.05 associated with standard parametric
analysis (for linear regressions) and permutation-based analysis (for MVPA).
Page 26 of 26British Journal of Anaesthesia
123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960
For Peer Review
Figure 4. SHAME (A) Whole brain map depicting regions implicated in painful outcomes of one’s errors (One’s – Others’ Painful Errors). (B) Linear regression of CI Rate. Surface rendering of a human brain
highlighting suprathreshold coordinates in which neural responses to One’s Painful errors explained nurses’ CI rate in univariate linear regression. Three subplots are also displayed. The left-low subplot describes the
linear relation between CI Rate and the average parameter extracted by the anterior Middle Cingulate Cortex (grey area refers to the 95% confidence interval). The remaining two subplots refer to data from
Multivariate Pattern Analysis (color-coded according to the machine-learning algorithm used). On the top the overall proficiency of LASSO and RF classifiers for prediction of the three clinical measures of interest. White
circles refer to mean square error (MSE) associated with out-of-subject predictions, superimposed with violin-plots of the permutation-based null distribution of MSE. The right-low subplot describes the linear
regression between nurses’ CI rate and the value predicted by each of the two classifiers. aMCC: anterior Middle Cingulate Cortex. MFG: Middle Frontal Gyrus. ***p < 0.001, *p < 0.05 associated with standard
parametric analysis (for linear regressions) and permutation-based analysis (for MVPA).
Page 27 of 26 British Journal of Anaesthesia
123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960
1
Supplementary Information
for the manuscript “Pain management decisions in emergency hospitals are predicted by brain
activity during empathy and error monitoring” by C. Corradi-Dell’Acqua, M. Foester, G. Sharvit,
L. Trueb, E. Foucault, Y. Fournier, P. Vuilleumier & O. Hugli
2
Supplementary Methods
Detailed description of the neuroimaging tasks
ED nurses were engaged in a neuroimaging study involving three experimental paradigms
testing empathy, risk taking, and error monitoring. The experiment took place in the Brain and
Behaviour Laboratory (http://bbl.unige.ch/) of the University of Geneva. All participants met
MRI safety requirements (no metallic objects in the body, no familial history of epilepsy, etc.),
and were placed supine in the scanner with the head fixated by firm foam pads. Stimuli were
presented on an LCD projector using either E-Prime 2.0 (Psychology Software Tools, Inc.) or
Cogent 2000 (Wellcome Dept., London, UK), and were observed through a mirror mounted on
the MRI headcoil. Key-presses were recorded on an MRI-compatible bimanual response button
box. The paradigms employed were the following.
Empathy for pain. In this task, 120 colour pictures were presented depicting hands in
either painful or non-painful situations. These pictures were sorted in four categories of 30
images each. Painful images described hands in pain, as visible by wounds/marks on the skin
and by the display of an object (scalpel, syringe, etc.) acting on the skin surface. Control images
were neutral stimuli matched with the previous category for hand laterality (right/left),
orientation, and associated visual content (presence of objects), but purged from any
painful/arousing feature. Arousing (and ArousingControl) images described hands in
emotionally aversive (and matched neutral controls), but painless situations (hands holding
knifes/guns, hands with handcuffs). Each of these 120 stimuli was presented for 2500 ms,
followed by an inter-trial interval that ranged from 2500 to 4100 ms (mean = 3300 ms) with
3
incremental steps of 320 ms. Participants were asked to perform a handedness task, i.e., press
one key if the stimulus depicted a right hand but another key if the stimulus was a left hand.
The 120 stimuli were presented in a randomized order together with 30 null-events, in which an
empty screen replaced the stimuli. This task was built using E-Prime 2.0 (Psychology Software
Tools, Inc.) and lasted about 15 minutes.
Balloon Analogue Risk Task (BART). Nurses had to press a key repeatedly in order to
inflate a virtual balloon as much as possible, and stop just before it exploded. If they stopped
before explosion, they received a monetary gain proportional to the volume of air pumped;
however they got nothing if the balloon exploded. Each trial started with an empty balloon
placed on a tip of an inflator. The balloon could then be inflated for a maximum of 11 times
through key-press, each of which was associated with an increasing probability of explosion
(from 0% to 100%), but also an increased monetary reward (from 0.1 to 4.3 CHF). In this
context, participants’ choice to inflate the balloon led to two possible feedbacks: a larger
balloon together with the graphical display of the current monetary gain (e.g., “+ 0.6 CHF”), or a
“negative” feedback in which a picture of the balloon explosion was displayed together with the
text “you have lost”. At each step, participants were free to discontinue the inflation, which led
to a “positive” feedback (“you have won”), and the money amount gained during this trial was
added to their overall earning.
The experimental session comprised 28 independent trials, each separated by an inter-
trial interval ranging between 2000 and 4000 ms. Within each trial, different inflations were
separated by an interval ranging between 1500 and 2500 ms. During that time, the inflator was
coloured in red, to signal participants to withhold any response until it got green. Win/loss
4
feedbacks lasted 2000 ms. Once every six trials, participants were displayed the question “to
which extent this task makes you feel anxious?” together with a visual analog scale that ranged
from “not anxious at all” to “extremely anxious”. Participants had 5 seconds to slide a marker
on the bar until it reached the position corresponding to their judgment. The overall
experimental session never exceeded 15 minutes. This task was built using Cogent 2000
(Wellcome Dept., London, UK).
Social Harm Avoidance Monitoring Experiment (SHAME). The nurse inside the MRI
scanner took turns with a colleague outside the scanner in performing and observing a dot-
counting task. The colleague was one of the other nurses of the experimental group, who was
matched with the participant based on availability (with no further constraint in terms of
seniority or interpersonal closeness). The experiment was organized in 14 blocks, 7 in which the
subject in the scanner performed the task (whilst the subject outside the scanner observed the
same display), alternating with 7 in which the task was performed by the subject outside. Each
block comprised 5 trials, each starting with the simultaneous presentation of two clusters of
white dots on a grey background separated by a vertical line (duration 500 ms, with inter-
stimulus interval ranging from 1.5–3.5 s). The participants in charge of playing indicated which
side contained the largest amount of dots. The trial difficulty was adjusted on-line throughout
the whole experiment (at participant unawareness) to ensure a comparable amount of correct
and incorrect trials for each participant1.
Critically, each response was followed by a thermal stimulation given to the arm of the
participant outside the scanner (3 s of raise time, 2 s of plateau, 3 s for returning to baseline).
Correct responses were always followed by a painless temperature (38°C), whereas incorrect
5
responses had 50% of probability to be associated with a painless or painful temperature (set
by subject-specific pain threshold; average 48°C ± 1.45; see below for details). The thermal
stimulus was associated with a visual feedback (5 s) informing about the performance in the
task and the painfulness of the temperature, and was followed by inter-trial-interval of variable
duration (from 2-9s). The thermal stimulation was delivered through a MSA Thermotest
thermode (SOMEDIC Sales AB, Sweden). This task was built using Cogent 2000 and lasted about
12 minutes.
Pain Thresholding Procedure
In line with previous studies, individual temperatures were determined through a double
random staircase (DRS) algorithm2 3. Our DRS procedure selected a given temperature on each
experimental trial according to the previous response of the participant in a pain
unpleasantness rating scale. Trials rated as more unpleasant than the given cut-off
(corresponding to 8 out of 10 on a visual analogic scale) led to a subsequent lower temperature
in the next trial; whereas trials rated as less unpleasant than the given cut-off led to a
subsequent higher temperature. This resulted in a sequence of temperatures that rapidly
ascended towards, and subsequently converged around, a subjective pain unpleasantness
threshold, which was in turn calculated as the average value of the first four temperatures
leading to a direction change in the sequence. In order to avoid participants anticipating a
systematic relationship between their rating and the subsequent temperature, two
independent staircases were presented randomly. Initial thermal stimulations for the two
staircases were 41°C and 43°C. Within each staircase, stimulus temperatures increased or
6
decreased with steps of 3°C, while smaller changes (1°C) occurred following direction flips in
the sequence. None of our subjects was stimulated at temperature larger than 52°C.
Post-Scanning quantitative debrief session
The nurses taking part to the neuroimaging study were subjected to a post-scanning debriefing
session of about 50 minutes, which comprehended standardized tests as well as a set of
custom-based items. In particular, participants underwent the inclusion of other in the self (IOS)
scale4 to assess to which extent the felt close with the “colleague” engaged in the SHAME
outside the MRI scanner. Furthermore, participants also rated the degree to which they felt
particular emotions (pain, fear, shame, guilt, sadness, and anger) when they were engaged in
the SHAME in the MRI scanner (but not when they were performing the task as confederates
outside the scanner). All ratings were carried out on a Liker scale ranging from 1 to 5, with the
exception of the rating of pain which was carried out on a verbal numeric rating scale ranging
from 1 to 10. Participants were asked to rate their subjective experience associated with any
kind of error event: i.e., they did a mistake or when they observed their colleague making a
mistake, either with a painful or painless outcome. See Koban et al.1 for more details about this
rating session.
Finally, participants were asked to rate each of the 120 stimuli used in the Empathy for
Pain paradigms in terms of familiarity (“how much is the content described in this picture
familiar to you?”), emotional intensity (“how intense is the emotion triggered by this image?”),
emotional valence (“does this image elicit positive or negative emotions?”), and pain (“how
intense is the pain felt by the hand depicted on this image?”). The rating session was carried
7
using E-Prime 2.0, and divided in four blocks, one for each question, during which all 120 stimuli
were rated on a Likert scale rating from 1 to 10 (with the exception of the emotional valence
rating, in which a Likert scale ranging from -4 to +4 was used). To avoid habituation biases due
to the presentation of the same stimuli four times, the order of the blocks and order of the
stimuli within each block was randomized across participants. See Corradi-Dell’Acqua et al.5 for
more details about the subjective rating session.
Imaging processing
Data Acquisition. Functional images were acquired using a 3T whole-body scanner (Trio
TIM, Siemens) with a 32-channel head coil. We used a multiplex sequence6, with TR = 650 ms,
TE = 30 ms, flip angle = 50°, 36 interleaved slices, 64 x 64 in-slice resolution, 3 x 3 x 3 mm voxel
size, and 3.9 mm slice spacing. The multiband accelerator factor was 4, and parallel acquisition
techniques (PAT) was not used. A fieldmap was also estimated through the acquisition of 2
functional images with a different echo times (short TE = 5.19 ms; long TE = 7.65). Finally, a
structural image was acquired using a T1 weighted 3D sequence (MPRAGE, TR = 1900 ms, TI =
900 ms, TE = 2.27 ms, flip angle = 9°, PAT factor = 2, 192 sagittal slices, 1 x 1 x 1 mm voxel size,
256 x 256 in-slice resolution.
Preprocessing. Statistical analysis was performed using the SPM12 software
(http://www.fil.ion.ucl.ac.uk/spm/). For each subject, all functional images were realigned and
unwrapped using a field map image, to account for geometric distortions due to magnetic field
inhomogeneity. Subsequently the functional images were normalized to a template based on
152 brains from the Montreal Neurological Institute (MNI) with a voxel-size resolution of 3 x 3 x
8
3 mm, using a deformation field estimated on a coregistered structural image. Finally, the
normalized images were smoothed by convolution with an 8 mm full-width at half-maximum
Gaussian kernel.
General Linear Model. Preprocessed images from each task were analysed using the
General Linear Model (GLM) framework implemented in SPM, consistently with previous
studies using the same paradigms1 5 7 8. For the Empathy for Pain task, trial time onsets from
each of the four conditions were modelled with a delta function. Additionally, for each
condition we also included an additional vector in which participants Response Times were
modulated parametrically5. For the BART, we modelled with a delta function all inflation events
(in which participants were prompted a decision), with the probability of explosion fed as
additional parametric regressor. Furthermore, we also modelled positive (win) and negative
(loose) feedbacks as separate regressors7 8. For the SHAME we modelled, separately for each
player, all trials in which participants were prompted with a judgment, as well as all kinds of
feedback (correct, incorrect painless, incorrect painful) with separate delta function1.
For all tasks, we accounted for putative habituation effects in neural responses of each
condition by using the time-modulation option implemented in SPM, which creates a regressor
in which the block/trial order is modulated parametrically. Furthermore, each regressor was
convolved with a canonical hemodynamic response function and associated with its first order
temporal derivative. To account for movement-related variance, we included six differential
movement parameters as covariates of no interest. Low-frequency signal drifts were filtered
using a cutoff period of 128 sec. Serial correlations in the neural signal were accounted through
exponential covariance structures, as implemented in the ‘FAST’ option of SPM12. Global
9
scaling was applied, with each fMRI value rescaled to a percentage value of the average whole-
brain signal for that scan.
Functional contrasts, testing differential parameter estimates images associated with
one experimental condition vs. the other were then fed in a second level, one-sample t-test
using random-effect analysis. Similarly, parameter estimates of conditions of interest were fed
to univariate linear regressions, using one of the three clinical measures of interest as
predictors. Effects were considered significant if exceeded p < 0.05, family-wise correction for
multiple comparisons at the cluster level (with an underlying height threshold of p < 0.001,
uncorrected). In addition, we report also effects surviving p < 0.05 small volume corrected for
masks of interests, defined through by previous studies in which independent lay populations
were engaged in the same studies implemented here1 5 8.
Region of interest masks. For each task, we identified an inclusive mask,
comprehending only those coordinates of theoretical interest, as obtained by reanalysing data
from independent researches employing the same paradigms1 5 8, under similar preprocessing
and modelling settings of the current study. The only exceptions were related to those datasets
in which a field map image was not available, for which no unwrapping was applied during the
preprocessing stage. Consequently, in these cases, the deformation field for the normalization
was estimated directly from the functional images (instead from a coregistered structural
volume), to minimize the impact of geometric distortions related the magnetic field
inhomogeneity9. Furthermore, as all these previous datasets were acquired with long repetition
time (> 2 sec), serial autocorrelations were accounted with standard first-order autoregressive
AR(1) model (as opposed to the FAST option for rapid sequences).
10
For the Empathy for Pain task, we reanalysed the data from Corradi-Dell’Acqua et al.5.
As this previous study employed larger database of pictures than ours, we considered as
conditions of interest those parameters estimated on the same sub-selection of images which
were used for the present study (the remaining images were modelled as separate conditions
of no interest), which were then used to identify brain regions associated with the contrast
Painful – Control (Table S1). For the BART, we reanalysed the study of Schonberg et al.8 (freely
available at https://openneuro.org/datasets/ds000001/). This study included an additional
control condition characterized by the inflation of balloon without risk of money loss8, which
was modelled in the first level as separate regressor of no interest. In keeping with results
reported by the original study8, we took into consideration regions implicated in the contrast
monetary loss – implicit baseline (in this specific dataset, no differential effect between
monetary loss – win were reported; see Table S2). Finally, for the SHAME, we reanalysed the
data from Koban et al.1 and selected features implicated in one’s erroneous performance with
painful outcome, compared with others’ erroneous trials with painful outcome (One’s – Others’
painful errors; see Table S3). In all these mask, we followed previous studies implementing
similar multivariate regression on whole brain data10 11, and excluded the coordinates in
occipital cortex that may be driven by distinctive visual features rather than on the information
of interest.
Multivariate Regression
We used Least Absolute Shrinkage and Selection Operator (LASSO)10–13 and Random Forest (RF)
regression14 to identify distributed patterns of activity across brain that could be predictive of
nurses’ professional behaviour in Emergency Department.
11
Feature Selection. For each task, we identified an inclusive mask, comprehending only
those coordinates of theoretical interest, as obtained from independent datasets in which lay
participants were engaged in similar paradigms than those used here1 5 8 (see above).
LASSO. For each of these independently-defined masks, we extracted the neural
activation associated with corresponding tasks in the present study. The resulting data matrixes
(e.g., for the Empathy for Pain Task, 33 nurses x 1077 voxels) were then fed into a LASSO
routine (lasso function from Matlab R2013b) to identify which components were jointly
predictive of nurses’ behaviour in their clinical practice (as recorded during the preceding 15
months). To optimize the modelling, but at the same time insure its generalizability to new
data, the LASSO regression was conducted in two nested 10-fold cross-validation loops. The
first (inner) was aimed at optimizing regularization hyper-parameter λ. The second (outer) was
aimed at predicting professional behaviour of a portions of nurse by a model optimized (also in
its hyper-parameters) on out-of-sample nurses.
Random Forest Regression. The same data matrixes were also fed to the regression
routines implemented in the Matlab-based RF toolbox14 15. This analysis involves the
implementation of decision trees, which perform recursive partitioning of the neural (feature)
space to lead to a non-linear predictive model. As decision tree-based models are susceptible to
small perturbations in the dataset, the variance in estimated prediction function was reduced
by the RF algorithm through 1000 bootstrap resampling of the original dataset, each of which
provides its own contribution (or vote) to the final prediction14. Generalizability of the
regression was conducted through a 10-fold cross-validation loop, in which the model
optimized in a portion of nurses was tested on out-of-sample nurses.
12
Permutation Analysis. The proficiency of the LASSO and RF procedures was assessed by
estimating the mean squared error (MSE), reflecting the deviation between nurses’ actual
behaviour and the behaviour estimated from the brain activity. This value was considered
significant if lower than the 5th percentile of the distribution of 1000 MSEs obtained by
rerunning the whole procedure on permuted datasets.
13
Supplementary Results
Empathy for Pain Task
Tables S4-5 report all behavioural (Table S4) and neuroimaging (Table S5) results associated
with this task. Behavioural effects include on-line accuracy and response time, but also post-
experimental rating sessions. Consistent with a previous study using the same paradigm5,
negative images (both painful and painless arousing) were associated with longer response
times, greater arousal, and lower accuracy, valence and familiarity, with respect to their
tailored controls (paired t-test: │t│ ≥ 1.94, p (1-tailed) ≤ 0.030; Wilcoxon sign-rank test: │Z│ ≥
2.38, p ≤ 0.017). In addition, painful images were associated with higher ratings of harm/pain,
than both their controls and painless arousing images (│t│ ≥ 7.84, p < 0.001; │Z│ ≥ 4.25, p <
0.001). However, unlike in our previous study on lay population, painful images were rated by
emergency nurses as more familiar, less negatively-valenced, and less arousing than painless
arousing images (the same for their corresponding controls – │t│ ≥ 2.40, p ≤ 0.027; │Z│ ≥ 2.24, p
< 0.025).
We then tested whether the behavioural responses to painful images could be
predictive of the three clinical indexes of interest. For this purpose we took both online (median
response times and accuracy) and offline (post-scanning ratings) measures for the condition of
interest (displayed in Table S4), and subjected them to massive univariate linear regression,
which led to no significant effects (│r│ ≤ 0.28, n.s.). When feeding all six measures to
multivariate regression using the same routines used for the analysis of brain data, we found
that a reliable prediction of the treatment application could be obtained using RF decision trees
(see Table S11). None of the other two indexes could be reliably predicted.
14
Balloon Analog Risk Task (BART)
Tables S6-7 report all behavioural (Table S6) and neuroimaging (Table S7) results associated
with this task. Behavioural data refer to the number of inflation in win trials, median response
times, the amount of money gained in the overall session, and the median subjective rating. For
each of these four measures, a linear relationship with the clinical indexes of interest was
tested with no significant results (│r│ ≤ 0.30, n.s.). Furthermore, when feeding all four
behavioural measures to multivariate regression using the same LASSO and RF approaches
employed for the analysis of brain signal, we found no reliable prediction (see Table S11).
Social Harm Avoidance Monitoring Experiment (SHAME)
Tables S8-9 report all the behavioural effects associated with the SHAME. Post-scanning ratings
obtained from all four kinds of errors suggest that nurses felt greater empathic pain, but also
greatest sadness, when observing an error with painful (vs. painless) outcome. Instead, greater
shame, guilt and anger were reported when nurses committed themselves an error (regardless
of its painful/painless outcome) relative to when they observed the confederate in the scanner
committing an error.
We then tested whether the behavioural responses to SHAME could be predictive of the
three clinical indexes of interest. For this purpose we took the ratings of pain, guilt, shame, fear,
sadness and anger (acquired in the post-scanning session) associated with One’s Painful Errors
(see Table S9). Furthermore we also took into consideration three online measures from when
the subject was playing the task in the MRI (see Table S8): the median response times,
accuracy, and median trial difficulty (as difficulty was automatically adapted according to
15
participants’ performance, the median trial difficulty throughout the experimental session is an
indirect measure of how challenging the task was for each subject). For each of these nine
measures, a linear relationship with the clinical indexes of interest was tested with no
significant results (│r│ ≤ 0.31, n.s.). When feeding all 9 measures to multivariate regression
using the same routines used for the analysis of brain data, we found that a reliable prediction
of the Documentation Rate could be obtained using RF decision trees (see Table S11). None of
the other two indexes could be reliably predicted.
Finally, Table S10 reports the regions involved when participants observed harmful
consequences of their own errors, relative to the condition in which pain was self-caused by the
colleague outside the scanner. We assessed whether these responses could be influenced by
the personal/professional relationship between the pair of nurses engaged in the task. Personal
closeness was assessed by the IOS questionnaire4 as implemented in the post-scanning debrief,
whereas professional closeness was assessed by calculating the absolute difference in age and
years of experience between the two nurses engaged in the task (no difference reflects
stronger similarity in professional status than a large difference). We then run a univariate
linear regression analysis, in which the neural responses to one’s painful errors were fitted
against each of these three measures (IOS, age difference, experience differences). No
significant effects of personal/professional closeness were found.
16
Supplementary References
1. Koban L, Corradi-Dell’Acqua C, Vuilleumier P. Integration of error agency and representation of others’ pain in the anterior insula. J Cogn Neurosci 2013; 25: 258–72
2. Corradi-Dell’Acqua C, Tusche A, Vuilleumier P, Singer T. Cross-modal representations of first-hand and vicarious pain, disgust and fairness in insular and cingulate cortex. Nat Commun 2016; 7: 10904
3. Sharvit G, Corradi-Dell’Acqua C, Vuilleumier P. Modality-specific effects of aversive expectancy in anterior insula and medial prefrontal cortex. Pain 2018; 159: 1529–1542
4. Aron A, Aron EN, Smollan D. Inclusion of Other in the Self Scale and the structure of interpersonal closeness. J Pers Soc Psychol 1992; 63: 596–612
5. Corradi-Dell’Acqua C, Hofstetter C, Vuilleumier P. Felt and Seen Pain Evoke the Same Local Patterns of Cortical Activity in Insular and Cingulate Cortex. J Neurosci 2011; 31: 17996–8006
6. Feinberg DA, Moeller S, Smith SM, et al. Multiplexed Echo Planar Imaging for Sub-Second Whole Brain FMRI and Fast Diffusion Imaging. PLOS ONE 2010; 5: e15710
7. Rao H, Korczykowski M, Pluta J, Hoang A, Detre JA. Neural correlates of voluntary and involuntary risk taking in the human brain: an fMRI Study of the Balloon Analog Risk Task (BART). NeuroImage 2008; 42: 902–10
8. Schonberg T, Fox CR, Mumford JA, Congdon E, Trepel C, Poldrack RA. Decreasing ventromedial prefrontal cortex activity during sequential risk-taking: an FMRI investigation of the balloon analog risk task. Front Neurosci 2012; 6: 80
9. Calhoun VD, Wager TD, Krishnan A, et al. The impact of T1 versus EPI spatial normalization templates for fMRI data analyses. Hum Brain Mapp 2017; 38: 5331–42
10. Chang LJ, Gianaros PJ, Manuck SB, Krishnan A, Wager TD. A Sensitive and Specific Neural Signature for Picture-Induced Negative Affect. PLoS Biol 2015; 13: e1002180
11. Krishnan A, Woo C-W, Chang LJ, et al. Somatic and vicarious pain are represented by dissociable multivariate brain patterns. eLife 2016; 5: e15166
12. Wager TD, Atlas LY, Lindquist MA, Roy M, Woo C-W, Kross E. An fMRI-Based Neurologic Signature of Physical Pain. N Engl J Med 2013; 368: 1388–97
13. Zunhammer M, Geis S, Busch V, Eichhammer P, Greenlee MW. Pain modulation by intranasal oxytocin and emotional picture viewing — a randomized double-blind fMRI study. Sci Rep 2016; 6: 31606
17
14. Breiman L. Random Forests. Mach Learn 2001; 45: 5–32
15. Jaiantilal A. Classification and regression by randomforest-matlab [Internet]. 2009. Available from: https://code.google.com/archive/p/randomforest-matlab/
16. Friston KJ, Worsley KJ, Frackowiak RSJ, Mazziotta JC, Evans AC. Assessing the significance of focal activations using their spatial extent. Hum Brain Mapp 1993; 1: 210–20
17. Chumbley JR, Friston KJ. False discovery rate revisited: FDR and topological inference using Gaussian random fields. NeuroImage 2009; 44: 62–70
18
Supplementary Tables
Table S1
Region of Interest mask for the Empathy for Pain Task. Regions included in both univariate and multivariate analysis defined from independent data5. The table lists the regions displaying differential activity for the contrast Painful – Control.
SIDE Coordinates
T(27) Cluster size X y z
Painful – PainfulControl
Anterior Insula R 39 29 -4 3.93 66*
Inferior Frontal Gyrus R 42 38 5 6.21
Middle/Posterior Insula R 42 5 -7 6.85 260***
Amygdala R 21 -4 -16 7.25
Anterior Insula L -33 23 -1 5.97
292*** Middle/Posterior Insula L -36 -4 8 4.26
Amygdala L -21 -4 -19 5.29
Supramarginal R 60 -22 38 8.56 301***
Postcentral Gyrus R 63 -16 29 7.08
Supramarginal L -54 -25 35 7.07 132***
Precentral Gyrus R 45 8 26 5.82 137***
***p < 0.001; **p < 0.01; *p < 0.05 family-wise corrected for the whole brain
19
Table S2
Region of Interest mask for the Balloon Analog Risk Task. Regions included in both univariate and multivariate analysis defined from independent data8. The table lists the regions displaying differential activity for the contrast monetary loss – implicit baseline.
SIDE Coordinates
T(15) Cluster size x Y z
Monetary Loss – Implicit Baseline
Anterior Insula R 42 23 -4 7.74 184***
Ventral Insula R 30 14 -19 4.74
Anterior Insula L -39 17 -1 5.30 156**
Ventral Insula L -33 14 -16 6.20
Pre-central Gyrus R 45 11 32 6.54 167**
Middle Frontal Gyrus R 42 23 47 5.62
Inferior Frontal Sulcus R 45 35 23 4.98 67*
Thalamus/Midbrain M -5 -31 37 5.77 85*
*p < 0.05 family-wise corrected for multiple comparisons at the cluster level
20
Table S3
Region of Interest mask for SHAME. Regions included in both univariate and multivariate analysis defined from independent data1. The table lists the regions displaying differential activity for the contrast One’s – Others’ Painful Errors. All clusters are displayed with a height threshold corresponding to p < 0.001 (uncorrected), and survive FWE16 or FDR17 correction for multiple comparisons for the whole brain at the cluster level.
SIDE Coordinates
T(15) Cluster size x y z
One’s – Others’ Painful Errors
Anterior Insula R 45 14 -7 5.21 58†
Putamen R 30 5 -1 5.22 71*
Anterior Insula L -39 11 -4 3.91 114**
Putamen L -21 8 5 5.60
Superior Frontal Gyrus R 36 38 53 5.38 140***
Superior Frontal Gyrus L -21 53 26 6.60 123**
Cerebellum R 27 -49 -28 5.24 59*
Anterior Middle Cingulate Cortex M 6 23 23 7.49 121**
Supplementary Motor Area M -9 14 58 7.70 111**
***p < 0.001; **p < 0.01; *p < 0.05 family-wise corrected for the whole brain; † p < 0.05
false-discovery-rate corrected for the whole brain
21
Table S4
Behavioural data from the Empathy for Pain Task. For each measure of interest, the average value associated with the four conditions is displayed together with 95% confidence intervals.
Painful Control Arousing ArousingControl
Resp. Times (ms) 1426 [1341, 1539] 1362 [1273, 1445] 1441 [1357, 1551] 1329 [1255, 1409]
Accuracy (%) 70.97 [64.02, 75.57] 83.97 [73.62, 89.61] 72.17 [64.81, 76.68] 81.55 [72.54, 87.06]
Arousal 6.15 [5.27, 7.04] 1.81 [1.48, 2.29] 7.01 [6.45, 7.64] 2.71 [2.24, 3.24]
Valence -0.73 [-1.20, -0.25] 1.23 [0.94, 1.69] -2.11 [-2.42, -1.78] 0.90 [0.65, 1.22]
Familiarity 5.29 [4.45, 6.12] 8.72 [8.27, 9.12] 2.24 [1.82, 2.75] 4.63 [4.05, 5.20]
Pain 8.92 [8.44, 9.28] 2.23 [1.58, 3.28] 3.26 [2.16, 4.69] 1.39 [1.13, 1.90]
22
Table S5
Neural Activations for the Empathy for Pain Task. Regions displaying differential activity for the
contrast Painful – Control Images, and increased activity for Painful Images with nurses’
Documentation Rate. All clusters survive correction for multiple comparisons for the whole
brain at the cluster level 16, or small-volume correction for a region of interest mask (described
in Table S1).
SIDE Coordinates
T(32) Cluster size X Y Z
Painful – Control Images
Middle/Posterior Insula R 42 -7 -1 6.64 256***
Amygdala R 24 -4 -16 8.39
Middle/Posterior Insula L -39 -1 -7 3.89 96*
Amygdala L -21 -7 -13 7.15
Inferior Frontal Gyrus R 45 38 11 8.71 133**
Inferior Frontal Gyrus L -42 32 14 6.94 116**
Precentral Gyrus R 48 8 26 6.87 160**
Precentral Gyrus L -45 5 23 7.38 167**
Inferior Temporal Gyrus R 51 -55 -10 9.23
2983***
Fusiform Gyrus R 30 -49 -13 9.07
Calcarine Gyrus R 18 -94 -4 10.02
Intraparietal Sulcus R 27 -64 44 8.04
Supramarginal/Postcentral Gyrus R 63 -22 26 9.39
Inferior Temporal Gyrus L -45 -55 -7 7.86
Fusiform Gyrus L -30 -49 -16 8.90
Calcarine Gyrus L -21 -94 -5 10.24
Periaqueductal Grey/Midbrain M -3 -31 -4 6.68
Intraparietal Sulcus L -21 -64 44 5.24 87*
Supramarginal/Postcentral Gyrus L -63 -25 32 8.20 133**
Painful Images*Documentation Rate
Postcentral Gyrus R 60 -16 32 4.50† 18
Middle Occipital Gyrus R 36 -73 14 4.51 89*
***p < 0.001; **p < 0.01; *p < 0.05 family-wise corrected for the whole brain; † p < 0.05
family-wise corrected for small volume
23
Table S6
Behavioural data from the Balloon Analog Risk Task. In keeping with previous studies, we measured the average number of inflations in each trial (excluding trials associated with negative outcome), the response times associated with each choice, and the overall money gained during the experimental session. Furthermore, we also considered subjects’ median anxiety rating collected along the whole experimental session. For each measure of interest, the average value is displayed together with 95% confidence intervals.
# inflations Response Times (ms) Money gained (CHF) Anxiety [0-100]
6.03 [5.79, 6.26] 584 [536, 643] 50.08 [45.23, 55.02] 31.25 [25.33, 37.42]
24
Table S7
Neural Activations for the Balloon Analog Risk Task. Regions displaying differential activity for
the contrast monetary loss – win, and increased activity for Monetary Loss with nurses’ CI Rate.
SIDE Coordinates
T(31) Cluster
size x Y z
Monetary Loss – Win
Anterior Insula R 39 17 -4 9.79
603*** Ventral insula R 39 -1 -16 10.97
Posterior Insula R 39 -10 -7 6.36
Anterior Insula L -36 20 -7 11.74
493*** Ventral Insula L -27 8 -22 9.54
Posterior Insula L -39 -4 -13 5.74
Temporo-Parietal Junction R 60 -40 20 4.65 68*
Middle Occipital Gyrus L -30 -88 14 4.93 168***
Inferior Occipital Gyrus L -45 -73 2 5.51
Precentral/Postcentral Gyrus R 39 -16 50 3.97
1025*** Supplementary Motor Area M 9 8 62 8.19
Middle Cingulate Cortex M -6 17 35 7.78
Calcarine Cortex R 18 -64 14 5.06
1802***
Lingual Gyrus R 27 -61 -7 8.74
Middle Temporal Gyrus R 46 -64 -4 5.09
Calcarine Cortex L -15 -70 11 6.10
Fusiform Gyrus L -27 -67 -7 7.67
Thalamus M 9 -28 -7 8.08
Monetary Loss*CI Rate
Ant. Insula/Inf. Frontal Gyrus R 24 32 -10 4.96 170***
Mid. Insula-Opercular Junction R 21 -13 14 5.06 69*
Putamen R 30 -4 11 4.98
Ant. Insular-Opercular Junction L -27 29 20 4.90 156**
Mid. Insula-Opercular Junction L -36 5 17 5.87 128**
Hippocampus R 30 -28 -7 4.55 104**
Thalamus R 21 -34 -1 4.21
Angular Gyrus R 39 -52 20 4.45 63*
Middle Occipital Gyrus R 45 -76 23 4.95 120**
Middle Occipital Gyrus L -33 -64 17 5.02 88*
Cerebellum R 12 -37 -34 5.49 759***
Cerebellum L -42 -58 -37 4.91
25
Middle Cingulate Cortex M -6 -4 29 5.33 261***
***p < 0.001; **p < 0.01; *p < 0.05 family-wise corrected for the whole brain.
Table S8
Behavioural data from the SHAME: online measures. For each measure of interest, the average value associated with the four conditions is displayed together with 95% confidence intervals.
Difficulty (dots diff) Accuracy (%) Resp. Times (ms) Pain Threshold (deg)
Subject in the scanner (self)
4.75 [4.57, 4.88] 52.01 [47.28, 54.27] 928 [848, 1001] ---
Subject outside the scanner
(other) 4.84 [4.61, 4.97] 51.85 [50.11, 55.17] 953 [877, 1024] 48.31 [47.78, 48.75]
26
Table S9
Behavioural data from the SHAME: post-scan rating measures. For each measure of interest, the average value associated with the four kinds of errors is displayed together with 95% confidence intervals. “Self” refers to the case in which the participant in the MRI scanner made a mistake, whereas “Other” refers to the case in which the participants in the MRI scanner observed the confederate making a mistake. “Pain” and “NoPain” refer to errors with painful and painless outcomes respectively.
Pain Fear Shame Guilt Sadness Anger
Self-Pain 2.61 [2.07, 3.35] 1.93 [1.50, 2.50] 2.81 [2.29, 3.30] 3.77 [3.13, 4.23] 2.42 [1.93, 2.94] 3.26 [2.68, 3.80]
Other-Pain 2.61 [2.03, 3.39] 2.45 [1.91, 3.03] 1.48 [1.22, 1.97] 1.42 [1.16, 1.85] 2.29 [1.76, 2.88] 1.74 [1.35, 2.29]
Self-NoPain 1.64 [1.28, 2.07] 1.45[1.16, 1.94] 2.52 [2.03, 3.03] 2.84 [2.28, 3.37] 1.74 [1.37, 2.23] 2.58 [2.06, 3.10]
Other-NoPain 1.45 [1.16, 1.85] 1.64 [1.30, 2.17] 1.42 [1.16, 1.91] 1.42 [1.14, 1.94] 1.84 [1.45, 2.34] 1.35 [1.07, 1.85]
27
Table S10
Neural Activations for the SHAME. Regions displaying differential activity for the contrast One’s
– Other’s Painful Errors, and increased activity for One’s Painful Errors with nurses’ CI Rate. All
clusters are displayed with a height threshold corresponding to p < 0.001 (uncorrected), and
survive correction for multiple comparisons for the whole brain, or small-volume correction for
a region of interest mask (described in Table S3).
SIDE Coordinates
T(30) Cluster
size x y z
One’s – Others’ Painful Errors
Anterior Middle Cingulate Cortex M -3 26 29 4.35† 25
One’s Painful Errors*CI Rate
Middle Frontal Gyrus L -33 8 53 6.79 104*
Anterior Middle Cingulate Cortex M 3 44 14 5.94 131**
**p < 0.01 *p < 0.05 family-wise corrected for the whole brain; † p < 0.05 family-wise
corrected for small volume
28
Table S11
Multivariate modelling of behavioural data. For each task, we modelled brain activity with multivariate regression based on LASSO and Random Forest (RF) approaches. The analysis of each task is described in terms of number of features fed to the multivariate regression, as well as measures of model’s proficiency (Mean Square Error [MSE]) at predicting each of the three indexes from the delegated analgesia protocol. Significant predictions are highlighted in bold, and significance cut-offs (5th percentile of a permutation-based MSE distribution) are displayed in squared brackets. Full details in Supplementary Methods.
Task # Features Algorithm Docum Rate CI Rate Treatment App.
From the neuroimaging session
Pain Images 6 LASSO 2.48 [2.23] 1.79 [1.55] 1.73 [1.49] ∙10-3
RF 2.28 [2.13] 1.92 [1.44] 1.29* [1.37]
Monetary Loss 4 LASSO 3.44 [3.10] 1.81 [1.71] 2.12 [1.87]
RF 3.91 [2.91] 1.93 [1.60] 2.24 [1.79]
One’s Painful Errors 9 LASSO 3.23 [3.14] 1.93 [1.68] 2.18 [1.98]
RF 2.86* [2.97] 2.10 [1.61] 2.30 [1.82]
* error lower than chance at p < 0.05