Review

Artificial intelligence applications in the intensive care unit

C. William Hanson III, MD, FCCM; Bryan E. Marshall, MD, FRCP, FRCA

T he amount of data acquiredelectronically from patientsundergoing intensive care hasgrown exponentially during

the past decade. Bedside equipment suchas pressure and flow transducers, infu-sion pumps, pulse oximeters, cardiac out-put monitors, and mechanical ventilatorsstore electronic data and are equippedwith computer interfaces. Modern bed-side monitors communicate with a hostof devices through data busses and inter-changeable, plug-in interfaces. Comput-erized intensive care systems interfacewith hospital databases including demo-graphic systems, electronic patientrecords, order-entry, laboratory, phar-macy, and radiology systems.

It is useful conceptually to recognizethat bedside data must be extracted andorganized to become information, andthat an expert must then interpret thisinformation before it becomes knowledgefor diagnostic and/or therapeutic pur-poses. This review is concerned primarilywith the second step in this cognitive

sequence, namely the use of computersto extract information from data and en-hance analysis by the human clinical ex-pert. We will argue that an as yet unre-alized role for the computer at thebedside is the extraction of informationfrom data rather than mere display. Avariety of novel, computer-based analytictechniques have been developed recently.Our purpose is to introduce these tech-niques and to discuss their potential forclinical applications in the intensive careunit (ICU).

A review of relevant literature cited inMEDLINE was performed between theyears 1966 and the present using Ovidsoftware. The search terms “critical care”and “intensive care” were exploded andcombined (18 and 423 “hits,” respectively).The terms “data warehouse” (18), “neuralnetwork” (4215), “genetic algorithm”(238), “fuzzy logic” (431), “case-basedreasoning” (55), “belief network” (47),and “data visualization” (42) were thenused as keywords to create individual datasets comprising all references to the spe-cific term. Each of these data sets wasthen individually combined (AND func-tion) with the critical care references. Acomprehensive review was made of allreferences cited in both data sets. In

addition, the smaller data sets (e.g., beliefnetwork) were thoroughly reviewed tofind abstracts or titles suggesting rele-vance to critical care practice. Finally,articles of historic significance (e.g., ref-erences to MYCIN) and appropriate textswere included for completeness.

Background

Nomenclature. It is useful to begin bydefining several terms that are commonlyused in describing computer systemsused for data analysis. A managementinformation system (MIS) has been de-fined as “a formalized computer informa-tion system that can integrate data fromvarious sources to provide the informa-tion necessary for management decisionmaking” (1).

The appropriate definition of artificialintelligence (AI) is controversial. AlanTuring, the English mathematician, de-vised what has become known as the Tur-ing test of computer intelligence. He sug-gested that a computer had artificialintelligence if it could successfully mimica human and thereby fool another hu-man.

An expert system is a computer pro-gram that simulates the judgment andbehavior of a human or an organization

From the Department of Anesthesia, Center forAnesthesia Research, University of PennsylvaniaHealth System, Philadelphia, PA.

Copyright © 2001 by Lippincott Williams & Wilkins

Objective: To review the history and current applications ofartificial intelligence in the intensive care unit.

Data Sources: The MEDLINE database, bibliographies of se-lected articles, and current texts on the subject.

Study Selection: The studies that were selected for reviewused artificial intelligence tools for a variety of intensive careapplications, including direct patient care and retrospective da-tabase analysis.

Data Extraction: All literature relevant to the topic was re-viewed.

Data Synthesis: Although some of the earliest artificial intel-ligence (AI) applications were medically oriented, AI has not beenwidely accepted in medicine. Despite this, patient demographic,clinical, and billing data are increasingly available in an electronicformat and therefore susceptible to analysis by intelligent soft-ware. Individual AI tools are specifically suited to different tasks,such as waveform analysis or device control.

Conclusions: The intensive care environment is particularlysuited to the implementation of AI tools because of the wealth ofavailable data and the inherent opportunities for increased effi-ciency in inpatient care. A variety of new AI tools have becomeavailable in recent years that can function as intelligent assis-tants to clinicians, constantly monitoring electronic data streamsfor important trends, or adjusting the settings of bedside devices.The integration of these tools into the intensive care unit can beexpected to reduce costs and improve patient outcomes. (CritCare Med 2001; 29:427–435)

KEY WORDS: intensive care unit; artificial intelligence; expertsystems; computer-assisted diagnosis; computer-assisted ther-apy; decision support techniques; neural networks; algorithms;fuzzy logic; data display; computer simulation; clinical decisionsupport systems; management decision support systems

427Crit Care Med 2001 Vol. 29, No. 2

with expert knowledge and experience ina particular field.

Data mining is the analysis of data forrelationships that have not previouslybeen discovered. The techniques used fordata mining can discover hidden associ-ations or sequences in data sets, cluster-ing of data points, and permit visualiza-tion of relationships among data orforecasting based on hidden patterns.Data mining is also known as knowledgediscovery, and derives its roots from sta-tistics, artificial intelligence, and ma-chine learning.

A data warehouse is a central reposi-tory for all or significant parts of the datathat an enterprise’s various business sys-tems collect. An alternative term is a datamart.

The Data Stream. The establishmentof modern data-intensive ICUs can prob-ably be traced to the introduction ofblood gas electrodes in the 1960s. Theoutstanding developments in bioengi-neering since then have resulted in animpressive amount of data being reliablyavailable at the bedside. For the mostpart, however, clinical practice has nottaken advantage of the rich, varied, andcontinuous stream of information. De-spite the wealth of electronically accessi-ble data, the synthesis and interpretationof information is done manually in manyof today’s ICUs with minimal preprocess-ing. Nurses laboriously transcribe infor-mation from monitors onto paperrecords. Data are lost, creatively inter-preted, averaged, and incorrectly tran-scribed in the process. Additionally, al-though a sicker patient requires morenursing interventions with consequentlyless time available for data transcription,it is these patients for whom accuratedata are most useful to clinicians for ret-rospective review and analysis.

There are several reasons for the du-rability of what would seem to be archaicmethods of data acquisition and tran-scription. Among the most often cited arethe legal and financial requirements fordocumentation of human observationsand decisions. These requirements arereal and are not likely to change. Theymandate written chart entries at regularintervals and at times of change, whichmust be made and signed by the respon-sible individual. The second basis for re-luctance by clinicians is the failed prom-ise of early expert systems, as discussedbelow.

Management InformationSystems

Management information systems canbe categorized as depicted in Table 1. Onegroup of decision support systems is themodel-driven or rule-based expert sys-tems (RBS). They are successful to theextent that they are able to represent thesubject material accurately and interfacewell with the user.

Rule-based systems can be thought ofas “top-down” systems. “Top-down” pro-gramming begins with a complex prob-lem and uses a reductionist approach,breaking the problem down into its con-stituent parts to arrive at the essentialcomponents that characterize it. This ap-proach can be used as a method for sim-ulating the thinking processes of the hu-man medical expert. In reality, however,experts make rapid, often intuitive, diag-noses by beginning with a few hypothesesselected by experience, followed by clini-cal or laboratory observations that fur-ther refine the differential diagnosis.

Rule-Based (Expert) Systems. Some ofthe first expert systems were developedfor medical care. Some notable early ex-pert systems were the MYCIN (2, 3), ON-COCIN (4, 5), and Internist/Quick Medi-cal Reference (6, 7) systems. MYCIN andONCOCIN were designed at Stanford andsimulated the performance of consultantsin infectious disease and oncology respec-tively. The Internist system was designedat the University of Pittsburgh to repro-duce the diagnostic behavior of an inter-nist. Unfortunately, these pioneering sys-tems have not been widely incorporatedinto the practice of medicine. One authorsuggests, “Artificial intelligence in medi-cine (AIM) has not been successful—ifsuccess is judged as making an impact onthe practice of medicine. Much recentwork in AIM has been focused inward(sic), addressing problems that are at thecrossroads of the parent disciplines ofmedicine and artificial intelligence. Now,

AIM must move forward with the insightsthat it has gained and focus on findingsolutions for problems at the heart ofmedical practice.” (8)

A variety of impediments have slowedthe general acceptance of medical expertsystems. They include the small marginof acceptable error in medical practice,the ready availability of experts in mostsettings, and the complexity of regulatoryrequirements. Many of these obstaclesmay disappear in the future. Fewer ex-perts are being trained, and expert sys-tems may become cost-effective in cer-tain environments. The performance ofexpert systems could improve to thepoint that they rival or exceed humanexperts. Finally, increased acceptance oftelemedicine may lower the barriers toacceptance of expert systems.

Data-Driven (Intelligent Assistant)Systems. A second, newer generation ofdecision support systems is data driven(9). Data-driven systems (DDS) take ad-vantage of the large quantity of data thatcan be acquired electronically to “discov-er” relationships and assume that futurebehavior can be predicted from past be-havior. They represent “bottom-up” sys-tems in which the data generated by asystem is used to describe the character-istics of the system. Data analytic toolssuch as these are typically less ambitiousthan expert systems in scope and scale,less expensive to develop and maintain,and well suited to act as intelligent assis-tants to human experts.

It is useful to contrast the model-driven and data-driven systems using afamiliar intensive care construct. A RBSwould include rules about the relation-ship between pulmonary artery occlusionpressure (PAOP) and cardiac output. Therules would be based upon well-under-stood and accepted physiologic principlessuch as the Frank-Starling mechanism,whereas a DDS might begin without pre-conceptions and then discover the Frank-Starling relationship for an individual pa-

Table 1. Decision support systems

Model Driven Decision Support Data Driven Decision Support

Monolithic ModularIncorporation of a body of knowledge Self-learningDesigned to reproduce the expert Typically act as intelligent assistants to an

expertAll solutions are preprogrammed Capable of arriving at novel, unexpected

solutions or observationsLikely to require substantial maintenance as

medical knowledge evolvesInherently autodidactic and therefore largely

self-maintained

428 Crit Care Med 2001 Vol. 29, No. 2

tient. Specific characteristics distinguishthe two approaches in this example. TheRBS embodies general rules about all pa-tients: “If the pulmonary artery occlusionpressure is less than or equal to 5 mm Hgand if the cardiac output is less than 2L/min, then give 10 mL/kg crystalloid.”Conversely, the DDS is used to discoverthe physiologic behavior of one individualfrom data acquired during continuousmonitoring of that patient: “When Mr.Smith’s PAOP decreases to less than 8mm Hg, his cardiac output decreases toless than 3 L/min.” The RBS imposesstructure on the data, whereas the DDSderives structure from the data.

Data Mining. New methods of dataanalysis and decision support have be-come available in the past decade, whichhave been euphemistically termed “datamining.” The techniques of data mininghave evolved from and depend on previ-ous generations of data analysis tools (Ta-ble 2). Several different techniques arecommonly used in data mining, or data-driven decision support. They includedata warehouses, neural networks, ge-netic algorithms, Bayesian or belief net-works, rule induction or case-based rea-soning, and machine learning. Fuzzylogic is another relatively new approachto programming that permits ambiguityin descriptions of data. Visualizationtechniques display large amounts of datain a comprehensive, comprehensiblefashion.

Data Warehousing. Data-driven deci-sion support benefits from the creation ofa data warehouse and an online analyticalprocessing system (OLAP). Decision sup-port systems consist of a well-organizeddatabase (the data warehouse) and an ac-cessible front-end (OLAP) that permitsflexible exploration and analysis of databy a nonprogrammer.

The medical director of an ICU coulduse an OLAP to acquire and analyze datafrom a data warehouse to answer ques-tions such as, “What is the length of stayof patients admitted (to my ICU) with thediagnosis of respiratory failure who areventilated for greater than 2 days.” Tra-ditionally, questions such as these wouldrequire a programmer to query a rela-tional database using a structured querylanguage, which in turn presupposed thatthe required data were available in a re-lational database. Using an OLAP, an ad-ministrator could answer the question atwhim using natural language rather thana specifically designed computer pro-gram. The administrator might then fol-low-up (or “drill down” in the vernacular)with a second question suggested by theanswer to the first, such as, “What wasthe mortality of those patients?”

Although data warehouses are widelyimplemented in industry, and to someextent in hospital administration, theyare essentially unavailable in the inten-sive care setting. Intelligent algorithmshave been developed to enhance the anal-ysis, functionality or display of informa-tion in the ICU as described below.

Neural Networks. Neural networks aredesigned to mimic the performance ofthe human brain. There are input nodes(or neurodes), output nodes, and a vari-able number of internal (or hidden) lay-ers. The nodes are connected with differ-ent architectures, but typically inputnodes are connected to hidden layernodes and they are in turn connected tooutput nodes. As the neural networklearns from (or “trains on”) a data set, theconnection weights are adjusted. In ef-fect, important connections are rein-forced (positively weighted) and unim-portant connections are punished(negatively weighted). Data are fed into

the input nodes, processed through thehidden layer(s), and the connectionweights to the output nodes are adjusted.

Neural nets are categorized based ontheir learning paradigm. In supervisednetworks, the outputs are known but theimportance of the relationship of a giveninput to an output is unknown beforetraining. In an ICU, a neural network canbe used to explore the relationshipsamong several physiologic variables. Forexample, Buchman used a neural net-work to evaluate the relationship of sev-eral demographic, pharmacologic, andphysiologic variables to ICU chronicity(Fig. 1) (10).

In unsupervised networks, the outputsare unknown and the system is encour-aged to find interesting, often unsus-pected, relationships among the data el-ements in large data sets. For example, ahypothetical unsupervised neural net-work might make the novel discoverythat a hypotensive episode of greater thanan hour’s duration immediately followingcardiac surgery is highly correlated withsubsequent development of pancreatitis.

Neural networks have been used in theICU setting in a variety of fashions, butmost extensively for outcome prediction.Neural networks have been shown to pre-dict length of stay in the ICU (10–12).Other neural network-based systemswere successful in predicting ICU mortal-ity (11, 13–15).

Another common application of neu-ral networks in the ICU is the real-timeanalysis of waveforms such as the electro-cardiogram and the electroencephalo-gram. One neural network-based algo-rithm identified cardiac ischemia withhigh sensitivity based on analysis of theST segment (16), whereas another diag-nosed myocardial ischemia in the emer-gency department patient (17–20). Neu-

Table 2. Evolution of data analytic techniques

Technique Typical Question Format Necessary Technologies Products/Vendors Characteristics

Data collection(1960–1970s)

How many patients were admittedto the ICU last year?

Data stored on tape/disks IBM Retrospective, static data interrogation,programmed queries

Data access (1980s) What was the mortality in ourcoronary artery bypass (CABG)last patients year?

Relational database andstructured querylanguage

Oracle, Sybase Retrospective, dynamic datainterrogation at record level,programmed queries

Data warehousing(1990s)

What was the CABG mortality lastyear? 3 Drill down to thosepatients with atrial fibrillation

Multidimensional databasesand online analyticalprocessing systems

Pilot, Microstrategy Retrospective dynamic datainterrogation at multiple levels,natural language queries

Data mining(emerging)

What’s causing the high rate ofatrial fibrillation in the ICU?

Massive data bases,multiprocessorcomputers

Lockheed, IBM Prospective, proactive informationdelivery

ICU, intensive care unit.

429Crit Care Med 2001 Vol. 29, No. 2

ral networks have also been used in theanalysis of electroencephalographic pat-terns in children (21) and adults sedatedwith midazolam (22). Finally, neural net-works have been used to analyze hemo-dynamic patterns in intensive care pa-tients (23, 24).

Neural networks can reveal unex-pected and otherwise undetectable pat-terns in large data sets. The major weak-ness in neural network solutions is thefact that the methods by which a relation-ship is discovered are hidden (or opaque)and therefore not readily understood orexplained.

Genetic Algorithms. Genetic algo-rithms were designed to find near opti-mal solutions to complicated problemsusing the principles of Darwinian selec-tion. For example, genetic algorithmshave been used to find a near optimalroute for the salesperson who needs totravel through several cities on a salestrip. The so-called traveling salespersonproblem was once considered noncom-putable as the number of cities becamelarge, but genetic algorithms provide abest approximation as an answer. Theprocess of optimization involves the fol-lowing: a) the creation of a number ofpossible solutions; b) competition amongthem using selection criteria (i.e., fastestroute, shortest route, least expensiveroute); and c) the elimination of “bad”solutions. Surviving solutions are then

permitted to mutate and cross-breed andcompete further. Ultimately, a highly de-sirable solution is selected from the set ofall possible solutions. Note that becausethey fail to explore all solutions, geneticalgorithms are quite efficient but cannotensure that the surviving solution is thebest possible choice.

An example of a hypothetical ICUproblem that might be susceptible to thisapproach is the determination of an op-timal staffing configuration for a group ofpatients with different acuity or nursingrequirements. Genetic algorithms havebeen used to determine the neural netconfiguration that was most accurate in

predicting prognosis in a group of 258ICU patients (14).

Fuzzy Logic. Strictly speaking, fuzzylogic is not a data-driven analytic ap-proach. Rather, it is a method of handlingdata that permits ambiguity, and as aresult, it is particularly suited to medicalapplications. One of the interesting iro-nies of medical practice is that its practi-tioners strive for objectivity and precisionwhile dealing with data that are inher-ently imprecise.

Fuzzy logic has proven to be wellsuited to a variety of industrial applica-tions, and fuzzy control strategies are, inmany cases, more efficient than tradi-

Figure 1. Supervised neural network: Neural net-work with three inputs, a hidden layer, and asingle output (O). The network is similar to onethat Buchman et al. (10) described for the pre-diction of chronicity in the intensive care unit.This network is supervised, i.e., it is trainedthrough feedback so that connections that createdesired outputs (better outcome prediction) arereinforced (thicker connections). HR, heart rate;BP, blood pressure.

Figure 2. Traditional vs. fuzzy classification schemes: Traditional classification of pulmonary arteryocclusion pressures into discrete sets compared to fuzzy classification using fuzzy sets. In the standardclassification scheme, any single pulmonary artery occlusion pressure (PAOP) (horizontal axis) hasfull membership (vertical axis) in only one set (low [L], normal [N], high [H]). Conversely, a singlePAOP can have simultaneous partial membership in more than one fuzzy set (very low [VL], very high[VH]). For example, when the occlusion pressure is 5.5 (dashed vertical line), it is classified into asingle classic set, whereas it is a “member” of two fuzzy sets (low and normal).

430 Crit Care Med 2001 Vol. 29, No. 2

tional alternatives. Fuzzy control systemsare used in applications as diverse as el-evator control, intelligent sensors, cli-mate control, and image stabilization invideo cameras.

Figure 2 shows the classically ac-cepted categorization of PAOPs as low,normal, and high. These sets are dis-crete—any single PAOP is a member ofonly one set. Fuzzy logic permits the useof overlapping sets, and therefore simul-taneous membership in more than oneset. Using fuzzy descriptors and controllogic in the ICU, very efficient infusioncontrollers have been designed that usefuzzy logical constructs like “if the bloodpressure is low, and the cardiac output islow normal, give dopamine slowly.”

Fuzzy control processes have beenused for the administration of anestheticsin the operating room (25–29). In theICU, fuzzy control strategies have beendesigned for the administration of fluid(30) and titration of oxygen therapy (31).Fuzzy controllers have also been de-

signed for the administration of vasodila-tors to control blood pressure in the peri-operative period (32–36) and duringdialysis (37, 38). Fuzzy logic has beenused to control mechanical ventilation(39, 40) and artificial hearts (41). Fuzzylogic diagnostic strategies have also beenused in the ICU to analyze physiologicdata during a simulated cardiac arrest(42), categorize oxygen destruction (43),interpret EEGs (21), and to distinguishreal alarms from artifacts in preterm in-fants (44).

Fuzzy logical systems are easy to con-figure and tune. Unlike neural networks,the logical constructs used in these sys-tems are easy to describe and closely ap-proximate the thinking processes used inclinical decision making.

Machine Learning. Machine learningalgorithms create rules or classificationschemes by searching through data forrelevant patterns. Unlike neural net-works, they generate rules and patternsthat can be evaluated and understood.

The success of machine learning al-gorithms requires the creation of a dataset with independent and dependentvariables. The machine learning algo-rithm is then used to explore the datafor interesting or unexpected relation-ships.

Figure 3 shows the use of a machinelearning algorithm to determine the re-lationship of five independent variables(preoperative beta-blocker therapy, se-rum magnesium level, serum potassiumlevel, central venous pressure, and medi-astinal drain output) to the developmentof atrial fibrillation (the dependent vari-able) in postoperative coronary artery by-pass patients.

Machine learning algorithms havebeen designed to derive rules for intelli-gent alarms on respiratory systems (45,46).

Case-Based Reasoning. Case-basedreasoning is a method of arriving at so-lutions by analogy. A case consists of aseries of attributes describing a situation

Figure 3. Data parsing by machine learning algorithm: A machine learning algorithm was used to partition this hypothetical database of cardiac surgicalpatients and demonstrate the correlation between two independent variables (low serum potassium and absence of beta blockade) with the dependentvariable (atrial fibrillation). Y, yes; N, no.

431Crit Care Med 2001 Vol. 29, No. 2

and a related solution. A series of repre-sentative cases are accumulated into acase base. The case base can then be ex-amined by the use of probes or test cases(to find all cases in the database like a testcase). The probe may use a subset of theavailable attributes and constraints onthe available solutions to find a set ofanalogous cases.

Case-based reasoning is well suited tothe derivation of solutions when there arediscontinuities in the case base. Case-based reasoning might be used to findcases similar to a hypothetical, newly pre-senting, undiagnosed patient with rapidlyevolving encephalitis, acute respiratorydistress syndrome, and renal failure (Fig.4). New diseases such as Hanta virus,acquired immune deficiency syndrome,and Legionnaire’s disease are discoveredthrough a process analogous to case-based reasoning. Although a case data-base might not contain a case with aspecific constellation of conditions, case-based systems can extrapolate from otherpatients with similar histories. Case-based reasoning simulates the reasoningprocesses of an expert who has a large

catalog stored in his brain and can rapidlyrecall analogous cases.

Case-based reasoning approaches havebeen used in the ICU to select a group ofpatients with similar characteristics froma demographic database (11). They havealso been used to select an antibiotic reg-imen (47, 48) and to project the course ofrenal function (49).

Bayesian (Belief) Networks. Thepower of Bayes’ Rule has been exploitedto develop extremely powerful, readilyunderstood learning algorithms knownas Bayesian networks or belief networks.They are versatile and have been used fora wide variety of functions including mil-itary applications such as the rapid iden-tification of incoming military targets(missiles, aircraft, vessels) and deploy-ment of counterattacks. Belief networkshave also been used to assist computersoftware support staff in diagnosing prob-lems on help lines. In the latter applica-tion, probability trees are constructed de-scribing the potential etiologies of acustomer problem.

The initial design of a belief networkrequires the configuration of a tree of

nodes describing the relationship of vari-ables to one another. Figure 5 shows aBayesian logic tree in which the branchesdescribe the relationships among thephysiologic variables that determine sys-temic oxygen delivery. An expert con-structs a table describing the pretestprobabilities that a given combination ofheart rate and wedge pressure will resultin a given cardiac output. The network istuned (and probabilities recalculated) byexposure to real data.

Bayesian networks are well suited tothe intensive care environment becauseof their speed and comprehensibility(transparency) in addition to the fact thatBayes’ theorem is well understood andaccepted by modern physicians. Beliefnetworks have been used in the intensivecare unit to evaluate EEGs (50) and toestablish prognosis in patients with headinjuries (51).

Data Visualization. Data visualizationis a term used to describe the intelligentdepiction of information using proximity,grouping, shape, color, animation, andother techniques to enhance data com-

Figure 4. Case-based reasoning: In case-based reasoning, a probe (test case) is designed to query adatabase for cases with similar attributes. Here a hypothetical test case has encephalitis, acuterespiratory disease (ARDS), and renal failure. Two other cases with similar attributes are retrieved, (1)encephalitis, ARDS, hepatic and renal failure and (2)encephalitis, renal failure, myocardial infarctionand ARDS), whereas two other dissimilar cases (3 and 4) are not.

Figure 5. Architecture of a belief network: A be-lief network describing the probabilities relatingseveral known variables to oxygen delivery in apatient who lacks a pulmonary artery catheter. Inthis example, the central venous pressure (CVP),left ventricular ejection fraction (EF), heart rate(HR), hemoglobin (HGB), and oxygen saturation(O2sat) are known and fall into predefinedranges. For example, the heart rate is known andfalls within the predefined normal range. There-fore the probability that the heart rate is low orhigh is zero, whereas the probability that theheart rate is normal is one. The probabilitiescharacterizing the unknowns, e.g., cardiac index(CI), depend on the probabilities of the parentnodes (CVP, EF, and HR) at a given point in time.The network can be configured with initial prob-abilities by an expert and then tuned with realdata. The patient described in the current state ofthis network has normal CVP, HR, and HGB, ahigh EF and low O2sat and therefore has an 82%probability of having normal oxygen delivery(O2del).

432 Crit Care Med 2001 Vol. 29, No. 2

prehension by the observer. Figure 6shows a multidimensional array usingthree axes (x, y, z) and size to depict thelikelihood that a disease process will betreated with a therapeutic modality.

Medical applications of visualizationtechniques have largely been restricted tothree-dimensional anatomical recon-struction for surgical applications. How-ever, it is possible to use data visualiza-tion for the display of hemodynamic data(52). Data visualization techniques havealso been used in an attempt to improvethe depiction of medical data in the pa-tient medical record (45, 53, 54).

CONCLUSIONS

There are a large variety of novel datamanagement tools that have becomeavailable over the past decade. They varyin their suitability to a given task, thedegree to which the achieved solution isunderstandable (transparent or opaque),and the ease of configurability. Becauseof the data rich nature of the ICU, theapplications to which these techniquescan be applied are varied, ranging fromwaveform analysis to outcome prediction.

There are obstacles to the rapid adop-tion of the data analysis tools described inthis manuscript. Although OLAPs havebeen widely adopted as business and fi-nancial decision support systems, theyhave essentially been ignored in clinicaldecision support. There are examples ofmedically oriented OLAP tools such asInfomine (Infocure, Atlanta, GA), how-

ever this is a data mining tool designedprimarily to allow financial analysis of amedical practice. The query tools avail-able with clinically oriented critical careinformation systems such as Picis Care-Suite (Picis, Arlington, VA) and SunriseCritical Care (Eclipsys, Delray Beach, FL)permit fixed format reports, retrospec-tive, dynamic data interrogation, or pro-grammed queries (Table 2). As patientinformation databases mature and be-come commonly available online, clini-cally oriented OLAPs will almost certainlybecome available. Greater physician lead-ership in the design and implementationof these systems will undoubtedly en-hance the speed with which they areadopted.

Many of the studies cited in this re-view are small and can be characterizedas proofs of concept or feasibility ratherthan true side-by-side comparisons of acomputer technique with the currentstandard of care. In addition, medicine isan appropriately conservative discipline,and regulatory agencies such as the Foodand Drug Administration have stringentrequirements for the approval of newdrugs and technologies to protect the in-dividual patient. Paradoxically, fuzzy con-trol systems and neural networks are be-ing used in automated trains and elevators,helicopters, and even spacecraft.

It is unlikely that intelligent softwarewill replace the clinician in the way thatthe original medical expert systems wereconceived. Future systems are morelikely to act as intelligent agents for spe-

cialized, complicated problems, and aregenerally intended to enhance the perfor-mance of a human expert. The applica-tions described above vary in their suit-ability to specific tasks, and it is likelythat they will be combined in the smartintensive care unit of the future.

Neural networks and fuzzy systemsare particularly useful for waveform anal-ysis. They will be integrated into bedsidemonitors and continuously analyze wave-forms for known patterns (cardiac isch-emia, hypovolemia).

Fuzzy controllers will be integratedinto bedside devices such as fluid andmedication infusion devices, mechanicalventilators, and dialysis machines.

Belief networks and neural networkswill be used in the development of smartalarms that integrate multiple datastreams (hemodynamics, laboratory data,other monitors), which display eventprobability (e.g., developing sepsis oracute respiratory distress syndrome).

Data visualization tools will permitthe clinician to interrogate and analyzelaboratory and hemodynamic trends inan individual patient at a glance.

Case-based reasoning, machine learn-ing algorithms, and visualization toolswill be used to analyze information fromdata warehouses describing the charac-teristic of an individual ICU. For example,a mini-epidemic of infections attributableto a resistant bacterial organism might beidentified and followed using tools suchas these.

Figure 6. Visualization used to summarize data. Visualization techniques are used here to demonstratethat this hypothetical intensive care unit cares for a large number (big sphere) of head-injured patientswho are rarely in need of continuous dialysis, mechanical ventilation, or invasive hemodynamicmonitoring. ARDS, acute respiratory distress syndrome.

D ata-driven deci-

sion support

tools will permit

the busy clinician (physi-

cian, nurse, respiratory ther-

apist) to function more effi-

ciently, caring for more

patients more safely in much

the same way that these

same tools have been used to

enhance the efficiency of

business applications.

433Crit Care Med 2001 Vol. 29, No. 2

Data-driven decision support tools willpermit the busy clinician (physician,nurse, respiratory therapist) to functionmore efficiently, caring for more patientsmore safely in much the same way thatthese same tools have been used to en-hance the efficiency of business applica-tions. Modern assembly lines are moreproductive and make fewer errors thanever before, partly through the applica-tion of data-driven decision support (8).Similar improvements can be expected inand will be demanded of the medical in-dustry in the immediate future.

REFERENCES

1. Telem M: MIS implementation in schools: Asystems socio-technical framework. ComputEduc 1996; 27:85–93

2. Shortliffe EH, Davis R, Axline SG, et al: Com-puter-based consultations in clinical thera-peutics: Explanation and rule acquisition ca-pabilities of the MYCIN system. ComputBiomed Res 1975; 8:303–320

3. Yu VL, Buchanan BG, Shortliffe EH, et al:Evaluating the performance of a computer-based consultant. Comput Programs Biomed1979; 9:95–102

4. Hickam DH, Shortliffe EH, Bischoff MB, etal: The treatment advice of a computer-basedcancer chemotherapy protocol advisor. AnnIntern Med 1985; 103:928–936

5. Shortliffe EH: Medical expert systems—Knowledge tools for physicians. West J Med1986; 145:830–839

6. Masarie FEJ, Miller RA, Myers JD: INTER-NIST-I properties: Representing commonsense and good medical practice in a com-puterized medical knowledge base. ComputBiomed Res 1985; 18:458–479

7. Miller RA, Pople HEJ, Myers JD: Internist-1,an experimental computer-based diagnosticconsultant for general internal medicine.N Engl J Med 1982; 307:468–476

8. Dhar V, Stein R: Seven Methods for Trans-forming Corporate Data into Business Intel-ligence. Upper Saddle River, NJ, PrenticeHall, 1997

9. Coiera EW: Artificial intelligence in medi-cine: The challenges ahead. J Am Med InfAssoc 1996; 3:363–366

10. Buchman TG, Kubos KL, Seidler AJ, et al: Acomparison of statistical and connectionistmodels for the prediction of chronicity in asurgical intensive care unit. Crit Care Med1994; 22:750–762

11. Frize M, Solven FG, Stevenson M, et al: Com-puter-assisted decision support systems forpatient management in an intensive careunit. Medinfo 1995; 8:1009–1012

12. Tu JV, Guerriere MR: Use of a neural networkas a predictive instrument for length of stayin the intensive care unit following cardiacsurgery. Comput Biomed Res 1993; 26:220–229

13. Doig GS, Inman KJ, Sibbald WJ, et al: Mod-eling mortality in the intensive care unit:comparing the performance of a back-propagation, associative-learning neural net-work with multivariate logistic regression.Proc Annu Symp Comput Applications MedCare 1993; 361–365

14. Dybowski R, Weller P, Chang R, et al: Pre-diction of outcome in critically ill patientsusing artificial neural network synthesisedby genetic algorithm. Lancet 1996; 347:1146–1150

15. Izenberg SD, Williams MD, Luterman A: Pre-diction of trauma mortality using a neuralnetwork. Am Surg 1997; 63:275–281

16. Maglaveras N, Stamkopoulos T, Pappas C, etal: An adaptive backpropagation neural net-work for real-time ischemia episodes detec-tion: development and performance analysisusing the European ST-T database. IEEETrans Biomed Eng 1998; 45:805–813

17. Baxt WG: A neural network trained to iden-tify the presence of myocardial infarctionbases some decisions on clinical associationsthat differ from accepted clinical teaching.Med Decis Making 1994; 14:217–222

18. Baxt WG: Application of artificial neural net-works to clinical medicine [see comments][Review]. Lancet 1995; 346:1135–1138

19. Baxt WG: Use of an artificial neural networkfor the diagnosis of myocardial infarction[published erratum appears in Ann InternMed 1992; 116:94] [see comments]. AnnIntern Med 1991; 115:843–848

20. Baxt WG, Skora J: Prospective validation ofartificial neural network trained to identifyacute myocardial infarction [see comments].Lancet 1996; 347:12–15

21. Si Y, Gotman J, Pasupathy A, et al: An expertsystem for EEG monitoring in the pediatricintensive care unit. Electroencephalogr ClinNeurophysiol 1998; 106:488–500

22. Veselis RA, Reinsel R, Sommer S, et al: Use ofneural network analysis to classify electroen-cephalographic patterns against depth of mi-dazolam sedation in intensive care unit pa-tients. J Clin Monitoring 1991; 7:259–267

23. Spencer RG, Lessard CS, Davila F, et al: Self-organising discovery, recognition and pre-diction of haemodynamic patterns in the in-tensive care unit. Med Biol Eng Comput1997; 35:117–123

24. Hanson CW, Weiss Y, Frasch F, et al: Neuro-fuzzy analysis of hemodynamic data. Abstr.Anesthesiology 1998; 89(suppl):474

25. Mason DG, Ross JJ, Edwards ND, et al: Self-learning fuzzy control of atracurium-induced neuromuscular block during sur-gery. Med Biol Eng Comput 1997; 35:498–503

26. Mason DG, Ross JJ, Edwards ND, et al: Self-learning fuzzy control with temporal knowl-edge for atracurium-induced neuromuscularblock during surgery. Comput Biomed Res1999; 32:187–197

27. Martin JF: Fuzzy control in anesthesia [Edi-torial; comment]. J Clin Monitoring 1994;10:77–80

28. Tsutsui T, Arita S: Fuzzy-logic control ofblood pressure through enflurane anesthesia[see comments]. J Clin Monitoring 1994; 10:110–117

29. Zbinden AM, Feigenwinter P, Petersen-FelixS, et al: Arterial pressure control with isoflu-rane using fuzzy logic. Br J Anaesth 1995;74:66–72

30. Hanson CW, Weiss Y, Frasch F, et al: A fuzzycontrol strategy for postoperative volume re-suscitation. Abstr. Anesthesiology 1998;89(suppl):475

31. Sun Y, Kohane I, Stark AR: Fuzzy logic as-sisted control of inspired oxygen in venti-lated newborn infants. Proc Ann Sym Com-put Applications Med Care 1994:756–761

32. Ying H, McEachern M, Eddleman DW, et al:Fuzzy control of mean arterial pressure inpostsurgical patients with sodium nitroprus-side infusion. IEEE Trans Biomed Eng 1992;39:1060–1070

33. Ying H, Sheppard L, Tucker D: Expert-system-based fuzzy control of arterial pres-sure by drug infusion. Med Prog Technol1988; 13:203–215

34. Ying H, Sheppard LC: Real-time expert-system-based fuzzy control of mean arterialpressure in pigs with sodium nitroprussideinfusion. Med Prog Technol 1990; 16:69–76

35. Oshita S, Nakakimura K, Kaieda R, et al:[Application of the concept of fuzzy logisticcontroller for treatment of hypertension dur-ing anesthesia] [Japanese]. Masui 1993; 42:185–189

36. Fukui Y, Masuzawa T. [Development of fuzzyblood pressure control system] [Japanese].Iyo Denshi to Seitai Kogaku 1989; 27:79–85

37. Nordio M, Giove S, Lorenzi S, et al: Projec-tion and simulation results of an adaptivefuzzy control module for blood pressure andblood volume during hemodialysis. ASAIO J1994; 40:M686–M690

38. Nordio M, Giove S, Lorenzi S, et al: A newapproach to blood pressure and blood vol-ume modulation during hemodialysis: anadaptive fuzzy control module. Int J ArtifOrgans 1995; 18:513–517

39. Nemoto T, Hatzakis GE, Thorpe CW, et al:Automatic control of pressure support me-chanical ventilation using fuzzy logic. Am JRespir Crit Care Med 1999; 160:550–556

40. Schaublin J, Derighetti M, Feigenwinter P, etal: Fuzzy logic control of mechanical venti-lation during anaesthesia. Br J Anaesth 1996;77:636–641

41. Kaufmann R, Becker K, Nix C, et al: Fuzzycontrol concept for a total artificial heart.Artif Organs 1995; 19:355–361

42. Goldman JM, Cordova MJ: Advanced clinicalmonitoring: considerations for real-time he-modynamic diagnostics. Proc Annu SympComput Applications Med Care 1994;752–755

43. Sailors RM, East TD, Wallace CJ, et al: Asuccessful protocol for the use of pulse oxim-etry to classify arterial oxygenation into fourfuzzy categories. Proc Annu Symp ComputApplications Med Care 1995; 248–252

434 Crit Care Med 2001 Vol. 29, No. 2

44. Wolf M, Keel M, von Siebenthal K, et al:Improved monitoring of preterm infants byFuzzy Logic. Technol Health Care 1996;4:193–201

45. Muller B, Hasman A, Blom JA: Building in-telligent alarm systems by combining math-ematical models and inductive machinelearning techniques. Part 2—Sensitivityanalysis. Int J Biomed Comput 1996; 42:165–179

46. Muller B, Hasman A, Blom JA: Evaluation ofautomatically learned intelligent alarm sys-tems. Comput Methods Programs Biomed1997; 54:209–226

47. Heindl B, Schmidt R, Schmid G, et al: Acase-based consiliarius for therapy recom-

mendation (ICONS): computer-based advicefor calculated antibiotic therapy in intensivecare medicine. Computr Methods ProgramsBiomed 1997; 52:117–127

48. Schmidt R, Pollwein B, Filipovici L, et al:Adaptation and abstraction as steps towardscase-based reasoning in the real medicalworld: case-based selection strategies for an-tibiotics therapy. Medinfo 1995; 8:947–951

49. Schmidt R, Heindl B, Pollwein B, et al: Mul-tiparametric time course prognoses bymeans of case-based reasoning and abstrac-tions of data and time. Med Inf 1997; 22:237–250

50. Gade J, Rosenfalck A, van Gils M, et al:Modelling techniques and their applica-

tion for monitoring in high dependencyenvironments–learning models. ComputMethods Programs Biomed 1996; 51:75–84

51. Nikiforidis GC, Sakellaropoulos GC: Expertsystem support using Bayesian belief net-works in the prognosis of head-injured pa-tients of the ICU. Med Inf 1998; 23:1–18

52. Hanson CW, Marshall C, Medsker C, et al: Agraphical display of clinical cardiopulmonarydata. Abstr. FASEB J 1996; 10(1698):295

53. Powsner SM, Tufte ER: Graphical summaryof patient status. Lancet 1994; 344:386 –389

54. Powsner SM, Tufte ER: Summarizing clinicalpsychiatric data. Psychiatr Serv 1997; 48:1458–1461

435Crit Care Med 2001 Vol. 29, No. 2