16/08/2018 1

Artificial Intelligence, Automatic Control, Operational Research for medicine, life

science and environment.

Alive Strategic Axis

Speaker: Louise Travé-Massuyès

16/08/2018 216/08/2018 2LAAS-CNRS/ Laboratoire d’analyse et d’architecture des systèmes du CNRS 2

Decision and Optimisation in the living and environment fields

> Applied mathematics, computer science, artificial intelligence, automatic control, operations research have an everyday increasing role to play § More and more technology§ More and more data

Rationale: Develop constructive theoretical results and efficient computational algorithms for proposing control, supervision, diagnosis and optimization solutions

> Develop novel, personalized intervention methods for prevention, treatments, and drug administration

> Provide Machine Learning models to support predictions on the future development of diseases and also predicting the responses to specific therapies and preventive measures

> Identify novel patterns and health correlations in complex data sets> Design AI-assisted diagnosis to help physicians decide about the relevant care at

early stages of disease detection

Some examples:

16/08/2018 316/08/2018 3LAAS-CNRS/ Laboratoire d’analyse et d’architecture des systèmes du CNRS 3

Automatic feedback-control of general anesthesia during surgery

Context : From anaesthesiologist-based open-loop control to automated closed-loop control (Multi-)objective:

- Medical point of view§ Fast induction§ Maintenance § Avoid drug overshoot§ Adapt to patient

characteristics

- Control theory point of view§ Positive system§ Slow/fast dynamics§ Patient uncertainty§ Actuator saturation§ Quantized sensor signal§ (aperiodic) sampling

Involved: I. QUEINNEC, S. TARBOURIECH, G. GARCIA (LAAS), M. MAZEROLLES (CHU Rangueil)

EEG

Bolus alloc function

(freq, conc, duration)

Positivity

Constraint continuous patient model

Pharmaco−kinetics

pro

po

fol

and

rem

ifen

tan

il

discrete/continuous

controllerStimulation

quantizerPupillarydilatation

flo

wra

tes

y1(t)

y2(k)

acquisitionconverterBIS

u0

u0

0−u

0−u u

sat(u)

0−u

u0

sat(u)−u

u

q(x)

x

Introduction Actuators and saturation Communication and sampling Conclusion

and a discrete-time dynamic output feedback controller

... interconnected with a discrete-time dynamic output feedback controller(

xc(tk+1) = Acxc(tk) + Bcy(tk) + Ec (u(tk))

u(tk) = Ccxc(tk) + Dcy(tk), 8k 2 N (10)

xc 2 Rnc , y 2 Rp and u 2 Rm are the state, the input and the output ofthe controller, respectively.

Ac , Bc , Cc , Dc , Ec are assumed to be constant and of appropriatedimensions.

Ec denotes an static anti-windup gain, and (u(tk)) is a vector-valueddecentralized dead-zone nonlinearity defined as follows:

(u(tk)) = sat(u(tk))� u(tk). (11)

26 / 34

16/08/2018 416/08/2018 4LAAS-CNRS/ Laboratoire d’analyse et d’architecture des systèmes du CNRS 4

Development of neural networks for 2D - quality control during radiotherapy treatment (Project DIVIN)

Context : External radiotherapy treatment.

(Electronic portal imaging

devEPIDice) è Now, not used for improving the control before (quality

control of the machine) or during the

treatment

è If used, possibility of determination- Absorbed dose truly received by the

patient during all treatment sessions

- Re-computing the treatment

plans in case of discrepancies

Objective : 2D(3D) absorbed dose reconstruction

è spatial configuration to reproduce spatial diffusion

inputs

outputs

Learning phase =

treatment planning

system (TPS)

absorbed dose

distribution considered

like the truth

EPID imageReconstructed

absorbed dose

(ANN)

Planned

absorbed dose

(TPS): target

Example : Intensity-modulated radiotherapy (10 iI/O datasets)

Detection of leaf mis-position

Comparison of gindex

Involved : F. CHATRIE (PhD, LAAS), MV LE LANN( LAAS/DISCO)

X. FRANCERIES (ONCOPOLE)

16/08/2018 516/08/2018 5LAAS-CNRS/ Laboratoire d’analyse et d’architecture des systèmes du CNRS 5J-C. Billaut – EURO, Vilnius, July 2012

Combinatorial Optimization for the analysis of big amounts of biomedical data

Involved : J. MONCEL (LAAS), N. BRAUNER, J. DARLEY (G-SCOP)

> Big data to be analyzed for biomedical diagnosis> Goal:

§ Characterize two classes: patients with pathology and patients without § Provide easily interpretable rules of the form IF condition THEN class 0/1§ Handle discrete and continuous variables

> Minimize the number of cutting points to binarize continuous variables> Find the support of observations (patterns covering all the set of clinical cases):

set cover problem> Reduce the support if too big: graph partitioning based on graph density to

identify groups of patterns and select one single pattern to represent all the group

16/08/2018 6

Artificial Intelligence, Automatic Control, Operational Research for medicine, life

science and environment.

Alive Strategic Axis

Speaker: Louise Travé-Massuyès