INGAR Instituto de Desarrollo y Diseño Consejo Nacional de Investigaciones Científicas y Técnicas Universidad Tecnológica Nacional Anomaly Detection in the Artificial Pancreas Luis Ávila, Ernesto Martínez INGAR (CONICET-UTN), Avellaneda 3657, Santa Fe S3002 GJC, Argentina {avilaol, ecmarti}@santafe-conicet.gob.ar 18 th September 2013 CAIS 2013-4º Congreso Argentino de Informática y Salud 18 al 20 de Septiembre de 2013 Universidad Nacional de Córdoba
Diabetes mellitus Artificial pancreas Faulty conditions Reference behavior Diabetes mellitus disease resulting from the impaired mechanism of insulin secretion from the pancreas In Type 1 Diabetes β-cells in the pancreas do not produce insulin (Insulin-dependent). In Type 2 Diabetes the cells in body do not utilize insulin properly (Insulin-independent). CAIS 2013-4º Congreso Argentino de Informática y Salud 2/16
Diabetes mellitus Artificial pancreas Faulty conditions Reference behavior Artificial pancreas A safety-critical design of the artificial pancreas should guarantee fast detection of performance degradation! CAIS 2013-4º Congreso Argentino de Informática y Salud 3/16
Diabetes mellitus Artificial pancreas Faulty conditions Reference behavior Faults and surprising events Unstable signal output, interference or poor calibration Dietary change, exercise, stress or pregnancy Changes due to constraint activation or tuning of control algorithm Obstruction of the infusion catheter or mechanical failures These drawbacks make clear the necessity for supervision of the control loop and early detection of performance degradation CAIS 2013-4º Congreso Argentino de Informática y Salud 4/16
Diabetes mellitus Artificial pancreas Faulty conditions Specified behavior Specified (optimal) behavior Detection of performance degradation in the regulation loop can be achieved by defining an optimal specified behavior and comparing it with real measurements obtained through Continuous Glucose Monitoring. Assuming an optimal control policy is used to control the glucoseinsulin dynamics over time, a reference behavior allows making a conjecture on expected BG levels. Significant deviationsfrom optimal behavior are evidence of changes in properly functioning of one or more components of the AP. CAIS 2013-4º Congreso Argentino de Informática y Salud 5/16
Patient model Error Grid Analysis Patient (stochastic) model Bergman s two-compartment model plus Ito s stochastic process to capture the patient variability using the variance parameter. ( ) + (, ) ( + )( + ) dg Gin t NHGB p3 I G p2ia p5 KM GX Gren = + dt V G ( K + G) V V G X M G G σε dt CGM sensors record values every 5 minutes Stochastic samples generated with an Ito s process CAIS 2013-4º Congreso Argentino de Informática y Salud 6/16
Bayesian surprise Error Grid Analysis Error Grid Analysis An AP should be regularly tested to assess on-line if it is performing well. Many tools have been proposed to gauge accuracy of continuous glucose sensors. Clarke: accounts for the distance of the current method to a reference and gives significance to deviations from the desired behavior. Consensus: based on the expertise of a large panel of clinicians to solve the problem of discontinuities presented by Clarke EGA. Continuous: evaluate not only sensor accuracy in terms of correct presentation of BGL but also the direction and rate of BG fluctuations. CAIS 2013-4º Congreso Argentino de Informática y Salud 7/16
Clarke s grid Clarke s grid PID and variability Closed-Loop EGA CL-EGA: variability CL-EGA: miscalibration It uses a Cartesian diagram where values estimated by the technique under scrutiny are displayed against the values provided by the reference method The diagonal line represents perfect agreement between the two readings. Points below and above the 45º line indicate over and underestimation of BG values. Region A: accuracy with respect to the reference. Region B: benign deviations. Region C: inadequate treatment. Region D: failure to detect hypo- or hyperglycemia. Region E: confusion of needed treatment. CAIS 2013-4º Congreso Argentino de Informática y Salud 8/16
Clarke s grid PID and variability Closed-Loop EGA CL-EGA: variability CL-EGA: miscalibration PID and variability Specified vs PID σ=0.10 Specified vs PID σ=0.50 CAIS 2013-4º Congreso Argentino de Informática y Salud 9/16
Closed-Loop EGA Avoiding hypoglycemia is the most important issue! Clarke s grid PID and variability Closed-Loop EGA CL-EGA: variability CL-EGA: miscalibration Through adaptation, the Consensusgrid can be used to assess an AP performance. y-axis:bgl when the loop behaves optimally. x-axis:bgl for current dynamics. Optimal control should guarantee normoglycemia, reference is limited between 70 and 180 mg/dl. CAIS 2013-4º Congreso Argentino de Informática y Salud 10/16
Clarke s grid PID and variability Closed-Loop EGA CL-EGA: variability CL-EGA: miscalibration CL-EGA: variability The modified CG-EGA assess pointand rateaccuracy Specified vs GPRL Closed loop performance for variability parameter σ=0.50. GPRL PID Specified vs PID CAIS 2013-4º Congreso Argentino de Informática y Salud 11/16
Clarke s grid PID and variability Closed-Loop EGA CL-EGA: variability CL-EGA: miscalibration CL-EGA: miscalibration Assessment of closed loop performance using calibration parameter ξ=10%. Specified vs GPRL Specified vs PID CAIS 2013-4º Congreso Argentino de Informática y Salud 12/16
Specified behavior Bayesian Surprise Specified behavior Developing an ability to rapidly detect surprising events is crucial in allowing safety critical systems to quickly identify potential risks and dangerous situations. The amount of surprise in new data from an AP by looking at the changes that take place in going from the specifiedto the current loop functioning CAIS 2013-4º Congreso Argentino de Informática y Salud 13/16
Specified behavior Bayesian Surprise Bayesian Surprise Any deviation from specified operation accounts for a surprising event in the closed-loop functioning Assessment of variability σ=0.50. GPRL PID Fuzzy. P KL Surprisedescribes pointwise differences between the optimal and the current loop operation T KL Surprisedescribes a progressive shift in the closed-loop dynamics CAIS 2013-4º Congreso Argentino de Informática y Salud 14/16
Summary Questions Summary Error Grid Analysis could be reused as a performance monitoring tool to detect glycemic variability and faulty conditions in an AP. We emphasize the importance of using well established methods in clinical practice to assess performance degradation. Our aim is to develop a generic on-line performance monitoring tool for the artificial pancreas that accounts for glycemic variability and faults using Bayesian inference. CAIS 2013-4º Congreso Argentino de Informática y Salud 15/16
Summary Questions Questions CAIS 2013-4º Congreso Argentino de Informática y Salud