Simulation assessment of a closed-loop controller designed by machine learning techniques.
O. Caelen*, G. Bontempi*, F. Clément§, E. Coussaert#, L. Barvais#
*Machine Learning Group, Computer Science, Université Libre de Bruxelles, Belgium; §Mexys-SA, Mons, Belgium #Erasmus Hospital Anaesthesia Departments, Université Libre de Bruxelles, Belgium;
Background and Goal of the Study:
BIS guided TCI anaesthesia archives were used to design a BIS-Propofol closed-loop controller by machine learning algorithms. The approach allows adaptive control capabilities and combines the information of several input variables measuring the patient condition and the operation state. The purpose of the study was to compare, in a simulation setting, the outputs of our controller with an expert-based closed-loop controller currently used in real conditions with acceptable performances. This expert-based closed-loop controller titrates Propofol and Remifentanil automatically according to a predefined range of BIS and hemodynamic values.
Materials and Methods:
Data collected on 965 chirurgical interventions were used to search the best values of the closed-loop parameters. To propose the best target of Propofol for controlling the BIS, our closed-loop uses the following variables: the current BIS value, the current target of Propofol, the current target of Remifentanil and the age and the weight of the patient. A Lazy Learning algorithm [1] was used as machine learning method. The control action proposed by our closed-loop controller was compared to the behavior of an expert-based closed-loop controller in 18 real archived TCI anaesthesias. Let PE (predictive error) be equal to the target of Propofol (μg/ml) proposed by the existing closed-loop minus the target of Propofol (μg/ml) proposed by our new closed-loop. To assess the actions of our new closed-loop, the following measures are computed: MDPE (median of the predictive error), MDAPE (median of the absolute predictive error) and the NMSE (normalized mean squared error).
Results:
MDPE |
MDAPE |
NMSE |
-6.1% |
10.1% |
0.19 |
Discussions:
The Propofol titration behavior of the two controllers is significantly different (P<<0.01) but the small value of the MDAPE means that this difference is reasonable. The average of the BIS values, when controlled by the existing closed-loop, is known to be sometimes too low and the negative sign of the MDPE is thus a promising result. The NMSE is a well known positive statistical measure of the prediction error where the value of one indicates the performance of the simplest prediction model and the small value of the NMSE is thus also a hopeful result.
Conclusion(s):
The simulation tests appear to be promising and the next step of this study will be a test in real conditions.
Reference:
[1] M. Birattari, al. (1999). Lazy learning meets the recursive least squares algorithm. Advances in Neural Information Processing Systems 11, MIT Press, Cambridge, MA, pp. 375-381.