Model building and LSTM-based system identification for implantable devices

Abstract

Implantable medical devices (IMDs) have aroused a wide research interest because of its increased ability of monitoring and recording signals from human organs and tissues. There are numerous issues that are under researching, one of which is the thermal control problem. Human body has a certain tolerance to high temperature (such as fever). However, long-term accumulation of the thermal effect if the temperature is higher than the safe limit leads to detrimental effect. Thus efficient thermal management methods are of significant importance for IMDs. Before designing controllers, the first aim is to build a reliable model for the thermal effects of the IMDs. In this thesis, three different system identification methods are explored for modeling thermal effects and their advantages and disadvantages are compared. First, a COMSOL model considering all the thermal effects of an IMD system is built. Then, a long short-term memory (LSTM) network is designed to predict the thermal dynamics of the IMDs both online and offline. For the validation of the LSTM algorithms, both COMSOL simulations and experiments are studied. The performance of the LSTM method is compared with that of a recursive predictor based subspace identification (RPBSID) method considering time complexity and prediction accuracy. The results based on COMSOL simulations indicate that the online LSTM algorithm outperforms the RPBSID algorithm in general except its higher computational cost. The offline LSTM algorithm has superiority for the time period when the convergence of the adaptive filters in the RPBSID algorithm is not yet achieved. Additionally, the results based on \textit{in vitro} experiments show that both online LSTM and offline LSTM triumph over RPBSID based on the metric - best fit rate (BFR).