A Performance Comparison of LSTM and Recursive SID Methods in Thermal Modeling of Implantable Medical Devices

This paper investigates application of long short-term memory (LSTM) and recursive system identification (RSID) algorithms to predict the thermal dynamics of bio-implants, e.g. UEA under certain assumptions. Both algorithms implemented in this paper predict the temperature readings of heat sensors using a window size of 10 data points. Simulations in COMSOL software as well as experiments using an in vitro experimental systems are utilized for validation and comparison of algorithm performances. Mean squared error (MSE) of prediction results based on the LSTM algorithm is compared against that of the competitive RSID algorithm for evaluation. Both simulation and experimental results indicate that the LSTM can accurately predict the thermal dynamics of the system and outperforms the RSID algorithm when certain conditions for inputs hold. According to COMSOL simulations and in vitro experiments, the LSTM algorithm returns more reliable predictions for the time period in which the convergence of the adaptive filters in the RSID algorithm is not yet achieved. Alternatively, once the adaptive filters converge, the performance of the RSID algorithm is significantly better than the LSTM for some cases due to its adaptive learning capabilities.