Online Implementation of LSTM for Thermal Management of Implantable Medical Device

As application of medical implant expand in complexity and position of usage, thermal regulation become a major issue with the device power expansion which led to elevating thermal energy. With brain implant, even a slight increase in device temperature may cause irreversible damage to the brain cell. Therefore, the goal of our project is to implement an online version of the LSTM system, building of a off-line LSTM system, with control scheme that can predict the temperature rise and regulate the device when temperature rises above threshold. Long Short-term memory system (LSTM), with use of the past data and prediction along with the forgetting mechanism to allocate computational resources to more valuable data, yield a high prediction accuracy while keeping the computational cost low. The COMSOL set up for the project contains six temperature sensors and two input sources, which are used for one-step-prediction and control scheme testing. To achieve low MSE with low computational cost, additional parameters and weight factor are included into the existing modules, while control scheme is introduced and tested to enable response from bioimplant when prediction reach above a defined threshold. In particular, a suspicion ratio function is implemented into the LSTM code which incorporates both a 1D suspicion ratio array for W[x-10] sliding window and adaptive suspicion point replacement into the training algorithm. To obtain suspicion ratio array, a sliding windows with length W[x-10] is generated and input into a normal distribution function to identify outliers and their associated suspicion ratio. The extrapolated information is then feed into a logistic sigmoid function to normalize its weight on the training model and tested via the different datasets. The resulting MSE will be calculated for each sensor reading and prediction. In addition, for the outliers identified, suspicion ratio prior to the point is extracted and rated for its frequency to determine whether the data point need to be dropped or replaced to prevent undesired fluctuation in the weight gradient curve. A matrix will also be added to the input and output arrays to incorporate the covariance parameters into the algorithm, while a non-linear predictive control is implemented onto the LSTM via SQP optimization. In summary, a low computational cost online LSTM system will be achieved through additional training parameters and control scheme which aim to make temperature prediction and control medical bioimplant to prevent potential tissue damages. The result will provide data for future analysis and improvement, giving researchers directions for possible optimization through the control of training factors of the model.