# Prediction and accuracy improvement of insulin pump in-fusion deviation based on LSTM and PID

## TL;DR

This paper improves the accuracy of an insulin pump by optimizing its structure and using LSTM and PID algorithms to reduce infusion deviations.

## Contribution

A novel combination of BP-PID control and LSTM prediction is used to enhance insulin pump accuracy.

## Key findings

- The minimum effective single infusion dose was reduced by 50.52% to 0.047 U.
- LSTM prediction reduced mean infusion deviation for large doses by 12.05%.
- The mean deviation decreased by 12.85% after 24h low basal rate injection.

## Abstract

In order to further improve the injection precision of the PH300 insulin pump, this paper optimizes and improves the mechanical structure and control algorithm of the PH300. The improved PH300 uses a proportional-integral-derivative controller based on back propagation neural network (BP-PID) algorithm to control operation, and the experimental results show that the minimum effective single infusion dose of the improved PH300 is 0.047 U, which is reduced by 50.52%. The deviation reduction of low-dose infusion (0.1U-0.9U) ranged from 1.47% to 10.87%, with a mean of 4.91%. The mean deviation of the improved PH300 decreases by 12.85% after a 24h low basal rate (0.5U/h) injection. In addition, Long Short-Term Memory (LSTM) was used to predict the deviation during injection, and the predicted values were uniformly compensated for in subsequent injection experiments. The LSTM model performed best with a training set of 85%, a test set of 15%, an epoch of 300, a batch number of 256, and 32 hidden layer neurons. After compensation, the mean infusion deviation for large doses was reduced by 12.05%, and the maximum deviation by 14.12%.