Smart seat cushion mobile application with on-device posture prediction using TensorFlow lite

Pressure injuries (PI) pose a significant risk for individuals with spinal cord injuries. While clinical guidelines recommend periodic pressure redistribution (PR), adherence is often low due to limited real-time monitoring and feedback. In this paper, we present an Android application, integrated with a machine learning-based posture prediction algorithm to enhance real-time monitoring and feedback in a smart seat cushion (SSC) system for wheelchair users. Data from 12 healthy non-wheelchair participants in nine seating postures were collected. Five deep leaning architectures - Multi-Layer Perceptron (MLP), Convolutional Neural Network (CNN), Long Short-Term Memory (LSTM), CNN-LSTM, and Multi-Headed Attention models were trained, and their test performances were compared. An Android application was then developed with Flutter for on-device deployment. The highest performing model (LSTM) was then integrated using TensorFlow Lite to enable real-time posture prediction. We found that LSTM gives an accuracy of 92%, outperforming the other architectures. Also, the Android app was tested on a Google Pixel tablet, which can successfully control seat cushion operations wirelessly, identify user's seating postures, visualize live pressure maps, generate statistics of user's seating habits and weight shifting maneuvers, as well as provide guidance during pressure relief protocols to improve adherence. The proposed system provides a solution to low adherence to weight shift protocols observed in other studies by providing a live pressure map view and real-time feedback, thereby promoting consistent PR practice. This innovation represents a significant advancement in the prevention of PI and supports improved user compliance with clinical guidelines.