Accessible FES Cycling: Leveraging Cost-Effective Crank Angle Measurement and Reinforcement Learning for Stimulation Optimization

The effectiveness of FES cycling in providing health benefits depends on the extent of training. Challenges arise for at-home training due to difficulties in determining the stimulation pattern, often requiring clinicians' assistance. Additionally, the cost can impede access to FES cycling equipment. Here, we present our two developments that help address the challenges in performing at-home training. The first development is a low-cost device for measuring the crank angle that combines a data-efficient technique to correct the error and a neural network to perform a real-time extended Kalman filter. The second development is the application of reinforcement learning and action branching architecture, denoted as QBR, that can efficiently learn the stimulation pattern. The complete framework enables a simple cycling exerciser to be used for FES cycling. We demonstrate and evaluate our framework on a simple cycling setup. Our device can give the measurement with an error below 5°, and QBR learns roughly 3 times faster than the method introduced in our previous work. In combination, our system enables FES cycling on the setup. These outcomes enable the potential of our proposed framework to improve the accessibility of FES cycling, especially for at-home training.