Physical PPG signal synthesis via reinforcement learning-based LED control

We introduce a hardware-in-the-loop framework for physical Photoplethysmography (PPG) signal synthesis and benchmarking, leveraging reinforcement learning (RL) as a control engine for the controllable, LED-based reconstruction of target PPG waveforms. Rather than focusing on physiological modeling or proposing a new RL algorithm, this work provides a repeatable testbed for sensor-output waveform reconstruction to facilitate device evaluation. To address the challenges of physical signal reconstruction, we comprehensively analyze how LED stimuli affect the amplitude and temporal characteristics of generated PPG signals and how LED change rates influence reconstruction performance. A standard Proximal Policy Optimization (PPO) agent is used as the underlying controller in this system to adjust LED intensity and control rate so that the sensor’s output matches the target signal. In particular, we design task-specific state representations and latency- and amplitude- aware reward functions that play a critical role in improving performance across waveforms with varying shapes and amplitudes. Evaluated across three representative PPG signals featuring low, moderate, and high amplitudes, our PPO-based control framework consistently reconstructs the target waveforms, outperforming alternative baseline controllers under the same standardized evaluation setup and the current tested conditions. Overall, our framework positions deep RL as a practical complement to classical control for hard-to-model hardware-coupled dynamics, highlighting the promise of data-driven paradigms for nonlinear system identification and control in bioengineering.