Enhancing the Accuracy of Low-Cost Pulse Oximetry via Edge AI: A Neural Network Approach Optimized for STM32 Microcontrollers

Abstract

Although pulse oximeters are widely used in clinical monitoring, low-cost sensors often exhibit reduced accuracy under altered perfusion and hypoxic conditions. This study proposes a deep learning-based Edge AI framework deployable on STM32 microcontrollers to enhance measurement reliability. Using the open-access OpenOximetry dataset, a neural network was trained with pulse-oximeter oxygen saturation (SpO2), Perfusion Index (PI), and Heart Rate (HR) as inputs, taking arterial oxygen saturation ( SaO2 ) as the reference standard. Evaluation on 3,918 independent test samples achieved a Mean Absolute Error (MAE) of 1.48% and R2 of 0.843, with a Bland-Altman mean bias of 0.16%. Post-training optimization using TensorFlow Lite reduced the model size to approximately 20 KB while maintaining inference latency below 2 ms. The results demonstrate that embedded deep learning can significantly improve low-cost pulse oximetry accuracy while enabling real-time, resource-efficient wearable monitoring.