Virtual PPG reconstruction from accelerometer data via adaptive denoising and cross-Modal fusion

Abstract

Accurate heart rate (HR) monitoring during high-intensity activity is essential for performance optimization and physiological tracking in wearable devices. While photoplethysmography (PPG) remains the standard for HR estimation, it is prone to motion artifacts, power constraints, and temporary signal loss. Accelerometers (ACC), by contrast, offer motion-resilient and energy-efficient sensing, but estimating HR from ACC alone remains a challenging task. In this study, we introduce a cross-modal virtual sensing framework for HR estimation and spectral reconstruction using only ACC signals. The framework includes: (1) a high-fidelity variational autoencoder (VAE) for offline PPG spectrum reconstruction from ACC input, and (2) a lightweight real-time attention-based denoising model for HR prediction. Both models are trained with a fusion-aware loss to enforce alignment between motion-driven and cardiovascular signal features. Experimental results on public and proprietary datasets demonstrate strong performance and generalization under varying sensor configurations and motion conditions. The real-time model achieves 7.0 BPM mean absolute error (MAE) with only 2.6K parameters, making it suitable for embedded deployment. While PPG remains superior under ideal conditions, the proposed system serves as a fallback modality when optical sensing is unreliable or unavailable-enabling gap-filling, post-processing correction, and low-power monitoring. More broadly, this work positions virtual PPG reconstruction as a proof-of-concept for physiological virtual sensing: a paradigm where one modality can be inferred from another, and potentially reversed, supporting robust multimodal inference in real-world mobile health scenarios.