AVAESA: Adaptive VAE With Self-attention and Learnable Signal Processing for Robust Radar-based Heart Rate Estimation

Abstract

Non-contact heart rate monitoring using radar sensors offers significant advantages for healthcare and automotive applications by preserving privacy while enabling continuous physiological assessment. Current Variational Autoencoder (VAE) approaches for radar-based vital sign monitoring, while superior to traditional neural networks, suffer from fixed preprocessing assumptions and inadequate temporal modeling that limit their generalization across diverse measurement conditions. This study introduces AVAESA (Adaptive VAE with Self-Attention and Learnable Signal Processing), a novel architecture that addresses these limitations through three key innovations: dual-stream in-phase/quadrature signal processing that preserves critical phase relationships, multi-head self-attention mechanisms for enhanced temporal dependency modeling, and adaptive signal preprocessing with learnable parameters that derive frequency bands and processing weights directly from input signal characteristics. The framework was evaluated on 1920 measurements from 10 participants across 48 measurement scenarios (4 distances × 3 angles × 4 orientations), assessing cross-scenario robustness under varied measurement conditions, with Polar H10 chest strap ground truth validation. Comprehensive comparison against multiple architectures (CNN, LSTM, Bi-LSTM, TCN, VAE) with statistical significance testing demonstrates substantial performance improvements, with mean absolute error reductions ranging from 17.3% under optimal conditions to 62.6% under challenging cross-scenario generalization scenarios. AVAESA maintains high accuracy (correlation coefficient > 0.86, R² > 0.84) even under challenging measurement conditions where baseline approaches exhibit degraded performance, demonstrating potential for contactless cardiac monitoring systems across diverse measurement environments through improved cross-scenario robustness.