Towards fair and trustworthy heart rate estimation from wrist-worn photoplethysmography: a multi-wavelength dataset and uncertainty-aware deep learning approach evaluated across skin tones, sexes, and motion conditions

Abstract Objective. To improve fairness (reduced disparities across skin tones and sexes) and trust (well-calibrated uncertainty metrics that indicate unreliable predictions) in wrist-worn photoplethysmography (PPG) heart rate estimation. While PPG offers a convenient, low-cost method for continuous heart rate monitoring, accuracy is often degraded by motion artefacts and demographic variation. Approach. We collected a multi-wavelength (blue, green, red, IR) wrist-worn PPG dataset covering diverse activities, skin tones, and sexes, and developed a multi-branch PPG-accelerometer fusion convolutional neural network quantifying both aleatoric and epistemic uncertainty (the latter via Monte Carlo dropout, concrete dropout, and deep ensembles). We compared wavelength sets, quantified demographic effects, and applied uncertainty-aware post-processing to filter out high-uncertainty predictions. Evaluation used leave-one-subject-out cross validation across both the collected multi-wavelength dataset and five commonly used single-wavelength datasets. Main results. Multi-wavelength fusion improved accuracy under motion, with blue-green-red-IR achieving the lowest mean absolute error (MAE). Baseline error was higher for participants with darker skin tones and for females; uncertainty-guided rejection reduced MAE and effectively eliminated these gaps (reducing parity gaps from 1.2 BPM to 0.0 BPM for skin tone). Concrete dropout provided the best calibration-accuracy-efficiency trade-off, achieving the lowest average miscalibration area (MA = 0.12 ) while maintaining competitive MAE. Significance. Combining calibrated uncertainty with group-aware evaluation yields a more fair and trustworthy wrist-worn PPG heart rate estimation pipeline, supporting responsible deployment in consumer and sports contexts and informing bias-aware clinical development.