Abstract Body (Do not enter title and authors here): Background: Continuous monitoring of heart rate (HR), heart rate variability (HRV), and respiratory rate (RR) provides valuable insight into autonomic and cardiovascular function. Photoplethysmography (PPG), widely available in wearables, enables large-scale, continuous tracking of these biomarkers in free-living conditions.

Objectives: This study aimed to:
- Develop and validate a lightweight, multi-task neural network for simultaneous estimation of HR, HRV (rMSSD), and RR from wearable PPG.
- Apply the model to a large wearable cohort to investigate associations between estimated physiological metrics and self-reported cardiovascular disease status.

Methods: A lightweight multi-task neural network combining convolutional and recurrent layers was trained to simultaneously estimate HR, HRV, and RR from 30-second PPG segments collected with the Oura Ring. Training data consisted of simultaneous ring PPG and polysomnography (PSG) data from 617 participants enrolled in a PSG sleep study. Ground-truth labels were derived from ECG (for HR, HRV) and respiratory inductance plethysmography (for RR). To assess clinical relevance, the model was applied to ~320,000 de-identified Oura Ring users with self-reported cardiovascular disease (CVD) status. Resting HR, HRV, and RR were compared between individuals with CVD and age- and sex-matched controls.

Results: On the independent test set (97,147 segments), the 76k-parameter multi-task model achieved coefficients of determination (R²) of 0.990, 0.890, and 0.672 with mean absolute errors (MAE) of 0.46 bpm, 4.06 ms, and 0.94 brpm for HR, HRV, and RR, respectively. Aggregation over 5-minute epochs improved R² to 0.995 (HR), 0.950 (HRV), and 0.793 (RR) with respect to PSG reference devices. Application of the model to PPG recordings from the cohort with known (self-report) CVD status revealed higher resting HR and RR, and lower HRV in individuals with CVD compared with controls; with variable effects depending on the type of CVD (e.g., participants with arrhythmia showed higher HRV than controls).

Conclusion: This study demonstrates that multi-task deep learning enables accurate, simultaneous estimation of cardiorespiratory biomarkers from wearable PPG. When applied at scale, these models reveal physiological signatures associated with self-reported cardiovascular disease, supporting the use of wearables for population-level risk stratification and long-term monitoring.