Abstract Body (Do not enter title and authors here): Background: Wearable photoplethysmography (PPG) offers a noninvasive means to continuously monitor cardiovascular signals and holds potential for the early identification of cardiovascular diseases.

Objectives: This study investigated:
- whether ring PPG embeddings from a self-supervised model can detect self-reported hypertension, coronary/peripheral artery disease, and type 2 diabetes.
- whether disease predictions are sensitive to longitudinal physiological changes after blood-pressure-lowering treatment initiation.

Methods: A neural network encoder was trained via self-supervised contrastive learning on more than 2 million 30-second PPG segments collected during sleep from ~230,000 Oura Ring users. Logistic regression models were then trained on top of the embeddings generated by the foundation model for downstream classification of cardiovascular diseases. The final classification dataset included self-reported disease status from ~320,000 de-identified Oura Ring users. A third of participants were held out for testing, and PPG embeddings from 7 nights were aggregated at the participant-level and used as inputs to the classifier. For each target condition, the area under the receiver operating characteristic curve (AUC) was computed on the test set, and compared to the AUC of a baseline classifier using age, sex, and BMI. Longitudinal changes in estimated probabilities of hypertension were analyzed for 605 hypertensives over a period of 22 weeks around the self-reported onset of blood-pressure-lowering treatment.

Results: On the test set, embedding-based classifiers achieved AUCs of 0.837 for hypertension, 0.885 for coronary and peripheral artery disease, and 0.891 for type 2 diabetes, which represents an average improvement in AUC of +0.029 compared to the baseline demographics model. Across all targets, false positive rates were significantly higher for users with self-reported comorbidities (e.g., obesity or sleep apnea). In the longitudinal analysis, initiating blood-pressure-lowering medication was associated with a statistically significant decrease in hypertension predictions.

Conclusion: Large-scale self-supervised models trained on wearable PPG data encode relevant features of cardiovascular health. When paired with simple classifiers, these embeddings facilitate preliminary detection of cardiovascular and metabolic conditions. Our findings support the utility of wearables as scalable digital biomarkers for early disease detection and monitoring.