Predicting cardiovascular disease risk using photoplethysmography and deep learning

Abstract

Cardiovascular diseases (CVDs) are responsible for a large proportion of premature deaths in low- and middle-income countries. Early CVD detection and intervention is critical in these populations, yet many existing CVD risk scores require a physical examination or lab measurements, which can be challenging in such health systems due to limited accessibility. We investigated the potential to use photoplethysmography (PPG), a sensing technology available on most smartphones that can potentially enable large-scale screening at low cost, for CVD risk prediction. We developed a deep learning PPG-based CVD risk score (DLS) to predict the probability of having major adverse cardiovascular events (MACE: non-fatal myocardial infarction, stroke, and cardiovascular death) within ten years, given only age, sex, smoking status and PPG as predictors. We compare the DLS with the office-based refit-WHO score, which adopts the shared predictors from WHO and Globorisk scores (age, sex, smoking status, height, weight and systolic blood pressure) but refitted on the UK Biobank (UKB) cohort. All models were trained on a development dataset (141,509 participants) and evaluated on a geographically separate test (54,856 participants) dataset, both from UKB. DLS's C-statistic (71.1%, 95% CI 69.9-72.4) is non-inferior to office-based refit-WHO score (70.9%, 95% CI 69.7-72.2; non-inferiority margin of 2.5%, p<0.01) in the test dataset. The calibration of the DLS is satisfactory, with a 1.8% mean absolute calibration error. Adding DLS features to the office-based score increases the C-statistic by 1.0% (95% CI 0.6-1.4). DLS predicts ten-year MACE risk comparable with the office-based refit-WHO score. Interpretability analyses suggest that the DLS-extracted features are related to PPG waveform morphology and are independent of heart rate. Our study provides a proof-of-concept and suggests the potential of a PPG-based approach strategies for community-based primary prevention in resource-limited regions.

Fig 1. Summary of study motivation and design.

The motivation of applying the PPG-based cardiovascular disease (CVD) risk assessment in the low-resource health systems. Non-office based information acquired from mobile-sensing technologies may help address the burden of cardiovascular disease risk screening and triage in resource-limited areas. In this study, we compare our developed model (DLS) with the existing office-based and lab-based CVD risk scores that have been developed for low-resource medical settings, including models refitting variables in the WHO and Globorisk scores, and office-based and lab-based Globorisk scores recalibrated on the same study cohort.

Fig 2. Flowchart of inclusion and exclusion of the cohort for developing the survival model.

Fig 3. Kaplan-Meier curves for the DLS with different definitions of high risk.

(A) Risk threshold corresponding to a specificity of 63.6%, (B) Risk threshold corresponding to a sensitivity of 55.4% (see Methods). The p-values were calculated by the log-rank test.

Fig 4. Calibration plot, showing observed and predicted 10-year MACE risk.

We discretized each model’s output into deciles and the slopes indicate the coefficient of a linear regression. METADATA: the risk model with age, sex, smoking status), OFFICE: the risk model using shared predictors from the office-based WHO/ISH risk chart and Globorisk score. DLS: the risk model with metadata and deep learning-based PPG features. DLS+: the risk model using all DLS predictors plus BMI. DLS++: the risk model using all DLS predictors plus BMI and systolic blood pressure.

Fig 5. DLS PPG features explainability visualizations.

The first row sorts PPGs by predicted risk, whereas the next 5 rows sort PPGs based on the 5 PPG feature values. The first column presents the average of 100 PPG waveforms sampled nearest to the following quantiles of the quantity mentioned on the left: 10th (red) 50th (green) and 90th (blue). The next 3 columns present the respective averaged PPGs along with the normalized saliency values based on integrated gradients. We observed that in general salient areas that most influence the predictions seem to be near the top of the systolic peak and the notch, independent of which quantile and feature/prediction the PPG was sampled from. Each PPG feature appears to correspond to different morphological aspects (see Table 4).