Cardiovascular Risk Assessment via Sleep Patterns and ECG-Based Biological Age Estimation

Figures

• 

  Figure 1. Age regression model architecture: Time-series data from nocturnal cardiac clusters and sleep-stage sequences are first embedded into a learnable space. These embeddings are processed through parallel convolutional blocks with varying kernel sizes to capture multi-scale temporal patterns. The resulting features are refined by neural network layers, enabling the model to predict biological age as a continuous value.

• 

  Figure 2. Construction of the heart cluster vectors: ECG signals are segmented using a sliding window and processed through a self-supervised learning framework, resulting in compact feature vectors representing cardiac activity. These vectors are then clustered using K-means clustering to identify distinct patterns of cardiac behavior. Finally, a temporal vector reflecting the sequence of clustered states is constructed, capturing the dynamic changes in heart rhythms across the entire sleep period.

• 

  Figure 3. Correlation plot between the biological age predicted by the heart cluster model and the clusterXsleep model showing the high correlation between the two predictions.

• 

  Figure 4. Mean difference across groups for all three experiments for cardiovascular diseases (CVDs).

• 

  Figure 5. Kaplan–Meier survival curves illustrating the probability of survival over time for individuals who developed heart failure, stroke, or myocardial infarction within the 5-year follow-up period. The healthy cohort is shown as a reference group. Shaded regions represent 95% confidence intervals.

• 

  Figure 6. Performance comparison of cardiovascular disease (CVD) risk classification models. The figure demonstrates how the addition of the biological age estimate to the traditional Framingham risk factors leads to improved metrics in both logistic regression and random forest classifier models.