Heart disease detection using an acceleration-deceleration curve-based neural network with consumer-grade smartwatch...

Area of Science:

• Biomedical Engineering
• Machine Learning in Healthcare
• Cardiovascular Medicine

Background:

• Cardiovascular disease (CVD) is a leading global cause of death and disability.
• Continuous monitoring using consumer smartwatches offers a novel approach for CVD detection.
• Analyzing large time-series data from wearables presents challenges due to sparse informative segments.

Purpose of the Study:

• To evaluate an acceleration-deceleration curve-based machine learning (ML) model for detecting cardiovascular diseases.
• To investigate the efficacy of novel data preprocessing and model aggregation techniques for CVD prediction.

Main Methods:

• Utilized data from the ME-TIME study (42 participants: 21 with CVD, 21 healthy controls).
• Applied per-subject normalization by peak inactivity curve values.
• Employed a neural network model with weekly prediction aggregation and contrastive loss.

Main Results:

• The model achieved 99% specificity and 40% sensitivity on the development set.
• The model demonstrated 100% specificity and 67% sensitivity on the test set.
• Acceleration-deceleration curves effectively ruled out CVD presence when data was properly processed.

Conclusions:

• Acceleration-deceleration curves are valuable for excluding cardiovascular disease.
• Careful preprocessing of curves and appropriate model selection are crucial for reducing variability and improving predictive accuracy.