Data Quality Degradation on Prediction Models Generated From Continuous Activity and Heart Rate Monitoring:...

Area of Science:

• Physiological monitoring
• Machine learning in healthcare
• Wearable technology

Background:

• Consumer wearable data accuracy concerns limit clinical integration.
• The impact of data degradation on predictive models is understudied.

Purpose of the Study:

• Simulate data degradation effects on prediction model reliability.
• Assess limitations of lower wearable device accuracy in clinical settings.

Main Methods:

• Trained a random forest model on step count and heart rate data.
• Evaluated model performance across 75 perturbed datasets (missingness, noise, bias).
• Compared perturbed data performance to the unperturbed dataset.

Main Results:

• Model performance (RMSE) was stable up to 20-30% data perturbation.
• Predictive capability was lost at 80% noise, 50% missingness, 35% combined perturbations.
• Systematic bias did not affect model RMSE.

Conclusions:

• Predictive models for cardiac competence are robust to declining data quality.
• Lower accuracy from consumer wearables may not preclude their use in clinical prediction.