Toward characterizing cardiovascular fitness using machine learning based on unobtrusive data.

Area of Science:

• Sports Medicine
• Biomedical Engineering
• Data Science

Background:

• Cardiopulmonary exercise testing (CPET) assesses maximal oxygen uptake (VO2max) for cardiovascular fitness (CF) but has accessibility limitations.
• Continuous, unobtrusive monitoring of CF is needed, driving interest in wearable sensors and machine learning (ML).

Purpose of the Study:

• To predict cardiovascular fitness (CF) using ML algorithms with data from wearable technologies.
• To evaluate the effectiveness of ML in estimating maximal oxygen uptake (VO2max) during daily activities.

Main Methods:

• Collected 7-day unobtrusive data from 43 volunteers using wearable devices.
• Utilized Support Vector Regression (SVR) with 11 inputs (e.g., heart rate, breathing rate, activity metrics, anthropometrics) to predict VO2max.
• Employed SHapley Additive exPlanations (SHAP) to interpret the ML model's feature importance.

Main Results:

• Support Vector Regression (SVR) successfully predicted cardiovascular fitness (CF).
• SHAP analysis identified hemodynamic and anthropometric factors as key predictors of CF.
• Wearable sensor data, combined with ML, proved effective for estimating VO2max.

Conclusions:

• Cardiovascular fitness can be accurately predicted using wearable technologies and machine learning.
• ML models analyzing data from unsupervised daily activities offer a viable alternative to traditional CPET.
• Hemodynamic and anthropometric data are crucial for ML-based CF prediction.