As machine learning shifts the clinical paradigm in the health care system, more clinicians and patients have been benefited from using big data such as personal health records (PHR). This paper presents lean body mass(LBM) prediction models with features extracted from wearable device data.
Recruited 300 subjects aged 20-65 were required to wear ActiHeart (Camntech Ltd., UK), and underwent a basic checkup and a treadmill test. After preprocessing the metabolic equivalents and heart rates obtained from the device, we implemented K-mean clustering (6 centroids) in this data and performed regression analysis with these centroids to extract informative features. Prediction models were designed with covariates such as age, sex, height, weight, slope, and intercept.
The results of 5-fold cross-validation of Linear regression and Random Forest models showed satisfactory performance in predicting LBM (with the coefficient of determination of 0.91 and 0.90, respectively). Root Mean Squared Error of LM was 2.98 and of RF is 3.15. Mean Absolute Error of LM was 2.21 and of RF was 2.35.
This paper shows the LM and RF models predicting LBM with additional features from ActiHeart data. Previous studies have found a linear relationship between LBM and peak VO2 and between HR and VO2. These linear relationships imply that the features (slope and intercept) used in the prediction models for LBM are reasonable. Continuous monitoring of body composition can motivate users to control weight and provide important insights for clinicians to interpret the association between diabetes and body composition.