FEV1 Prediction From Sensor Data of a Physical Activity Armband in COPD Patients: Results From the COSYCONET Cohort

Forced Expiratory Volume in one second (FEV1) is a critical clinical marker used to assess the severity and progression of Chronic Obstructive Pulmonary Disease (COPD). Physical inactivity is an important predictor of mortality in patients with COPD and can be reliably measured over a few days through wearable sensors. To enable continuous respiratory monitoring outside the clinic, this study introduces a machine learning framework that uses daily life sensor data and demographics to predict lung function, specifically FEV1. Using the COSYCONET cohort, we analyzed 707 week-long wearable recordings from 423 patients. A total of 144 engineered features, derived from accelerometry, step count, energy expenditure, and physical activity level, were used to predict FEV1. We applied several interpretable and regularized models, including linear regression with forward and backward subset selection, Lasso regression, Support Vector Machines with Recursive Feature Elimination (SVM-RFE), and Random Forests. Results showed that models using feature selection or regularization (particularly backward subset selection and Lasso) significantly outperformed baseline approaches (Mean Absolute Error ≈ 15%), achieving prediction errors within clinically meaningful thresholds. Feature importance analyses identified accelerometry-based variability measures, age, and walking metrics as the most informative predictors. These findings demonstrate the potential for non-invasive, real-time monitoring of respiratory health in COPD and provide a basis for future personalized health applications aimed at early detection of exacerbations and reduced healthcare burden.