Use of Wearable Devices to Detect Sedentary Behaviour in an Office Environment

Abstract

Introduction: Sedentary Behaviour (SB) is a significant public health problem associated with chronic illness, poorer quality of life, and increased healthcare costs. Fitbit-type wearable devices provide continuous monitoring of activity, heart rate, and sleep, creating an opportunity to detect sedentary behaviour in office-based and free-living contexts. Methodology: This study presents an interpretable, subject-aware, and prototype-oriented SB detection framework using a public Fitbit dataset. Five supervised ML models (Logistic Regression, SVM, KNN, Random Forest, and XGBoost), two baseline DL models (FNN and LSTM), and exploratory 1D-CNN and CNN-LSTM sequence models were evaluated using leakage-safe, subject-independent validation. Activity, sleep, heart rate, step intensity, engineered, and subject-normalised features were considered, with SHAP used for explainability. Results: Under participant-grouped five-fold cross-validation, XGBoost achieved the strongest aggregate binary results, with accuracy 0.9366 ± 0.0183, macro-F1 0.8082 ± 0.0333, balanced accuracy 0.7962 ± 0.0577, ROC-AUC 0.9203 ± 0.0430, and PR-AUC 0.9890 ± 0.0091. Threshold sensitivity supported the 600-minute cut-off because it produced the highest macro-F1 and balanced accuracy, while participant-level bootstrap intervals quantified uncertainty around the estimates. Conclusion: The revised framework also includes a Streamlit prototype that supports Fitbit CSV upload, automated preprocessing, probability-based sedentary risk scoring, low/moderate/high risk interpretation, and SHAP-based explanations. It is a research prototype rather than a validated workplace deployment. The results show that tree-based ML models remain more reliable than sequence-based DL models for this small and imbalanced Fitbit dataset, while the interface provides a route for further evaluation.