Creative Commons source reading · Prediction and forecasting
Interpretable Feature Selection and Hybrid Deep Learning Models for Depressive Symptoms Prediction from Wearable Device Data
Ko J; Oh S; Enkhbayar D; Lee JK; Chung MK; Shin T; Kim MH; Lim HS; Urtnasan E; Key J
Originally published in J Med Syst · 2026-03-03
From the article
Selected source text. JournalOS removes markup and shortens excerpts where indicated. The full article provides the complete context.
Abstract
AbstractEarly detection and prediction of Depressive Symptoms is essential for improving mental health outcomes. This study proposes a hybrid deep learning and machine learning framework that utilizes tabular data collected from wearable devices, including sleep patterns, physical activity, and health-related indicators. Three ensemble learning models were used to identify influential predictors through the application of explainable artificial intelligence techniques such as SHAP and LIME. Based on the selected important features, two hybrid models were developed combining 1D-Convolutional Neural Network and Multi-layer Perceptron with LightGBM. The experimental results showed that models trained on selected features consistently outperformed those using the full feature set. The highest classification accuracy, 93.43%, was achieved…
Result
In this study, we applied three ensemble learning models Random Forest, XGBoost, and LightGBM to identify and analyze the most important features associated with depression. To improve the interpretability of the feature selection process, we employed XAI techniques, including SHAP and LIME, which enabled the quantitative assessment of each feature’s contribution to model predictions. Through this process, we identified candidate biomarkers potentially associated with depressive symptoms. Using these selected features, we developed two deep learning-based hybrid models MLP + LGBM and 1D-CNN + LGBM to predict depression outcomes and compared their performance with models trained on the full set of features. All Models were implemented using the PyTorch framework. The dataset was divided into 60% training, 20% testing, and 20%…
Discussion
In this study, a hybrid deep learning and machine learning model was designed and evaluated to predict depressive symptoms using tabular data collected from wearable devices. Specifically, SHAP and LIME were employed to identify and interpret key features contributing to model predictions. A central focus of this work was on feature importance analysis, which not only provided insights into the most influential predictors but also enabled a clearer understanding of their directionality and individual-level variability. The SHAP analysis revealed that ‘Night Sleep Time’, ‘Day Sleep Time’, and ‘Religion Activity’ consistently emerged as high importance features across all ensemble models. Among them, Night Sleep Time demonstrated a robust and stable protective effect, with durations close to the median being associated with a lower risk…
Editorial reading guide
- What data created and tested the model, and was evaluation independent?
- What is the cost of a false positive, false negative or out-of-distribution decision?
