Predicting Mental Health Using Lifelog Data: Application of Median Resampling and Data Augmentation Techniques

Lifelogging, the practice of recording and analyzing personal data through wearable devices and mobile applications, offers a rich source of information for predictive analytics. This study explores the use of median sampling techniques and data augmentation to handle missing and outlier data in lifelog data for predicting mood, stress, and sleep. The dataset, sourced from the 3rd ETRI Human Understanding Paper Competition, includes smartphone and smartwatch data from four participants over periods of 105 and 115 days. Key features extracted include GPS-based mobility metrics, heart rate variability (HRV) features. To generalise the model, we performed data augmentation using jittering. We used four machine learning models, optimizing hyperparameters through cross-validation techniques. To address class imbalance, SMOTE was applied. Model performance was evaluated using F1 scores and AUROC. SHapley Additive exPlanations (SHAP) analysis was conducted to interpret feature importance. The median sampling approach provided stable and representative values, improving the robustness of predictions. We found that data augmentation with jittering improved performance for all models except LR. Random forest exhibited the highest performance for most targets, particularly in predicting emotional states and sleep metrics. SHAP analysis highlighted HRV and day location entropy as key features.