Comparative Performance of Single-Task and Multi-Task Learning Models for Predicting Schizophrenia Symptoms Using Smartphone-Based Digital Phenotyping.

This preprint presents a comparative study of Multi-Task Learning (MTL) and Single-Task Learning (STL) for predicting mental health symptoms within a digital phenotyping framework. The analysis leverages passive smartphone sensing data and self-reported Ecological Momentary Assessment (EMA) measures from the CrossCheck dataset to evaluate how effectively MTL generalizes across multiple symptom prediction tasks compared with STL. To ensure methodological consistency across all predictive tasks, L1-regularized models are employed, using LassoCV for STL and MultiTaskLassoCV for MTL. The choice of regularization is motivated by strong diurnal patterns in daily activities and high collinearity among passive-sensing features, both of which necessitate sparsity control and robust feature selection. Results indicate that MTL consistently outperforms STL in predicting the ten EMA symptom trajectories, particularly when appropriate hyperparameter tuning is applied. Statistical validation using the Wilcoxon signed-rank test confirms that the performance gains of MTL over STL are significant (p = 0.0039). Symptom-level analysis shows that negative symptoms, such as hearing voices, perceptual disturbances, and harm-related thoughts, are the most difficult to predict, likely due to lower behavioral variability. In contrast, positive affective symptoms, including hope, calmness, and social engagement, exhibit greater variability and benefit more from model tuning. The findings suggest that multitask learning offers a promising direction for building more accurate and robust digital phenotyping models for mental health monitoring. At the same time, the study highlights important limitations, including a small dataset, missing data, and the absence of temporal sequence modeling. Future work is encouraged to explore larger datasets and time-series deep learning approaches for predictive performance and generalizability.