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

This study compares the comparative effectiveness of Multi-Task Learning (MTL) and Single-Task Learning (STL) in predicting schizophrenia symptoms using smartphone-based digital phenotyping data. Using the CrossCheck dataset, which combines passive smartphone sensing with Ecological Momentary Assessment (EMA) responses from 75 participants diagnosed with schizophrenia, LassoCV and MultiTaskLassoCV models were used. Analysis revealed that MTL consistently outperformed STL for generalized symptom prediction across ten EMA-based symptoms, with statistical validation provided by the Wilcoxon signed-rank test, confirming a significant improvement (p = 0.0039) when hyperparameter tuning was applied. Negative symptoms (voices, perceptual disturbances, harm-related thoughts) showed the lowest prediction errors, indicating more stable behavioral correlates, while positive affective symptoms (hopefulness, social engagement) proved more challenging to predict. The findings demonstrate that leveraging shared structure among correlated symptom dimensions through MTL can improve predictive accuracy in digital mental health monitoring systems, offering potential for scalable, data-driven tools that augment traditional clinical assessment and enable earlier detection of symptom deterioration. Keywords: schizophrenia, digital phenotyping, multi-task learning, machine learning, ecological momentary assessment, passive sensing, symptom prediction, LassoCV,