Comparative Analysis of Deep Learning Models on Depression Prediction

Abstract

Major depressive disorder (MDD) is a severe psychiatric condition characterized by persistent emotional, cognitive, and functional impairments. Traditional diagnostic approaches heavily rely on subjective self-report scales and clinical observations, which are inherently vulnerable to recall biases and social desirability effects. While recent advancements in artificial intelligence (AI) offer promising paradigms for enhancing predictive accuracy and enabling personalized mental healthcare, implementing these models effectively across heterogeneous clinical environments remains highly challenging. This study conducts a horizontal comparison of DNNs, RNNs, and CNNs in MDD prediction, focusing on the correspondence among model architecture, training data type, and predictive accuracy. The findings indicate that DNNs are more suitable for structured clinical and questionnaire variables, RNNs for longitudinal and time-series data, and CNNs for image-based and spatial-representation data. However, model accuracy is highly dependent on the characteristics of the processed data and the sources of training samples, which broadly constrain existing findings. Future research should further develop multimodal transformer-based modeling to improve the capacity of AI models to process multimodal data.