Abstract
Depression has severe impacts, such as causing labor loss, increasing the risk of lifestyle‐related diseases, and raising suicide risk, with potential effects extending to future generations. This underscores the importance of early prediction and intervention. Advances in machine learning are expected to improve prediction accuracy, with applications expanding to classifying depression, predicting treatment response, and assessing recurrence risk. This paper reviews recent trends and challenges in machine learning research on early prediction of depression. Despite numerous studies, the accuracy of predicting new onset of depression remains low. There is also a lack of approaches that comprehensively utilize indicators related to abnormalities in emotional and reward systems‐mechanisms linked to the onset of depression‐alongside traditional demographic and psychosocial factors. Achieving high accuracy with fewer features remains a challenge. Moreover, AI tuning is often insufficient, highlighting the need for efficient optimization methods such as Bayesian optimization. Additionally, the imbalance in depression prevalence necessitates the use and optimization of PR curves and F1 scores as evaluation metrics. This paper discusses the future prospects for building practical prediction models in light of these challenges.
