What machine learning can and cannot do for mental health
- Article
- Mental Health
Diagnosing a mental health condition still relies mostly on interviews, self-reports, and questionnaires. These tools work, but they capture a snapshot, depend on how a person describes themselves, and often miss the early stages of an illness. A 2024 review by Kamala Devi Kannan, Senthil Kumar Jagatheesaperumal, Rajesh N. V. P. S. Kandala, Mojtaba Lotfaliany, Roohallah Alizadehsani, and Mohammadreza Mohebbi surveys how machine learning and deep learning are being applied to fill that gap.
What kinds of data
The review groups the work by the data a model learns from. Medical imaging, mainly MRI and EEG, gives a picture of brain structure and activity. Genetic data and biomarkers point to predispositions. Behavioral assessments, including speech, typing patterns, phone use, and social media activity, capture how a person acts day to day. Each of these has been used to build models for depression, bipolar disorder, and schizophrenia, the three conditions the review focuses on.
What kinds of models
Most of the studies covered use supervised learning: the model is trained on records where the diagnosis is known and then asked to predict it for new patients. Ensemble methods, which combine several models and average their output, are common because they tend to be more accurate and less sensitive to noise than any single model. Deep learning is used where the data is high-dimensional, such as brain scans or long recordings, and where hand-picking features is impractical.
The review also covers predictive modeling beyond diagnosis: risk scores that estimate who is likely to develop a condition, and longitudinal studies that try to forecast how an illness will progress in a given patient. These are the areas the authors see as most promising for individualized treatment, since a model that flags a worsening trend early gives a clinician time to act.
The open problems
The authors are direct about what holds the field back. Studies use different data, different definitions, and different evaluation methods, so results are hard to compare and often do not replicate. Combining data of different types, such as imaging with behavior, is still an unsolved engineering problem. And the ethical questions are large: mental health data is among the most sensitive there is, models trained on one population may fail on another, and a wrong prediction can carry real consequences for the person it is about.
Their conclusion is that these tools should be built with clinicians, ethicists, and patients at the table, not handed to them afterward. The review is available as a preprint on arXiv.