The story of machine learning (ML) begins with several possible conceptions. Arthur Samuel, a computer scientist at IBM, coined the phrase ‘machine learning’ in 1959 to describe his pioneering work programming machines to essentially ‘teach’ themselves by applying mathematical and statistical models. His breakthrough was building the first computer-based program of the strategy game checkers, using scoring functions to serve as rudimentary algorithms to predict the probability of winning.

The alternative text for this image may have been generated using AI.
Credit: Marina Spence and Jackie Niam / iStock / Getty Images Plus

But Samuel was not alone in carving out ML as the starting point of what is considered a central component of artificial intelligence. Donald Hebb’s 1949 theory describing the coordinated or associative activity of neurons as a function of synaptic plasticity that occurs during learning, elegantly put as ‘neurons that fire together, wire together’, has remained the foundation influencing the development of artificial neural networks. The fact that ML was, in essence, born out of emulating brain mechanisms — identification of relationships, inductive reasoning, and prediction of the behavior of others — makes it a uniquely powerful tool for investigating human behavior and brain health.

Like many journals that feature mental health and psychiatry research, we have seen an influx of primary research studies using ML techniques, as well as commentary on its potential clinical applications. Back-of-the-envelope estimates using databases such as PubMed indicate a surge in ML and psychiatry publications, growing from approximately 200 in 2023 to over 1,200 in 2024. It is a trend that is likely to continue. But with that rapid rise comes the risk that technological and computational applications will outpace the general mental health readership’s ability to parse these developments. As the editors on our team have become more familiar with ML, we have come to appreciate how these evolving applications can yield powerful insights and have the potential to improve our understanding of mental health conditions, but are aware of how important it is that these studies and their findings and implications are presented in a way that researchers in other disciplines can grasp and use in their own work.

The January 2025 issue of Nature Mental Health includes several papers that advance the agenda of making ML in mental health and psychiatry research more accessible. A Review by Lucasius and coauthors provides a comprehensive overview of the specific processes involved in ML, but is framed to also capture the entire life cycle of an ML project. This is an important distinction from previous reviews and perspectives on ML applications in psychiatry, in that it is intended to spur collaboration among clinicians, mental health researchers and data scientists in order to leverage their respective expertise. The paper also considers some fundamental issues that ML in psychiatry must contend with, such as the reception and adoption of ML models of diagnostic predictions to aid clinicians, as well as the gaps in data sources, including electronic health records, that can be used for training models. The authors have also included an interactive Jupyter notebook with which readers can practice the concepts using an open dataset.

Previously, synthesizing and analyzing the methods, definitions and assessments used in a large set of studies required standard systematic reviews and meta-analytic techniques. ML techniques have provided an additional means for augmenting more conventional or established assessments of evidence. In their Analysis, Blekic and colleagues conducted a systematic review of 30 studies that used ML techniques to predict post-traumatic stress disorder. The findings of the systematic review indicate that ML approaches can identify different risk profiles and unexplored predictors beyond those identified by standard analysis and therefore can improve the risk stratification of people with post-traumatic stress disorder. The authors argue that an integrative model of data-driven components and theory-based models may be a promising new avenue for investigating this complex disorder.

The issue also includes an Article by Vannucci and colleagues that demonstrates how ML techniques can be used to surmount issues such as replicability and generalizability inherent in some mental health studies. The authors explored ML applied to caregiving-related early adversities, experiences that comprise a wide range of environmental risk factors for psychopathology later in life and may be affected by individual differences and disparate developmental trajectories. The findings suggest that earlier caregiver-related adversities and those of longer duration were associated with the development of a mental health disorder; in addition, the overall presence or absence of these adversities was useful for estimating transdiagnostic risk.

These three papers provide a glimpse into some of the exciting, novel and practical ways researchers are incorporating ML into their investigations. It is noteworthy that each emphasizes the need to use ML as a complement to or enhancement of other existing methodologies. This is a crucial caveat, given that ML techniques are in a nearly constant state of development and refinement. The value of applying ML within more established research programs is the ability to bring to the surface the relationships and factors that may not be observable with conventional data analysis. But an inherent limitation of ML applications is that without standard statistical methods for comparison, the interpretability and impact of these models is hampered. Together, these studies underscore what a ripe time it is for mental health and psychiatry research stakeholders to hypothesize and collaborate expansively.