Cautions, concerns, and future directions for using machine learning in relation to mental health problems and clinical and forensic risks: A brief comment on "Model complexity improves the prediction of nonsuicidal self-injury" (Fox et al., 2019).

Machine learning (ML) is an increasingly popular approach/technique for analyzing "Big Data" and predicting risk behaviors and psychological problems. However, few published critiques of ML as an approach currently exist. We discuss some fundamental cautions and concerns with ML that are relevant when attempting to predict all clinical and forensic risk behaviors (risk to self, risk to others, risk from others) and mental health problems. We hope to provoke a healthy scientific debate to ensure that ML's potential is realized and to highlight issues and directions for future risk prediction, assessment, management, and prevention research. ML, by definition, does not require the model to be specified by the researcher. This is both its key strength and its key weakness. We argue that it is critical that the ML algorithm (the model or models) and the results are both presented and that ML needs to be become machine-assisted learning like other statistical techniques; otherwise, we run the risk of becoming slaves to our machines. Emerging evidence potentially challenges the superiority of ML over other approaches, and we argue that ML's complexity significantly limits its clinical utility. Based on the available evidence, we believe that researchers and clinicians should emphasize identifying, understanding, and explaining (formulating) individual clinical needs and risks and providing individualized management and treatment plans, rather than trying to predict or putting too much trust in predictions that will inevitably be wrong some of the time (and we do not know when). (PsycINFO Database Record (c) 2020 APA, all rights reserved).