The risks of risk. Regulating the use of machine learning for psychosis prediction

Recent advances in Machine Learning (ML) have the potential to revolutionise psychosis prediction and psychiatric assessment. This article has two objectives. First, it clarifies which aspects of English Law are relevant in order to regulate the use of ML in clinical research on psychosis prediction. It is argued that its lawful implementation will depend upon the legal requirements regarding the balance between potential harms and benefits, particularly with reference to: (i) any additional risks introduced by the use of ML for data analysis and outcome prediction; and (ii) the inclusion of vulnerable research populations such as minors or incapacitated adults. Second, this article investigates how clinical prediction via ML might affect the practice of risk assessment under mental health legislation, with reference to English Law. It is argued that there is a potential for virtuous applications of clinical prediction in psychiatry. However, reaffirming the distinction between psychosis risk and risk of harm is paramount. Establishing psychosis risk and assessing a person's risk of harm are discrete practices, and so should remain when using artificial intelligence for psychiatric assessment. Evaluating whether clinical prediction via ML might benefit individuals with psychosis will depend on which risk we try to assess and on what we try to predict, whether this is psychosis transition, a psychotic relapse, self-harm and suicidality, or harm to others.