Can machine learning be used to assess risk of suicide?

Problem

Suicide is the leading cause of death for those aged 20 to 34 in the UK. General practitioners play a central role in suicide risk assessment and management, with 91% of individuals consulting their GP on at least one occasion in the year before their death by suicide. Despite significant research in the area there is no clinical risk scoring tool for suicide approved by The National Institute for Health and Care Excellence. A validated clinical tool for risk assessment could be applied to enable earlier intervention in those at higher risk, but so far reliable methods have been lacking. The NHS Long Term Plan identifies machine learning as a key area for development in healthcare – can machine learning aid in assessing suicide risk?

Approach

The study protocol is available on PROSPERO. A search strategy was developed and searches made in Embase, Medline, PsycINFO, and Web of Science. Studies were included where machine learning methods were used to generate predictive models for suicide or suicide attempts in a population who were self or clinician-identified as being at risk of suicide or who were deemed to have died from suicide.

Findings

Four databases yielded 1306 records. After removal of duplicates there were 854 records. Screening of abstracts left 91 records. Full text screening left 40 for qualitative synthesis. The majority of studies used cross sectional data from electronic health records. Populations included both general community, emergency department, and patients engaged with both inpatient and outpatient psychiatric care. A significant number of studies used the US Veterans Health Administration data. The preliminary analysis from data extraction shows there have been significant advances in the accuracy of machine learning techniques for suicide risk assessment in recent years.

Consequences

Accurate clinical tools for suicide risk assessment based on machine learning could help with identification and early intervention for those most at risk. With the ongoing analysis of the data from this review we hope to identify the current strengths and weaknesses in the use of machine learning for suicide risk prediction.