Background: Machine learning techniques are now being widely used to probe electronic medical records and other databases to identify people at high risk of suicide or self-harm. Our aim is to identify the positive predictive value (PPV) of these instruments.
Methods: Systematic review and meta-analysis of all studies that have developed risk prediction instruments for suicide or self-harm using machine learning. We calculate the pooled PPV of these instruments for suicide and self-harm outcomes.
Results: We will report on the number of relevent studies identified, the characteristics of those studies and the pooled PPV for suicide, self-harm and suicide or self-harm outcomes. Results will be stratified by study type: case-control, case-cohort and cohort. For the case-control studies we will report the pooled PPV after adjusting for prevalence.
Interpretation: There is now considerable effort to use machine learning methods to develop new risk prediction instruments for suicide. This study will shed light on the value of that endeavour by identifying whether it offers improved prediction over traditional methods of risk classification. We will discuss the implications of our findings for clinical practice.