Reconsidering false positives in machine learning binary classification models of suicidal behavior

We posit the hypothesis that False Positive cases (FP) in machine learning classification models of suicidal behavior are at risk of suicidal behavior and should not be seen as sheer classification error. We trained an XGBoost classification model using survey data from 173,663 Norwegian adolescents and compared the classification groups for several suicide-related mental health indicators, such as depression, anxiety, psychological distress, and non-suicidal self-harm. The results showed that as the classification is made at higher risk thresholds - corresponding to higher specificity levels - the severity of anxiety and depression symptoms of the FP and True Positive cases (TP) become significantly more similar. In addition, psychological distress and non-suicidal self-harm were found to be highly prevalent among the FP group, indicating that they are indeed at risk. These findings demonstrate that FP are a relevant risk group for potential suicide prevention programs and should not be dismissed. Although our findings support the hypothesis, we account for limitations that should be examined in future longitudinal studies. Furthermore, we elaborate on the rationale of the hypothesis, potential implications, and its applicability to other mental health outcomes.