SCIENTIFIC SESSION

Background: Suicide accounts for more than 720,000 deaths worldwide each year. Traditional risk assessment methods have limitations when it comes to early detection. Machine learning is emerging as a promising tool for improving the prediction of suicidal and self-harming behaviour.
Objective: To develop and validate machine learning-based predictive models to identify individuals at high risk of suicidal thoughts and self-harming behaviours, prioritising interpretability and generalisation.

Methods: Two studies were conducted. The first (MINDCOVID) analysed 8,996 Spanish healthcare professionals during the COVID-19 pandemic to predict suicidal ideation. The second (CSRC-Epi) used health records of patients admitted to psychiatric hospitals in Catalonia (2015–2018) to predict self-harm following discharge.
Results: The model for suicidal ideation achieved an AUCROC of 0.87 and an AUCPR of 0.52, identifying the following predictors: previous suicidal ideation, panic attacks, intrusive thoughts and pandemic-related factors, with gender differences. The self-harm models achieved AUCROC = 0.76–0.82 and AUCPR 5–7 times higher than chance, highlighting: depressive episode, adjustment disorder, schizophrenia, previous self-harm and use of SSRIs/antipsychotics. The models demonstrated good temporal generalisation and generalisation across subgroups.

Conclusions: Machine learning is a promising approach for predicting suicidal behaviour in vulnerable populations, offering interpretable tools that can be applied in prevention and to support clinical and public health decision-making.