Bipolar disorder (BD) is a recurrent disorder, causing functional impairment and raised mortality, particularly due to suicide. However, the difficulty in predicting suicidal behaviors relies in the lack of clear biomarkers.

Machine learning (ML) has emerged as a promising tool to enhance suicidal prediction. However, most ML studies focused on lifetime attempts, without having a predictive time window, and did not employ time-dependent variables. Moreover, most studies lie on cross-sectional databases, without including more than one time-point.

First, we aimed to predict 12-months suicide attempts in a naturalistic sample of BD patients, using clinical and demographic data.

Second, we aimed to improve the prediction by including information from intermediate visits (1, 3, and 6 months), mimicking more closely the clinician’s way of thinking and the multiple observations a patient receives.

A sample of 163 BD patients (53% females, mean age 44.7, SD 15.3) were recruited.

Based on EHR, 56 clinical and demographic features were extracted, including hospitalizations, suicidal behaviors lifetime and in the last 12 months, along with comorbidity, family history, work, and therapies. Patients were followed up for 12 months.

Support Vector Machine (SVM) was used to differentiate subjects who attempted suicide versus those who did not in a 12-month time window, within a repeated nested Cross-Validation. The SVM was optimized weighting the hyperplane for uneven group sizes.

Then, we repeated the analysis including information from intermediate visits (1, 3, 6 months after the first contact). For each visit, we created a composite score based on current therapy, new admissions, and ER presentations. To avoid circularity, all the information (ER, admission etc.) related to a suicide attempt were not included.

During the 12-months follow-up, 9.8% of patients attempted suicide. The results from the 12-months suicide prediction model obtained an Area Under the Curve of 0.71(with a Balanced Accuracy (BAC) of 68%).

After incorporating the composite scores based on intermediate visits in the model, the prediction raised to an Area Under the Curve of 0.78 (BAC 73%), suggesting that including intermediate visits is a valid method to improve prediction.

The features that contributed the most to the prediction were the composite score at 6-month visit, lifetime number of suicide attempts, suicide attempts in the last 12 months, substance of abuse (other than cannabis), and antipsychotics.

ML proved a good prediction accuracy for suicide in a 12-months time window, and the prediction was improved by including data from intermediate visits. The model showed the importance of time-dependent features, such as attempts in the last 12 months. Our analysis might help in identifying early clinical risk factors and underlies the importance of multiple evaluations in populations at risk.

None Declared