F49 Machine learning approach in analysis of enroll-hd data for suicidality prediction in huntington disease

Abstract

Background Suicidal ideation and suicidal behaviour are frequently reported, severe features in Huntington disease gene expansion carriers (HDGECs), but it is difficult to predict who are at increased risk. So far, no suicidality prediction models have been developed using machine learning approach (MLA). Objective To develop a model for prediction of suicidal ideation or suicidal behaviour in HDGEC based on Enroll-HD data using MLA. Design/methods We have developed a prediction model based on MLA using the third Enroll-HD study periodic dataset (PDS3). Suicidal ideation/behaviour was measured with the Columbia-Suicide Severity Rating Scale (C–SSRS). HDGECs with no or ‘passive’ suicidal ideations [state 1] at their first visit, who at the follow-up (FUP) either stayed in state 1 or worsened to ‘active’ suicidal ideations and/or suicidal behaviour [state 2] were analyzed. The PBAs scale was used to assess the presence of behavioural symptoms. Prediction algorithm was based on Boosted Trees (implementation from XGBoost Library for Python) and contained 48 variables from the PDS3. We also used Fisher Exact test, Mann–Whitney U-test, and Holm method. Results For 377 HDGECs (114 pre-manifest; 161 males; median age 50 [20;78]; median nCAG=43 [38;65]) C-SSRS data of two consecutive visits were available. At the FUP, 316 remained in state 1 and 61 HDGECs had worsened to state 2. Sixty four percent of the HDGECs who remained in state 1 at FUP were accurately classified (probability as having state 2 <30%). HDGECs who worsened to state 2 were correctly predicted in 38% cases (probability of being classified as having state 2 >60%). We then compared the poorly (probability <30%; 31 subjects) with the well (probability >60%; 23 subjects) classified groups in state 2 at FUP and found significant difference in the PBAs total scores for depression, anxiety, aggression, and apathy, with more severe baseline scores in the well classified HDGECs. However, regression analysis did not show a significant relationship of these behavioural symptoms and the probability of being classified as subject in state 2 at FUP. Conclusions Our model showed moderate accuracy. Further research is needed to understand the risk for development of suicidal ideation/behaviour in HDGECs with mild behavioural symptoms.