Identification of suicidal ideation in major depressive disorder: A machine learning approach with multimodal digital features

Vollständiger Abstract

Worum geht es in dieser Arbeit?

Current detection models for suicidal ideation (SI) among depressed patients have primarily relied on clinical and biological features. This study aims at enhancing the real-world applicability of these models by using multimodal digital features. Patients self-administered the Hospital Anxiety and Depression Scale (HADS) and 20-item Toronto Alexithymia Scale (TAS-20). Their multimodal features (facial, acoustic and linguistic features) were obtained using ecological momentary assessment (EMA) with our self-developed smartphone application over 7 days. The Structured Interview Guide for the Hamilton Depression Rating Scale (HDRS) were conducted to identify clinically rated SI. Mood description recordings obtained from EMA were processed through natural language processing, Openface, and OpenSmile to obtain linguistic, facial and acoustics features, respectively. Multimodal features and questionnaires were then used for training and testing machine learning algorithms. The study was conducted and reported under the TRIPOD+AI guideline. Among the recruited 99 patients with major depressive disorder (MDD), 38 (mean age = 49.3 ± 9.7 y, 76% female) had SI while 61 (mean age = 51.5 ± 11.5 y, 76% female) did not have SI. Among the 10 machine learning models evaluated, the CatBoost classifier demonstrated the strongest detection performance, achieving an AUC of 0.79 (p < .001), accuracy of 0.79 and F1-score of 0.73. K-Nearest Neighbours (KNN), Artificial Neural Network (ANN) and Naive Bayes models also provided comparable results. Multimodal features were correlated with SI among MDD patients. Machine learning algorithms have modest performance in identifying SI among patients with MDD using multimodal digital features, including facial expressions, vocal characteristics and language use. However, larger and more diverse datasets are needed to enhance the generalizability and accuracy of these machine learning approaches in real-world clinical settings.