2025 Theses Doctoral

Leveraging Recurrent Neural Networks for Predicting Suicidal Ideation: Advancing the Analysis of Ecological Momentary Assessment Data

Choo, Tse-Hwei

Introduction: Understanding the temporal dynamics of mental health conditions, such as suicidal ideation (SI), is critical for advancing research and clinical interventions. Ecological Momentary Assessment (EMA) is a method of increasing relevance for capturing such data over time, within participants daily lives. However, traditional analytical methods often fail to capture the episodic nature and complex temporal dependencies inherent in EMA mental health data. This project investigates the application of recurrent neural networks (RNNs) to EMA data to improve the prediction and understanding of SI, leveraging their ability to model sequential, high-dimensional data.

Methods: Data for this study were drawn from a randomized controlled trial examining the effects of dialectical behavioral therapy (DBT) and SSRI medication on SI in individuals with borderline personality disorder. Participants provided EMA responses multiple times per day over a period at baseline and again post-treatment. RNNs were trained on a portion of each participant’s baseline EMA data with EMA as the outcome, using various baseline and time-varying predictors. Predicted EMA SI values were then generated for a baseline EMA testing dataset, and for the post-treatment EMA period, These predicted SI values were examined to assess the accuracy of the RNN modeling. Baseline testing accuracy was compared to traditional mixed-effects models (MEMs) to demonstrate RNNs feasibility as an alternative for learning and predicting SI time series. Additionally, simulated EMA data was generated in order to describe the data conditions under which RNNs are most useful in modeling EMA data. Furthermore, post-treatment EMA SI predictions were explored to assess the long-term predictive capabilities of RNNs and investigate the prospects of using RNNs to draw causal or mechanistic insights.

Results: Key findings underscore the potential of RNNs in mental health research. At baseline, RNNs consistently outperformed MEMs in predicting SI, demonstrating their ability to model complex temporal dependencies and account for within- and between-subject variance. The simulated data analysis highlighted conditions under which RNNs excel, including the use of time-varying predictors and the availability of sufficient longitudinal data, offering guidance for future RNN use. The post-treatment analysis revealed that RNNs continued to provide reasonably accurate predictions, showcasing their robustness even when data were temporally distant from the training period, and following treatment interventions. Furthermore, differences in prediction accuracy between the DBT and SSRI treatment groups suggested that these interventions may uniquely influence SI dynamics in ways that RNN predictions may help to illuminate. Variables associated with prediction error differences provided further insight into treatment-specific mechanisms, highlighting the potential for RNNs to uncover nuanced effects not readily captured by traditional methods.

Conclusion: This dissertation advances the understanding of SI as a dynamic and context-dependent mental health outcome. By integrating EMA with RNN-based modeling, it addresses critical gaps in the analysis of temporal mental health data, offering novel insights into both the evolution of SI and the effects of therapeutic interventions. These findings underscore the potential of machine learning techniques to enhance EMA's utility, paving the way for future research and clinical applications aimed at improving mental health outcomes.