Leveraging Artificial Intelligence to Detect Short-term Suicidal Ideation from Ecological Momentary Assessment

Although artificial intelligence (AI) shows promise for suicidal ideation (SI) detection, accurate short-term prediction remains challenging. This reflects a misalignment between the clinical nature of SI and the design of existing AI studies, which typically rely on static or one-time assessments. According to Fluid Vulnerability Theory, SI fluctuates dynamically over hours or days, highlighting the importance of short-term detection for suicide prevention. Systematic reviews show that ecological momentary assessment (EMA) effectively captures such fluctuations with reduced recall bias. Within EMA, repeated text responses provide richer contextual and linguistic information than rating scales with low participant burden, while being more clinically relevant than social media text. However, achieving accurate short-term prediction requires understanding the complex semantic and sequential structure of EMA text, which is rarely modelled directly. Furthermore, meta-analytic evidence shows that SI rarely occurs in isolation, and comorbid symptoms such as sleep disturbance are associated with short-term fluctuations in SI. But the predictive value of comorbidity-related linguistic markers for improving short-term SI detection remains underexplored. To address these gaps, this thesis proposes three objectives: (1) To validate EMA text as a clinically meaningful data source, we established a dataset from seven-day repeated text responses of Cantonese-speaking adults; (2) To capture short-term SI fluctuations, we developed a large language model-based framework that models EMA text at semantic, local, and sequential levels to predict one-week SI outcomes; (3) To examine whether symptom comorbidity improves prediction and reveals shared linguistic mechanisms, we applied a deep learning framework, multi-task learning, to jointly model SI and insomnia. The proposed framework is expected to outperform static, single-task baselines and clinical benchmarks. This work aims to provide an interpretable, low-burden digital screening tool for real-world settings, extending assessment beyond one-off screening to short-term SI transitions.