Leveraging timeline integration methods for effective suicidal ideation detection

• We introduce Timeline Integration Methods (TIMs) to detect suicidal ideation. • TIMs model temporal dependencies across posts written by a user. • TIMs improve zero-shot LLMs performance in suicidal ideation detection (SID). • We identify TIMs best suited to each model by size and reading comprehension. • Medical LLMs with TIMs outperform general-purpose counterparts on SID. Suicidal Ideation Detection (SID) is a critical task that involves analyzing a user’s social media posts to assess the severity of suicidal ideation (SI) and enable early intervention. Recent research has leveraged the advanced language understanding capabilities of large language models (LLMs) to perform this task more effectively. However, existing methods often process multiple posts simultaneously, which may overlook temporal dependencies between posts and exceed input length limitations. To address these challenges, we introduce timeline integration methods (TIMs), which model temporal dependencies across posts by incorporating information from earlier time points through prompt components such as the predicted label distribution of previous posts or recent post texts. This approach enables LLMs to track the progression of SI risk and make more accurate severity assessments. Furthermore, it remains unclear whether domain-specific LLMs for the medical field outperform general-purpose LLMs in SID tasks. We apply TIMs across various LLMs and benchmark them against existing methods. Our key findings indicate that: (1) TIMs significantly improve the performance of LLMs on the SID task, particularly on long user posts; (2) the effectiveness of TIMs varies depending on the model size and its capability in reading comprehension; and (3) domain-specific LLMs for the medical field show superior performance in SID, particularly when enhanced with TIMs.