Abstract
During the implementation of our research project, we were faced with challenges regarding the security of patient privacy and the technical implementation of therapy recordings to reassure sufficient data quality for the data analysis. Therefore, in addition to improving suicidality prediction with multimodal methods, we will develop an end-to-end workflow for further AI research in the clinical context.
Introduction
Suicide remains a critical global health challenge, yet identifying individuals at immediate risk is notoriously difficult. Current clinical practice relies heavily on subjective self-reports and unstructured interviews, which often suffer from low interrater reliability. While AI has shown promise in detecting mental health markers in social media data, a significant "data gap" exists in actual clinical settings.
The SPEAK-SAFE project addresses three primary challenges: the lack of high-quality non-English (specifically German) clinical datasets, the need for multimodal analysis (combining content with acoustic nuance like pitch and rhythm), and the necessity of a secure, GDPR-compliant pipeline for sensitive patient data.
Methodology: Multimodal Fusion and Secure Workflow
The SPEAK-SAFE study protocol outlines a sophisticated approach to analyzing therapeutic exchanges. The system processes two distinct data streams to capture the full nuance of patient communication:
- Textual Stream: Utilizes Large Language Models (LLMs) to analyze linguistic patterns from transcripts.
- Acoustic Stream: Employs the
openSMILEtoolkit to extract 88 specific features—such as jitter, shimmer, and speech rate—known to correlate with depression and suicidality.
To ensure clinical reliability and transparency, the researchers utilize open-source LLMs (GPT-OSS family), comparing a massive 120B parameter model against a more efficient 20B parameter model. This efficiency is crucial for local deployment on standard hospital hardware (requiring as little as 16GB of memory). The pipeline also includes manual pseudonymization and local data storage to meet strict German data protection laws (BDSG).
Data Collection and Clinical Measures
The project is building a new, trans-diagnostic German psychiatric corpus involving 200 planned interviews with patients diagnosed with depression, bipolar, or psychotic disorders. AI predictions are benchmarked against a "gold standard" assessment provided by board-certified psychiatrists using validated scales.
As detailed in the table above, clinical data ranging from the Beck Depression Inventory (BDI-II) to the Clinical Global Impression of Suicide Severity (CGI-SS) provides the ground truth for training and validating the AI models. This ensures that the AI's objective analysis is grounded in established psychiatric practice.
Discussion and Future Directions
As a study protocol, SPEAK-SAFE focuses on establishing a robust methodology. Power analysis indicates a required sample size of $n=162$ to detect significant effects. While preliminary benchmarks for LLMs in suicidality detection are promising—with some models reaching 87.5% accuracy in identifying risk from admission notes—real-world implementation faces hurdles like audio quality in noisy clinical environments.
Future work will focus on integrating "Ecological Momentary Assessment" (EMA) via smartphones to capture data between therapy sessions, ensuring that AI serves as a tool to augment, rather than replace, clinical expertise. By treating privacy as a foundation rather than a hurdle, the SPEAK-SAFE workflow enables a new era of secure AI research in sensitive clinical contexts.
Clinical Trial Registration: DRKS00027878
Summary updated