Developing a Suicide Risk Prediction Algorithm Using Electronic Health Record Data in Mental Health Care: Real-World Case Study
Bibliographic record
Abstract
Background: Artificial intelligence (AI) offers potential solutions to address the challenges faced by a strained mental health care system, such as increasing demand for care, staff shortages, and pressured accessibility. While developing AI-based tools for clinical practice is technically feasible and has the potential to produce real-world impact, only a few are actually implemented into clinical practice. Implementation starts at the algorithm development phase, as this phase bridges theoretical innovation and practical application. The design and the way the AI tool is developed may either facilitate or hinder later implementation and use. Objective: This study aims to examine the development process of a suicide risk prediction algorithm using real-world electronic health record (EHR) data through a qualitative case study approach for clinical use in mental health care. It explores which challenges the development team encountered in creating the algorithm and how they addressed these challenges. This study identifies key considerations for the integration of technical and clinical perspectives in algorithms, facilitating the evolution of mental health organizations toward data-driven practice. The studied algorithm remains exploratory and has not yet been implemented in clinical practice. Methods: An exploratory, multimethod qualitative case study was conducted, using a hybrid approach with both inductive and deductive analysis. Data were collected through desk research, reflective team meetings, and iterative feedback sessions with the development team. Thematic analysis was used to identify development challenges and the team's responses. Based on these findings, key considerations for future algorithm development were derived. Results: Key challenges included defining, operationalizing, and measuring suicide incidents within EHRs due to issues such as missing data, underreporting, and differences between data sources. Predicting factors were identified by consulting clinical experts; however, psychosocial variables had to be constructed as they could not directly be extracted from EHR data. Risk of bias occurred when traditional suicide prevention questionnaires, unequally distributed across patients, were used as input. Analyzing unstructured data by natural language processing was challenging due to data noise, but ultimately enabled successful sentiment analysis, which provided dynamic, clinically relevant information for the algorithm. A complex model enhanced predictive accuracy but posed challenges regarding understandability, which was highly valued by clinicians. Conclusions: To advance mental health care as a data-driven field, several critical considerations must be addressed: ensuring robust data governance and quality, fostering cultural shifts in data documentation practices, establishing mechanisms for continuous monitoring of AI tool usage, mitigating risks of bias, balancing predictive performance with explainability, and maintaining a clinician "in-the-loop" approach. Future research should prioritize sociotechnical aspects related to the development, implementation, and daily use of AI in mental health care practice.
Fetched live from OpenAlex and de-inverted. Abstracts are not stored in this database: the inverted indexes are 8.6 GB of the frame’s 9.3 GB of text, and the host has 13 GB free.
How this classification was reachedexpand
Full frame machine prediction
Teacher imitationNot calibrated prevalence, not ground truth. Human validation pending. The Gemma side is a direct model label for every work in the frame, read from the title-only record. The Codex side is a classifier learned from the 10,348 direct Codex labels and calibrated to design-weighted sample rates; fields without enough sample support carry no Codex call. Candidate is the union of the two sides; consensus is their intersection. These outputs are machine_predicted_unvalidated and are not human labels.
Distilled classifier scores by category (both heads)
| Category | Codex | Gemma |
|---|---|---|
| Metaresearch | 0.025 | 0.062 |
| Meta-epidemiology (narrow) | 0.001 | 0.001 |
| Meta-epidemiology (broad) | 0.000 | 0.001 |
| Bibliometrics | 0.002 | 0.001 |
| Science and technology studies | 0.004 | 0.003 |
| Scholarly communication | 0.003 | 0.003 |
| Open science | 0.003 | 0.004 |
| Research integrity | 0.003 | 0.003 |
| Insufficient payload (model declined to judge) | 0.001 | 0.000 |
Machine scores (provisional)
The two teacher heads of the student model, read on this work. A score orders the frame for review; it never asserts a category, and the validation status ships verbatim with every row.
Baseline scores from an immature model (maturity gate not passed, 7 training rounds). Scores rank; they never assert a category.
score_only:v0-immature-baseline · verbatim from the scoring run: score_only means the number may rank works, and no category label ships from itClassification
machine, unvalidatedMachine predicted; a candidate call from one source (direct Gemma or distilled Codex), not a consensus.
How this classification was reached, model by model and score by score, is at the end of the page under "How this classification was reached".