JMIR Publications

ABSTRACT

Background:

Artificial intelligence (AI) approaches to suicide risk stratification increasingly rely on electronic health record data and clinical notes. However, psychiatric documentation is not a neutral transcript of the clinical encounter. What becomes documented may be shaped by clinical judgment, workload, confidentiality concerns, information-system design, and sociocultural context. These processes may be particularly consequential in settings where psychiatric information is highly sensitive and where incomplete or indirect documentation can alter the data available for AI modeling.

Objective:

This study aimed to examine how psychiatric documentation practices shape the readiness of clinical data for AI-supported suicide risk stratification at Sultan Qaboos University Hospital (SQUH), Sultanate of Oman, and to explore clinicians' perspectives on the most appropriate role for AI in suicide risk assessment and clinical workflow.

Methods:

A qualitative exploratory study was conducted using semi-structured interviews with 9 psychiatric clinicians, comprising 6 residents and 3 consultants involved in suicide risk assessment at SQUH. Interviews were conducted between September 2025 and January 2026, audio-recorded, transcribed, deidentified, and analyzed using a mixed inductive-deductive thematic approach. Arabic and bilingual interviews were translated into English and independently reviewed. A small number of deidentified clinical notes were reviewed as supplementary contextual material.

Results:

Documentation was described as selective and clinically filtered, with immediate suicide risk and management information generally prioritized while broader psychosocial, historical, and contextual information was less consistently recorded. Privacy was a major influence on this process. Eight of 9 participants raised concerns about the accessibility of psychiatric notes beyond the immediate treating team. Within a close-knit sociocultural setting, concerns about stigma, family reputation, and unintended disclosure contributed to clinicians generalizing, withholding, or verbally communicating sensitive information rather than documenting it explicitly. Documentation was further shaped by workload, shift timing, free-text documentation, clinician seniority, training, continuity of care, and differences in risk interpretation. Importantly, 7 of 9 participants viewed AI as more immediately useful for supporting interviewing and documentation, including transcription, note structuring, and reducing administrative burden, than for independently predicting or stratifying suicide risk.

Conclusions:

The clinical record available for AI is a selectively constructed representation of the psychiatric encounter, shaped by clinical, organizational, technological, and sociocultural factors. Consequently, important clinical information may be deliberately omitted or recorded in less specific terms because of privacy, cultural, or clinical considerations. This means that the clinical notes may not fully represent the information available to the clinician, which can limit their reliability as a data source for AI-based suicide risk modeling. Before prioritizing autonomous suicide risk prediction, health systems should strengthen psychiatric confidentiality, improve documentation structures and data continuity, and evaluate culturally and linguistically appropriate AI tools that support the interview and documentation process while preserving clinician responsibility for risk formulation.