This page was translated automatically from Turkish. Read the original

Bu sayfa, 26. Dünya Psikiyatri Kongresi'nde sunulan çalışmanın yapay zekâ desteğiyle hazırlanmış Türkçe özetidir. Çalışmanın sahibi sunan yazarlardır; atıf için orijinal kaynağa başvurun.

EV0445

A multilayer contextual annotation framework for large language model–assisted structuring of Japanese psychiatric records

A MULTIGRANULAR CONTEXTUAL ANNOTATION FRAMEWORK FOR LLM-ASSISTED STRUCTURING OF JAPANESE PSYCHIATRIC RECORDS

A TEI/XML framework that explicitly labels the source of information, whom it concerns, its function within the note, its certainty status, and the links between assessments and supporting statements enables human review of large language model outputs.

The aim of the study is to develop a TEI/XML framework that preserves the original text and allows human review of annotations generated by a large language model (LLM). The rationale is that psychiatric intake notes combine patient statements, family reports, referral letters, and clinicians’ observations and inferences; these distinctions become important when notes are reused for research or when information extracted by an LLM is compared with the original text. While existing approaches encode whether a finding is present, absent, or uncertain, and sometimes whether it concerns someone other than the patient (Real-MedNLP v8; the 2010 i2b2/VA schema; ConText), the proposed framework jointly addresses the source (source), whom the information concerns (about), its function within the note (function), its certainty status (stance), and the evidential links between assessments and supporting statements (evidence). In the poster, patient statements, family reports, clinician observations, and clinician assessments were separately annotated in a simulated intake note; a diagnostic assessment marked as uncertain was linked to two supporting statements through evidence links. The framework specification and reference tools in Python/TypeScript for validation, extraction, and editing have been implemented. In an ongoing pilot study, three simulated intake interviews (transcripts and sample records) are being independently annotated by two annotators, with the same texts also being annotated by local and commercial LLMs; outcome measures include dimension-specific inter-annotator and human–LLM agreement (F-score, κ), as well as error analysis. The authors state that the framework provides an auditable foundation for LLM-assisted structuring, while cross-language transfer has yet to be tested.

Tüm WCP 2026 özetleri