This page was translated automatically from Turkish. Read the original

This page is an AI-assisted Turkish summary of the study presented at the 26th World Congress of Psychiatry. The study belongs to the presenting authors; please refer to the original source when citing it.

EV0477

An AI pipeline for recognizing and structuring retrospective psychiatric discharge records from digitized archives

ARTIFICIAL INTELLIGENCE PIPELINE FOR RECOGNIZING AND STRUCTURING RETROSPECTIVE PSYCHIATRIC DISCHARGE RECORDS FROM DIGITIZED ARCHIVES

A five-stage pipeline combining OCR and a large language model converted paper psychiatric discharge summaries into a structured research database of 14,554 records, achieving high completeness across key clinical fields.

The aim was to develop and evaluate an end-to-end artificial intelligence pipeline that converts heterogeneously digitized psychiatric discharge records into a validated relational research database suitable for epidemiological analysis. The source material consisted of paper discharge summaries from two inpatient units at a federal mental health research center in Moscow: 112 multi-record PDFs from 2011–2023, together with JPG and HEIC photographs of pages, totaling 429 unique source files and covering admissions from 2003–2023; consecutive inclusion was used, with no probability sampling. The five-stage pipeline comprises conversion of pages into images; optical character recognition using Yandex Vision (OCR, with Tesseract as a local fallback); structuring into JSON using YandexGPT according to a field-level system prompt (temperature 0.2, maxTokens 6,000, an input limit of 45,000 characters, and documents exceeding 50,000 characters split at institutional headers and then reassembled); validation (date normalization, mandatory patient-name checks, and deduplication using file_hash; responses with fewer than 30% of fields populated are quarantined rather than entered into the database); and storage in SQLite. The pipeline produced 14,554 records, extracting more than 40 structured fields per admission. Field completeness in the extracted records was as follows: patient name 100.0%, admission date 98.5%, discharge date 97.9%, ICD-10 code 94.5%, diagnosis text 93.9%, mental status 90.0%, treatment administered 89.3%, life history 85.8%, age at admission 83.1%, date of birth 79.0%, treatment outcome 67.1%, and family history 56.6%. Through prompt engineering, the system prompt was reduced from approximately 43 KB to approximately 12 KB, and maxTokens from a range of 6,000–8,000 to 4,000; this yielded an approximately 70% reduction in tokens per document without a measurable loss in field completeness. The authors report that the same setup can be transferred to other centers and languages, and that deduplication is essential (otherwise, multi-patient PDFs silently generate duplicate records). The stated limitation is that completeness does not imply validated accuracy, and field-level agreement with the source document has not yet been quantified.

All WCP 2026 summaries