DBNet/EasyOCR converted image-based Thai hospital polysomnography reports to text, followed by task-specific ChatGPT-3.5 extraction. Prompt refinement reduced errors and showed potential for real-world evidence workflows, but numerical, symbol and encoding errors remained common and clinically consequential. The abstract gives no sample size or accuracy estimates.
Key findings
- The pipeline extracted features and task-specific prompting reduced errors. Numerical, symbol and character-encoding errors dominated, but no report count, sensitivity, precision or error rate is provided in the abstract, preventing readiness assessment.
Why this matters globally
Image and scanned records obstruct real-world evidence worldwide. Separating OCR from LLM extraction helps locate errors and design quality controls.
Thai researcher contribution
Mahidol, Ramathibodi and Bangkok University researchers addressed a real Thai sleep-laboratory data bottleneck.
Limitations to consider
The abstract omits dataset size, document diversity, gold-standard accuracy and field-specific performance. ChatGPT-3.5 is version-dependent, one centre limits transfer, and external-service use requires strict privacy governance.