Feature extraction from real-world polysomnography reports of obstructive sleep apnea cohort using large language...

Electronic health records (EHRs) are a valuable resource for generating real-world evidence. However, their utilization can be challenging as these reports are largely unstructured texts stored in image formats or accumulated scans or images, thereby hindering efficient data feature extraction. With the development of computer vision and large language models (LLMs), there is a growing opportunity to explore their application in overcoming these challenges. This paper explores the potential use of computer vision and LLMs to extract data from text-containing images of polysomnography (PSG) reports obtained from the sleep laboratory center of a tertiary care hospital in Thailand. We utilized a two-phase approach: (1) extracting text from image-based PSG reports using Differential Binarization Network (DBNet) within EasyOCR Python library, and (2) deriving feature values from the extracted text using ChatGPT-3.5 through task-specific prompt strategies. Performance was measured across different stages of the conversion process. Results show that computer vision and LLMs have the potential to substantially enhance the efficiency of feature extraction for evidence synthesis. The most common errors encountered in both phases were numerical, symbol, and character encoding errors. ChatGPT-3.5 reliably extracted features from sleep reports, with further error reduction achieved through improved prompt strategies. Although promising, we emphasize the need for extensive testing across diverse document qualities and conditions to fully understand the challenges and pitfalls of using computer vision and LLMs for feature extraction in real-world scenarios.