#3420 Application of ChatGPT 4o to implement local registries in low-resourced countries. A proof of concept for the Jamaican Registry Initiative

Abstract Background and Aims Middle-income and low-income countries face substantial challenges to build local registries for chronic and end-stage kidney disease (CKD/ESKD) patients that allow to characterize disease and build evidence to inform policy makers. The primary challenges stem from limited access to technology and inefficient communication systems. As a result, the use of printed forms to collect data via handwritten registries remains a more feasible option in healthcare settings that lack advanced technological infrastructure. Conversely, mobile and cellphone technologies are increasingly being leveraged to address communication gaps and facilitate clinical research in healthcare settings. In a proof-of-concept study, we aimed to use ChatGPT to extract patient information from pre-designed hand-written medical record forms and to transfer the data to a database. Method We used a pre-designed patient information form from the Nephrology Department at Grenada General Hospital in Grenada and the University of West Indies Mona in Jamaica. The form was created to establish a local, national, and international CKD Caribbean registry. Each form comprised 28 data fields, such as name, address, diagnosis, etc. We distributed the form to healthy volunteers and requested them to fill out the form in English with hand-written sham data. Photographs of the filled-out forms were then imported into ChatGPT, version 4o, and a request for data extraction was prompted. The ChatGPT generated data extracts were then exported into an Excel spreadsheet. To assess the accuracy of the process, the extracted data were then compared with the original hand-written data. We assessed the global discrepancy rate between original form and Excel data base entry. In addition, we categorized discrepancies by four groups: a) Letters, b) Numbers, c) Special characters/symbols, and d) Checkbox discrepancies. Descriptive analyses were done with SAS On Demand for Academics 3.1.0. Results Twenty-two forms with a total of 616 entries (= 22 x 28 data fields) were evaluated. Discrepancies between original and abstracted data occurred in 48 (8%) instances (Fig. 1). The median discrepancy count per form was 2 (interquartile range: 2). Most frequent discrepancies occurred with numbers (54% of all discrepancies), followed by checkbox discrepancies (23%; Fig. 2). Conclusion The extraction of hand-written medical data from pre-defined medical record forms using ChatGPT showed a satisfactory performance in English language with a median error rate of 8%. Human error research indicates spreadsheet cell entry error rates between 1% and 5%. However, these studies did not consider handwritten source data but entry of printed data into spreadsheets (see panko.com for extensive literature and discussion). Additionally, performance may vary according to language and alphabet used. Efforts to improve the writing of numbers by hand and attention to detail when checking boxes are important steps to improve accuracy. Adapting the available resources for the establishment of local registries in low-resourced countries is key for collecting evidence on kidney disease in disadvantaged areas.