Assessing the Feasibility of Processing a Paper-based Multilingual Social Needs Screening Questionnaire Using Artificial Intelligence

Abstract The collection of Social Determinants of Health (SDoH) data is increasingly mandated by healthcare payers, yet traditional paper-based methods pose challenges in terms of cost effectiveness, accuracy, and completeness when manually entered into Electronic Health Records (EHRs). This study explores the application of artificial intelligence (AI), specifically using a document understanding model (Microsoft Azure Document Intelligence) and large language models (OpenAI’s GPT-4 Turbo and GPT-3.5 Turbo), to automate the conversion of paper-based Social Determinants of Health (SDoH) questionnaires into structured, machine-readable formats that could theoretically be incorporated into EHRs. Using a dataset of synthetic and scanned examples, the study compares the performance of the GPT-3.5 and 4 Turbo base models and fine-tuned GPT-3.5 Turbo models on this task. Findings indicate that GPT-4 Turbo outperforms GPT-3.5 Turbo in accuracy and consistency, with fine-tuning enhancing GPT-3.5 Turbo’s performance and consistency in several languages. These results suggest that AI could prove to be an accurate alternative to manual data entry, with important implications for improving how SDoH data is incorporated into EHRs. Future research should address data privacy, security concerns, cost considerations, and the technical aspects of incorporating AI-generated data into EHRs. Description This study explores the application of artificial intelligence (AI), specifically using a document understanding model (Microsoft Azure Document Intelligence) and large language models (OpenAI’s GPT-4 Turbo and GPT-3.5 Turbo), to automate the conversion of paper-based Social Determinants of Health (SDoH) questionnaires into structured, machine-readable formats that could theoretically be incorporated into a patient’s electronic health records (EHRs).