Machine Learning Application: Congenital Syphilis Classifier

Abstract

Objectives: Congenital syphilis (CS) is resurging in Thailand, which is challenging the nation's elimination status. Current surveillance relies on manual expert review of paper-based case investigation forms (CIFs), resulting in significant backlogs which hinder timely public health intervention. This study aimed to develop and evaluate an automated system using Large Language Models (LLMs) to digitize and classify CS cases.

Methods: We conducted a retrospective diagnostic study using 143 validated CIFs from Eastern Thailand during October 2024 to October 2025. The system utilized Google Gemini 2.5 Pro and Flash models for optical character recognition (OCR) and a rule-based algorithm for case classification. Cost was estimated by Google Gemini application programming interface (API) pricing. Performance was benchmarked against an expert committee's consensus.

Results: The system efficiently reduced processing time from months to under five minutes at a cost of approximately $0.006 per case. It achieved a data extraction accuracy of 94.2% with a Character Error Rate (CER) of 4.32% for complex handwriting. The overall classification accuracy was 80.0%, with sensitivity 83.3% and specificity of 75.0%. Automated "Red Flags" detection identified inadequate maternal treatment (69.7%) and late antenatal care (32.9%) as primary drivers for confirmed and probable cases.

Public Health Recommendations: The LLM and rule-based classifier offer a scalable "Paper-to-Digital Bridge" for resource-limited settings. While human oversight remains necessary for complex cases, the system functions as an effective high-volume triage tool, transforming retrospective surveillance into real-time actionable intelligence to support the elimination of mother-to-child transmission.