Introduction: Diagnosing acute rheumatic fever (ARF) remains challenging, with delays often leading to missed opportunities for timely evaluation. To address this issue, a chatbot-based screening tool was developed to identify children at high risk of ARF and facilitate referrals for echocardiogram screening.
Methods: The Chatbot's clinical questions were designed for completion by healthcare professionals or family members via WhatsApp at primary care centers. A retrospective analysis applied these questions to 1,109 children from the ARC project across Brazil, Malawi, Timor-Leste, and Pakistan. Children classified as high risk by the Chatbot were compared to the gold standard ARF diagnosis, determined through ARC adjudication based on imaging, clinical history, and lab findings.
Results: Among 1,109 children, the chatbot flagged 164 for possible ARF (15%). A combination of three key symptoms—suspected chorea, joint pain, and mood changes—along with fever and pain, was strongly associated with an ARF diagnosis. The model demonstrated good discriminatory performance, with a C-statistic of 0.802 (95% CI: 0.766–0.839), sensitivity of 86.6%, and specificity of 64.3%. While coefficients varied across countries, suspected chorea consistently showed the strongest association with ARF (OR: 11.70, p < 0.001), followed by joint pain (OR: 4.26, p < 0.001) and mood changes.
Conclusion: The model effectively discriminates ARF cases, with key variables consistent across countries. Further refinement maybe needed to improve sensitivity for non-chorea dominant presentations with additional subanalysis of performance underway. Overall, the tool shows significant potential for broader application, with opportunities for additional adjustments.