Reducing invasive RSV diagnostic testing with machine learning: A retrospective validation study

OBJECTIVE: To evaluate whether a machine learning (ML) based screening algorithm can optimize respiratory syncytial virus (RSV) testing while maintaining high diagnostic accuracy in pediatric patients. STUDY DESIGN: We conducted a retrospective analysis of pediatric patients under 2 years old who presented with respiratory infection symptoms and received RSV testing at Yokohama Municipal Citizen's Hospital (2009-2015). The cohort was divided into training (2009-2013; n = 3587) and validation (2014-2015; n = 587) sets. Using patient-reported symptoms and background characteristics from structured electronic questionnaires, we collected clinical symptoms and patient characteristics to build an ML model for predicting RSV testing necessity according to established clinical guidelines, focusing on hospitalized patients and those with underlying conditions. RESULTS: The median age was 11.2 and 11.7 months in the training and validation sets, respectively, with hospitalization rates of 45.4 % and 43.1 %. The ML model showed good performance, achieving a sensitivity of 77.1 % and specificity of 73.4 % in the training dataset, with improved sensitivity (85.1 %) and comparable specificity (71.2 %) in validation. Implementation could potentially reduce unnecessary testing by 77.9 % (98.5 tests annually) for cases requiring hospitalization and 72.9 % (17.5 tests) for patients with underlying conditions, with negative predictive values of 97.0 % and 100 %, respectively. CONCLUSION: This study demonstrates that ML-based screening using symptom data could substantially reduce unnecessary invasive RSV testing while maintaining high diagnostic accuracy. The approach offers promising clinical utility by potentially minimizing patient discomfort and optimizing resource allocation in pediatric respiratory care.