AI-Driven Early Detection of Severe Influenza in Jiangsu, China: A Deep Learning Model Validated Through The Design of Multi-Center Clinical Trials and Prospective Real-World Deployment

Influenza causes about 650,000 deaths worldwide each year, and the high mortality rate of severe cases is closely related to subjective bias in clinical assessment and inconsistent diagnostic and treatment standards. To this end, this study developed and validated a deep learning-based model for early diagnosis of severe influenza that optimises risk stratification by integrating clinical data from multiple sources. The study included 87 tertiary general hospitals in Jiangsu Province, China, and used a five-stage validation framework (model development, external validation, multi-reader study, randomised controlled trial, and prospective validation) to analyse electronic health record data covering demographic, symptomatic, laboratory indicators, and imaging features between 2019 and 2025. Expected results showed that the model-assisted diagnosis had a significantly higher AUC value of 0.18 (95% CI: 0.14-0.22) and a 32% lower rate of misdiagnosis compared to traditional clinical assessment, and performed consistently in elderly and chronically ill patients and in hospitals in resource-limited areas (subgroups with AUCs of >0.82 in all cases). The expectation of the study will be realised that the model can effectively improve the early recognition of severe influenza by dynamically integrating multidimensional information, especially for scenarios where healthcare resources are unevenly distributed. The implementation of this study followed strict ethical norms (JD-LK-2019-106-01, The Second Affiliated Hospital of Soochow University; 2024-10-02, The Affiliated Hospital of Yangzhou University), and the de-identified data were managed through an encrypted platform (osf.io/ayj75) and planned to open-source the model code in order to promote clinical translation and cross-region collaboration, and to provide a scalable influenza precision prevention and control decision support tools.

Trial: ChiCTR2000028883

Registration DOI: https://doi.org/10.17605/OSF.IO/SC93Y