Staged identification of CAP in fever patients across epidemic environments: modeling & validation

Diagnosing community-acquired pneumonia (CAP) relies on costly imaging, posing challenges in resource-limited settings. Traditional tools focus on diagnostic tests for clinicians rather than patient use. Additionally, classification of subtypes in traditional Chinese medicine (TCM) lacks criteria. We developed a multimodal fusion model using machine learning algorithms and clinical variables from basic information, medical records, and lab tests to assess CAP risk in fever patients. The model integrates top-performing models via ensemble learning to predict pneumonia probability. We trained on 2,193 visits at Beijing Traditional Chinese Medicine Hospital's fever clinic from Dec 2021 to Dec 2022, and validated on 300 visits from Jan to July 2024. Use unsupervised learning to classify subtypes. The training cohort included 1,781 CAP and similar patients, with 210 in the external validation cohort. CAPs were diagnosed via chest CT. The α model, based on pre-visit medical records, performed well (AUC_internal=0.80, 95%CI 0.77-0.83; AUC_external=0.80, 95%CI 0.71-0.87). The β model added four lab indicators, optimizing performance (AUC_internal=0.93, 95%CI 0.92-0.95; AUC_external=0.81, 95%CI 0.70-0.90). Two models were developed into online calculators. Latent class analysis distinguished Cold/Heat syndrome as subtypes. Despite the single-center, retrospective design, two final models performed good for identifying CAP across epidemic environments.