Development and validation of a machine learning-based combined model for malignancy risk prediction in pulmonary subsolid nodules: a dual-center study

Accurate characterization of pulmonary subsolid nodules (SSNs) is important for selecting surveillance, biopsy, or surgery. Existing clinical models were developed in broader nodule populations and may not fully reflect the radiological and biological features of SSNs. This dual-center retrospective study included 225 patients with pathologically confirmed SSNs from January 2018 to January 2024. The development center contributed 167 patients, split into internal training ( n = 118) and internal validation ( n = 49) cohorts, and the external center contributed 58 patients. Ten machine-learning algorithms were compared within the development cohort. The final combined model used four routinely available predictors: spiculation, consolidation-to-tumor ratio (CTR), CYFRA 21 − 1, and nodule diameter. Discrimination, calibration, decision-curve analysis, pairwise model comparison, and sensitivity analyses were performed. The combined model achieved AUCs of 0.932 (95% CI, 0.887–0.977), 0.954 (95% CI, 0.895-1.000), and 0.930 (95% CI, 0.864–0.995) in the internal training, internal validation, and external validation cohorts, respectively. In external validation, the corresponding AUCs for the Mayo, VA, Brock, and PKUPH models were 0.750, 0.560, 0.800, and 0.840. Calibration was evaluated in all three cohorts, and decision-curve analysis showed net benefit across threshold probabilities of approximately 0.12–0.68 in external validation. A four-variable combined model integrating imaging features and CYFRA 21 − 1 showed consistent discrimination across the internal and external validation cohorts. The web calculator may support research-stage risk estimation, but prospective multicenter validation is needed before clinical implementation.