Clinical variables and quantitative CT radiomics for mortality risk stratification after acute exacerbation of idiopathic inflammatory myopathy-associated interstitial lung disease

Background Acute exacerbation (AE) is the most lethal event of idiopathic inflammatory myopathy-associated interstitial lung disease (IIM-ILD). This exploratory study compared clinical, radiomics, and combined predictor strategies across five survival-learning algorithms for survival risk stratification. Methods We retrospectively analyzed 78 patients with AE-IIM-ILD and divided them into a training cohort (n = 54; 20 deaths) and a held-out internal testing cohort (n = 24; 9 deaths), stratified by observed mortality event status. Three predictor strategies were evaluated: a clinical strategy comprising PaO 2 /FiO 2 ratio, neutrophil percentage (NEU%), and arthritis-related symptoms; a radiomics strategy comprising two selected radiomics features; and a combined strategy comprising PaO 2 /FiO 2 ratio, arthritis-related symptoms, and a composite radiomics score (Rad-score). Each strategy was evaluated using five survival-learning algorithms. Harrell’s concordance index (C-index) was the primary discrimination metric, with mean time-dependent area under the curve (AUC) used as a secondary measure. Results During the 1-year follow-up, 29 deaths were observed. Performance rankings varied across predictor strategies, algorithms, and evaluation metrics. The combined-Random Survival Forest configuration demonstrated the highest training-cohort C-index among the 15 configurations. Its mean time-dependent AUC was 0.923 (95% CI: 0.860–0.970) in the training cohort and 0.915 (95% CI: 0.736–0.986) in the testing cohort, with corresponding C-indices of 0.874 (95% CI: 0.803–0.940) and 0.842 (95% CI: 0.673–0.971), respectively. The model stratified patients into high- and low-risk groups with distinct observed survival trajectories. Conclusion In this exploratory cohort, integrating selected clinical variables with a quantitative CT Rad-score using RSF showed promising numerical discrimination and risk-stratification performance. These findings suggest that clinical and quantitative imaging information may provide complementary prognostic information after AE-IIM-ILD.