Exploratory multimodal prognostic subtyping of severe pneumonia using CT radiomics and baLF mNGS: a single-center retrospective derivation cohort study

Severe pneumonia is etiologically heterogeneous, and the prognostic value of integrating CT radiomics with bronchoalveolar lavage fluid (BALF) metagenomic next-generation sequencing (mNGS) remains uncertain. We conducted an exploratory analysis to derive multimodal prognostic subtypes and compare the apparent prognostic discrimination of single-modality models within a single-center derivation cohort. This single-center retrospective derivation study included 185 patients with severe pneumonia who underwent chest CT and BALF mNGS. Consensus clustering was used to derive exploratory subtypes based on selected CT radiomics features, mNGS findings, Acute Physiology and Chronic Health Evaluation II (APACHE II) scores, and infection indicators. Separate Cox models based on CT radiomics alone, mNGS variables alone, and the APACHE II score alone were evaluated within the same derivation cohort. A randomly selected subset of 20 patients underwent repeat region-of-interest segmentation for intra-rater reproducibility assessment. Mortality associations were subsequently assessed among the derived subtypes. Using a performance–parsimony criterion, the top-20 ranking cutoff, which yielded 10 features shared by Random Forest, XGBoost, and LightGBM, was selected for subsequent analyses. Within the derivation cohort, the CT radiomics-only model showed higher apparent prognostic discrimination (AUC, 0.912; 95% CI, 0.870–0.953) than the mNGS-only model (AUC, 0.627; 95% CI, 0.544–0.710) and the APACHE II-only model (AUC, 0.719; 95% CI, 0.644–0.794). In the 20-patient repeat-segmentation subset, 83 of 107 features (77.6%) met the prespecified ICC threshold of ≥ 0.75, and all 10 features retained in the final model were stable (ICC range, 0.779–0.981). Consensus clustering generated three exploratory multimodal subtypes. Compared with Subtype 1, Subtype 2 (HR, 1.59; 95% CI, 1.01–2.50) and Subtype 3 (HR, 1.83; 95% CI, 1.09–3.08) were associated with higher mortality risk. Descriptively, Gram-positive, Gram-negative, and fungal/viral mNGS detections were enriched in Subtypes 1, 2, and 3, respectively; these distributions do not establish pathogen-specific causality. In this single-center retrospective derivation cohort, CT radiomics showed higher apparent prognostic discrimination than mNGS or the APACHE II score. The integrated analysis generated three exploratory, etiologically enriched multimodal subtypes with hypothesis-generating associations with mortality. These subtypes should be regarded as research-oriented risk-stratification and diagnostic-prioritization phenotypes, but should not be used alone to guide clinical decisions. Prospective multicenter studies with independent external validation and clinical-impact evaluation are required before clinical implementation.