Phenotype-aware explainable machine learning and process optimization for individualized risk stratification in glioma

Abstract

Glioma prognosis is heterogeneous, but risk communication often compresses molecular state, tumor biology, and clinical context into one opaque score. We tested whether an explainable, phenotype-aware workflow could improve 2-year mortality stratification and reveal non-obvious biological signals.We analyzed public CGGA mRNA-seq cohorts: mRNAseq_693 for development and mRNAseq_325 for external validation. Expression was summarized into 10 prespecified modules using 81 genes and combined with age, grade, IDH, 1p/19q, MGMT, treatment, and recurrence. A regularized gradient-boosting classifier was compared with a clinical logistic-regression baseline using 5-fold out-of-fold evaluation. The validation cohort was held out. Discrimination, Brier score, calibration, decision curves, SHAP explanations, Kaplan–Meier curves, and an immune-balance/IDH interaction were evaluated. Of 693 development and 325 validation samples, 632 and 296 were eligible for the landmark endpoint. The integrated model achieved out-of-fold and validation AUCs of 0.839 and 0.876, compared with 0.807 and 0.862 for the clinical baseline. Validation Brier score improved from 0.155 to 0.144. IDH, grade, proliferation, T-cell, recurrence, and myeloid features led global explanations. Contrary to a simplistic protective interpretation, higher immune balance was associated with higher adjusted hazard (β = 0.149, P = .001), without significant IDH effect modification (P = .671). Decision curves showed positive validation net benefit across plausible thresholds. A compact biological model improved external discrimination and supplied patient-level explanations. The adverse immune-linked signal is hypothesis-generating, not evidence that immune activation causes treatment resistance. Prospective multimodal validation is required before deployment.