Retrospective development and external validation of machine learning models for COVID-19 severity prediction in lung and hematological malignancies: evidence of subtype-specific feature selection bias in joint training

Le résumé fourni par la source

Patients with lung or hematological malignancies face an elevated risk of severe COVID-19 outcomes. While machine learning frameworks offer predictive utility, training a single joint clinical-biological model on heterogeneous cancer populations can introduce structural feature selection bias, limiting clinical generalizability. Utilizing a multicenter retrospective cohort (n = 340) for model development and an independent external cohort (103) for validation, we evaluated an end-to-end machine learning pipeline. Initially, a joint 7-variable signature—comprising age, ECOG performance status, lymphopenia, BMI, albumin, creatinine, and SARS-CoV-2 serology—was analyzed using logistic regression (LR) and random forest (RF) algorithms. Phenotypic deconstruction via cluster-level ablation and cancer-type-stratified optimal feature searches were subsequently executed to evaluate population-specific feature mechanics. The primary joint LR model achieved an optimism-corrected internal AUC of 0.75 (95% CI: 0.69–0.79) and an external validation AUC of 0.72 (95% CI: 0.62–0.82). However, cluster ablation and subgroup stratification revealed a critical biological divergence between the cancer types. The joint model's feature selection prioritized variables that acted as predictive noise for specific populations; notably, the apparent predictive contribution of albumin and creatinine in the joint model diminished or reversed after subtype-specific stratification, suggesting these variables largely reflect mixed-population effects rather than robust subtype-specific biological signal; their removal improved lung cancer model performance (albumin ΔAUC +0.043, creatinine +0.036) and did not materially benefit hematological patients either. The two cancer-type-optimal signatures shared age, ECOG performance status and lymphopenia, but diverged in their remaining components — BMI and albumin for hematological patients versus diabetes and immunodepression status for lung patients — indicating that the joint model's discriminative variables are only partially transferable across subtypes. While a unified 7-variable framework provides acceptable predictive baseline performance, joint training on mixed cancer populations introduces systematic feature selection bias. These exploratory findings suggest that cancer-type-stratified models may better capture subtype-specific biological signatures than a single joint model; however, independent prospective validation with pre-specified subtype stratification is required before any clinical deployment. Not applicable.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.