An interpretable multi-task whole-slide histopathology AI model for non-small cell lung cancer: Cross-cohort...

Tumour-node-metastasis staging does not fully explain prognostic heterogeneity in non-small cell lung cancer. We evaluated whether haematoxylin-and-eosin whole-slide images could estimate histological subtype, pathological stage probabilities, survival risk and spatially grounded biological associations.

SparseAGE-MTL, a weakly supervised multi-task multiple-instance learning model with a shared projection-topology encoder and endpoint-specific heads, was trained and benchmarked in 954 The Cancer Genome Atlas cases using seven pathology feature spaces and 18 comparator models. External evaluation used 948 tissue-microarray and 324 whole-slide cases. Attention maps were co-registered with 10x Visium spatial transcriptomics and integrated with bulk transcriptomics, immune-infiltration estimates and ESTIMATE scores. Analyses included paired model comparisons, false-discovery-rate correction, Cox models, calibration assessment and decision curve analysis.

In the CONCH feature space, SparseAGE-MTL achieved 93.73% accuracy, 98.19% area under the receiver-operating-characteristic curve and 93.08% F1-score for adenocarcinoma/squamous cell carcinoma classification in internal benchmarking; external area-under-the-curve values were approximately .91 and .82. Stage estimation had lower discrimination, with external overall area under the curve approximately .70 and cohort-dependent calibration. Risk-score-defined groups differed in overall survival in both histological subtypes and showed similar external trends. High-attention regions were enriched at tumour-stroma or tumour-immune interfaces and were associated with B-cell, fibroblast, C1QC, COL1A1, epithelial-mesenchymal transition, metastasis and hypoxia signals. Higher risk cases showed malignant pathway activation, lower immune/stromal scores, higher tumour purity and subtype-specific immune/stromal differences. Adding the risk score to the clinical model increased external pooled concordance index from approximately .620 to .672.

In retrospective cohorts, SparseAGE-MTL generated subtype-classification, stage-probability and survival-risk outputs from routine pathology images. Subtype classification had higher numerical performance than stage estimation. Survival-risk and attention outputs were associated with outcome and spatial/transcriptomic features, but prospective, treatment-annotated validation is required before clinical use.

SparseAGE-MTL is a weakly supervised multi-task MIL framework that jointly performs NSCLC subtype classification, stage prediction, and survival risk estimation from routine H&E slides using only slide-level labels. The model achieves stable competitive performance across seven feature spaces and 18 comparators, with external validation demonstrating robust generalization to independent WSI and TMA cohorts. Attention hotspots co-localize with tumor-stroma/immune interfaces and spatial transcriptomic signatures of EMT, hypoxia, and C1QC/COL1A1 enrichment, providing biologically grounded interpretability. High-risk groups exhibit activated malignant pathways, lower immune/stromal scores, higher tumor purity, and incremental prognostic value beyond clinical variables.