Cross-attention AI predicts GIST imatinib response with 0.99 AUC
A multimodal deep learning framework hits 0.99 AUC internally but struggles to generalize externally.
A team of researchers from multiple European medical centers has developed a cross-attention multimodal deep learning framework that integrates computed tomography (CT) imaging and clinical variables to predict patient response to neoadjuvant imatinib in gastrointestinal stromal tumors (GIST). The model was built using a retrospective cohort of 935 patients for pretraining and 213 patients for prediction from four tertiary centers between 2000-2023. Two training strategies were evaluated: self-supervised pretraining with low-rank adaptation and training from scratch, with hyperparameters optimized via SMAC3. Internal cross-validation yielded AUC scores up to 0.99, demonstrating the model's ability to accurately classify responders from non-responders within the training distribution.
However, external testing revealed a sharp drop in performance (AUC 0.60-0.63), highlighting generalization challenges common in medical AI. Clinical-only models achieved a moderate AUC of 0.66, while imaging-only models ranged from 0.56-0.66. Explainability analyses identified significant differences in feature importance between responders and non-responders, including biomarkers CD117, BRAF, PDGFRA, along with age, sex, disease status, and comorbidities (FDR-adjusted Pā¤0.036). The study underscores the potential of cross-attention mechanisms for improving treatment response prediction while also revealing the limits of current multimodal approaches when applied to broader, unseen patient populations.
- Cross-attention model achieved up to 0.99 AUC internally but only 0.60-0.63 AUC on external validation, indicating generalization gaps.
- Key predictive features identified included KIT mutations (69% vs 56.7%), larger tumors (112 vs 89 mm), and higher mitotic index (3 vs 0).
- Study used 1,148 patient records across four centers, with two training strategies and SMAC3 hyperparameter optimization.
Why It Matters
Improves imatinib response prediction for GIST patients, but external validity remains a critical hurdle for clinical deployment.