Interpretable multimodal deep learning for predicting spontaneous intracerebral hemorrhage progression: a multicenter study

Hematoma expansion (HE) is a critical predictor of poor outcomes in patients with spontaneous intracerebral hemorrhage (ICH), yet reliable early prediction tools remain limited. This study aimed to develop an interpretable deep learning model for predicting the risk of HE after initial medical intervention in acute ICH. We retrospectively reviewed 1353 patients with ICH from three medical centers. A total of 721 patients were included and divided into a training cohort ( n = 319) and an external validation cohort ( n = 402). Clinical variables, non-contrast CT (NCCT) images, and manually segmented hematoma regions were collected. Three predictive models were developed: a machine learning model, a deep learning model, and a multimodal deep learning model. Model performance was evaluated using AUC, calibration curves, and decision curve analysis. SHAP and Grad-CAM were applied to enhance interpretability. The multimodal deep learning model achieved the best performance, with an AUC of 0.94 (95% CI 0.91–0.96) in the training set and 0.85 (95% CI 0.78–0.91) in the external validation set. It outperformed both the machine learning model (AUC = 0.81) and the deep learning model (AUC = 0.62). The deep learning model's limited generalizability reflects scanner heterogeneity and the inherent difficulty of predicting HE from NCCT images alone, underscoring the value of multimodal integration. SHAP and Grad-CAM provide quantitative and visual explanations of both baseline and early therapeutic determinants. The interpretable multimodal deep learning model reliably predicts HE in acute ICH, outperforming traditional models. It may facilitate early risk stratification and intervention to improve patient outcomes. Prospective validation is warranted to confirm its clinical effectiveness in improving patient outcomes.