Integrating multimodal data and deep learning for functional assessment and rehabilitation prediction after cerebral hemorrhage.
Authors
Affiliations (5)
Affiliations (5)
- Department of Rehabilitation, Shengjing Hospital of China Medical University, Shenyang, China. Electronic address: [email protected].
- Department of Rehabilitation, Shengjing Hospital of China Medical University, Shenyang, China; Department of Physical Medicine and Rehabilitation, The Second Clinical College, China Medical University, Shenyang, Liaoning, China. Electronic address: [email protected].
- Department of Rehabilitation, Shengjing Hospital of China Medical University, Shenyang, China. Electronic address: [email protected].
- Department of Rehabilitation, Shengjing Hospital of China Medical University, Shenyang, China. Electronic address: [email protected].
- Department of Rehabilitation, Shengjing Hospital of China Medical University, Shenyang, China; Department of Physical Medicine and Rehabilitation, The Second Clinical College, China Medical University, Shenyang, Liaoning, China. Electronic address: [email protected].
Abstract
Intracerebral hemorrhage (ICH) is a leading cause of long-term disability, particularly in China. Post-stroke motor recovery exhibits considerable heterogeneity, presenting substantial challenges for clinicians in establishing realistic goals. While integrative AI paradigms have shown success in other complex medical domains, existing models often rely on single-modality data, limiting their predictive accuracy and clinical utility. This study aims to develop and validate a multimodal predictive model integrating CT imaging, clinical data, and rehabilitation assessments to simultaneously predict motor recovery and global rehabilitation outcomes following cerebral hemorrhage. We conducted a retrospective study involving 739 patients (315 for motor function prediction and 424 for rehabilitation outcome assessment) who received rehabilitation therapy after cerebral hemorrhage. To predict motor function, we constructed a late-fusion deep learning model leveraging 3D-DenseNet for CT neuroimaging and Multi-Layer Perceptron (MLP) for clinical and laboratory features. To predict rehabilitation outcomes, a Gradient Boosting Decision Tree (GBDT) model was developed and validated using 5-fold and 10-fold cross-validation, comparing it against other machine learning algorithms, including SVR, Random Forest and AdaBoost. Model performance was assessed using metrics including AUC and R². Additionally, univariate and multivariate regression analysis were performed to identify significant factors influencing motor recovery and rehabilitation outcomes. A total of 739 patients were included. The multimodal fusion model achieved an AUC of 0.856 (95 % CI: 0.741-0.971) and an F1 score of 0.897 (95 % CI: 0.819-0.975), significantly outperforming the imaging-only (AUC: 0.833) and clinical-only (AUC: 0.749) models. For rehabilitation outcome prediction, the GBDT model achieved an R<sup>2</sup> of 0.849 (95 % CI: 0.803-0.887), demonstrating superior stability and accuracy over other models. Additionally, multivariate analysis revealed that serum albumin (ALB), neutrophil percentage (NEUT%), triglycerides (TG), and thrombin time (TT) were independent predictors of motor recovery, while age, admission mBI, and time to start rehabilitation significantly influenced functional outcomes. This study confirms that a multimodal deep learning framework integrating routinely available CT imaging and clinical biomarkers provides high predictive value for simultaneously forecasting motor recovery and global functional outcomes after ICH. This proof-of-concept approach offers a reproducible, data-driven tool for early risk stratification, facilitating the formulation of individualized rehabilitation strategies and optimizing resource allocation in clinical workflows.