Multimodal Digital Twin Framework for Personalized CyberKnife Dosimetry and Survival Prediction in Non–Small Cell Lung Cancer

Precision radiotherapy and patient-specific survival prediction in non–small cell lung cancer (NSCLC) demand computational frameworks that are accurate, interpretable, and reproducible. This study introduces a comprehensive digital twin framework for CyberKnife dose prediction and survival estimation, integrating heterogeneous clinical, radiomic, and genomic data through a synergy of classical machine learning, deep learning, and ensemble modeling. The framework begins with a Ridge–Long Short-Term Memory (LSTM) hybrid model that combines the interpretability of linear regression with the temporal–spatial learning capability of recurrent networks. Subsequent enhancements incorporate multilayer perceptrons (MLP) and XGBoost to address structured data constraints and improve generalization through ensemble and bridge stacking architectures. Furthermore, multimodal fusion of radiomic and genomic data via late fusion emphasizes the prognostic value of molecular features while exposing the limitations of naïve averaging strategies. Rigorous preprocessing, multimetric evaluation, and 3D synthetic dose visualizations ensure model transparency, reproducibility, and clinical applicability. Comparative analyses reveal that model performance is modality-dependent: LSTM excels in temporal dependency modeling, while MLP and XGBoost yield superior results for tabular data; ensemble approaches consistently enhance generalization and resilience to outliers. Although phantom-based dosimetry validates the spatial feasibility of dose prediction, future integration with real patient dose–volume histograms (DVHs) is expected to further strengthen clinical translation. Overall, this work establishes a robust, extensible, and clinically aligned digital twin framework that advances precision radiotherapy planning, survival modeling, and multimodal biomedical data integration for NSCLC patient care.