The discrepancy in reliability and evidence conflict in multimodal data fusion for medical image prediction significantly undermines the accuracy of clinical decision-making. To address this challenge, we propose an Adaptive Evidence-Gated Fusion Network (AEGFN) based on Dempster-Shafer (DS) evidence theory. This framework models the evidence quality and cognitive uncertainty of CT images, image sequences, and clinical data using the Dirichlet distribution. We innovatively introduce an Evidence-Attention Gate (EAG) to dynamically adjust fusion weights for high-conflict modalities (conflict > 0.6), enabling conflict-aware uncertainty compensation. Additionally, a hybrid loss function combining KL divergence regularization with uncertainty-weighted cross-entropy is designed to balance model confidence and generalization. Evaluated on colorectal cancer (656 cases) and radiation pneumonitis (117 cases) datasets for binary classification tasks (predicting patient death and predicting RP occurrence), AEGFN achieves classification accuracies of 95.04% (AUC 0.97) and 82.34% (AUC 0.8312), outperforming the state-of-the-art method DDEF by 0.66% and 1.94%, respectively. This work provides a robust and interpretable solution for multimodal medical prediction, enhancing the reliability of clinical decision support systems.

Radiation pneumonitis (RP), one of the most common and severe complications in locally advanced non-small cell lung cancer (LA-NSCLC) patients following thoracic radiotherapy, presents significant challenges in prediction due to the complexity of clinical risk factors, incomplete multimodal data, and unavailable slice-level annotations in pre-radiotherapy CT images. To address these challenges, we propose a multimodal fusion framework based on Dual-Layer Attention-Based Adaptive Bag Embedding Multiple-Instance Learning (DAAE-MIL) for accurate RP prediction. This study retrospectively collected data from 995 LA-NSCLC patients who received thoracic radiotherapy between November 2018 and April 2025. After screening, Subject datasets (n = 670) were allocated for training (n = 535), and the remaining samples (n = 135) were reserved for an independent test set. The proposed framework first extracts pre-radiotherapy CT image features using a fine-tuned C3D network, followed by the DAAE-MIL module to screen critical instances and generate bag-level representations, thereby enhancing the accuracy of deep feature extraction. Subsequently, clinical data, radiomics features, and CT-derived deep features are integrated to construct a multimodal prediction model. The proposed model demonstrates promising RP prediction performance across multiple evaluation metrics, outperforming both state-of-the-art and unimodal RP prediction approaches. On the test set, it achieves an accuracy (ACC) of 0.93 and an area under the curve (AUC) of 0.97. This study validates that the proposed method effectively addresses the limitations of single-modal prediction and the unknown key features in pre-radiotherapy CT images while providing significant clinical value for RP risk assessment.