Accurate classification of medical device risk levels is essential for regulatory oversight and clinical safety. Manual classification is labour-intensive and prone to inconsistency, creating bottlenecks in regulatory processes. To address this, we introduce a Transformer-based multimodal framework trained on a real-world corpus of 1,005 Chinese NMPA-registered devices with paired narratives and product images. Our model integrates textual and visual modalities through a cross-attention mechanism and employs a confidence-based self-training loop to enhance generalization under limited supervision. Under 5-fold stratified cross-validation, the best multimodal configuration (SVM with self-training) achieves 91.6% accuracy, 98.86% AUROC, and 87.5% F1, substantially outperforming unimodal baselines (78.4% accuracy text-only; 51.5% accuracy image-only). Ablation studies confirm the complementary benefits of cross-modal attention and robustness analyses show that text-only predictions remain competitive when images are unavailable. Beyond technical gains, we highlight applications in AI-assisted regulatory triage, compliance verification, and support for UDI and GMDN-based device matching within the NMPA context. While validated on Chinese regulatory data, the framework is not assumed to generalize to FDA or EMA contexts without jurisdiction-specific retraining and validation. Our findings suggest that domain-adapted multimodal models, when designed with transparency and calibration, can serve as supporting tools for pre-compliance triage within the originating regulatory jurisdiction.
Han et al. (Sun,) studied this question.