Toward Automated Regulatory Decision-Making: Trustworthy Medical Device Risk Classification with Multimodal Transformers and Self-Training

TL;DR

This paper tackles automated regulatory risk classification for medical devices by integrating textual narratives and visual imagery using a cross-attentive Transformer and a self-training loop. It introduces modality-specific projections and a cross-modal encoder to fuse text and images into a shared representation, augmented by high-confidence pseudo-labels to expand the training set. Key contributions include a hybrid textual feature pipeline, domain-adapted EfficientNet-B4 image representations with attention, a cross-modal fusion strategy, and a robust self-training protocol, achieving up to $0.904$ accuracy and $0.979$ AUROC on a real-world NMPA-derived dataset. The findings suggest meaningful practical impact for regulatory decision-support and highlight future work on multilingual deployment, structured metadata integration, and explainability to enhance trust and accountability in regulatory AI.

Abstract

Accurate classification of medical device risk levels is essential for regulatory oversight and clinical safety. We present a Transformer-based multimodal framework that integrates textual descriptions and visual information to predict device regulatory classification. The model incorporates a cross-attention mechanism to capture intermodal dependencies and employs a self-training strategy for improved generalization under limited supervision. Experiments on a real-world regulatory dataset demonstrate that our approach achieves up to 90.4% accuracy and 97.9% AUROC, significantly outperforming text-only (77.2%) and image-only (54.8%) baselines. Compared to standard multimodal fusion, the self-training mechanism improved SVM performance by 3.3 percentage points in accuracy (from 87.1% to 90.4%) and 1.4 points in macro-F1, suggesting that pseudo-labeling can effectively enhance generalization under limited supervision. Ablation studies further confirm the complementary benefits of both cross-modal attention and self-training.

Figures (2)

• Figure 1: Multimodal Model Training and Evaluation Pipeline
• Figure 2: Predication Accuracy Rate Across Different Models