NMPA-MedDevice: A Multimodal Regulatory Dataset of 43,000 Chinese Medical Devices for Risk Classification Research

NMPA-MedDevice is a large-scale regulatory dataset derived from China's National Medical Products Administration (NMPA) Unique Device Identification (UDI) registry (full release: July 2024). It comprises two complementary components:1. **Labeled multimodal subset** -- 1,005 medical devices with product descriptions, regulatory metadata, and manufacturer-assigned risk classification labels (Class I / II / III).2. **Unlabeled text-only corpus** -- Approximately 42,000 unique device records (after deduplication and quality filtering from the full 66,472-record registry) with complete regulatory text fields.Risk class labels are programmatically derived from the ninth character of the NMPA registration number, ensuring objective and reproducible annotation:- `1` = Class I (low risk)- `2` = Class II (moderate risk)- `3` = Class III (high risk) ### Image ModalityFor privacy and copyright reasons, raw product images are **not** distributed with this dataset. Instead, we provide **pre-extracted feature vectors** using standard deep learning backbones, following established practice in the scientific community (cf. UNI2-h features [Chen et al., Nature Medicine 2024], LAION-5B embeddings [Schuhmann et al., NeurIPS 2022]):Text features: BERT-base-Chinese [CLS] token embeddings (768 dimensions per sample)Image features:EfficientNet-B5 average-pool embeddings (2,048 dimensions per sample)These pre-computed embeddings enable full reproducibility of all downstream classification experiments reported in the companion ESWA paper. Users requiring original images may use the provided `scripts/image_retrieval.py` to download them directly from manufacturer websites.