NMPA-MedDevice: A Multimodal Regulatory Dataset of 52,000 Chinese Medical Devices for Risk Classification Research

The subset has been rebuilt rather than patched, and further problems in the earlier files were corrected (see codebook.pdf).</p> <p>The dataset covers the 66,472 records of the NMPA UDI full release of 1 July 2024; a cleaned corpus of 52,251 of them with a ninth-character risk label, the recommended route-derived label and a non-standard-identifier flag; a rebuilt image-linked subset of 1,013 devices (Class I/II/III = 50/495/468, route-derived) with precomputed BERT text and EfficientNet-B5 image embeddings and UDI-grouped five-fold cross-validation indices; an external temporal validation set of 300 later-registered devices; and records of every device attempted during sampling.</p> <p>This deposit contains no NMPA registry records.

The NMPA website statement asks that its content not be republished in its original form and that content supplied by other units be republished only with their authorisation, so records are identified by their UDI-DI and rebuild_from_official.py retrieves their registry fields from an official NMPA full release (udi.nmpa.gov.cn/download.html). The July 2024 release is no longer offered by the NMPA; rebuilding from the release of 1 May 2026 retrieved 99.3% of the records, 92.0% with an unchanged description.