From raw hospital records to an AI-ready surveillance dataset: a FAIR-compliant data pipeline for healthcare-associated infection research in a Chinese district hospital

The proliferation of hospital surveillance data in China has not been matched by publicly available, machine-readable datasets suitable for artificial intelligence (AI) research. We describe the design, implementation and validation of a reproducible Python pipeline that transforms 39 months of raw, Chinese-language healthcare-associated infection (HAI) surveillance records from a 1,100-bed district general hospital into a de-identified, FAIR-compliant, AI-ready benchmark dataset. The pipeline addresses five real-world data-engineering challenges: heterogeneous Excel schemas with month-to-month format drift, mixed Chinese free-text fields requiring domain-specific bilingual translation, composite risk-factor strings requiring decomposition into 19 binary flags, incomplete temporal denominators, and privacy-preserving de-identification with k-anonymity enforcement. The resulting dataset (1,240 HAI episodes, 604 culture-positive, 57 variables, 115-entry codebook) is deposited in Zenodo (DOI: 10.5281/zenodo.20725167) under CC-BY 4.0 with full provenance metadata, representing the first publicly available, FAIR-compliant HAI surveillance dataset from a Chinese district-level hospital. Data completeness was high (>95% for 54 of 57 variables), bilingual translation achieved 100% coverage of 156 categorical values with 98.7% semantic-equivalence verification of the complete dictionary, and automated risk-factor decomposition demonstrated sensitivity 0.92–1.00 and specificity >0.97 against human expert annotations. The pipeline code is released under the MIT licence to enable replication at other Chinese hospitals facing similar data-harmonization challenges. As a proof-of-concept demonstration of dataset utility for downstream AI research, we benchmark three machine-learning classifiers on antimicrobial-resistance phenotype prediction, achieving area under the receiver operating characteristic curve (AUROC) of 0.74–0.82 across the three standard classifiers (best: random forest 0.82, 95% CI 0.74–0.89), demonstrating that the dataset contains learnable signal using standard off-the-shelf methods despite a modest sample size. The publicly accessible dataset, codebook, and pipeline enable: (1) development of AI models for HAI prediction without requiring access to restricted hospital data, (2) benchmarking of new algorithms against a transparent baseline, and (3) replication of the pipeline methodology at other Chinese healthcare institutions.