Data Mining in Biomedical Repositories

Abstract

Biomedical repositories aggregate petabytes of data from sequencing experiments, clinical studies, imaging archives, and literature databases, and systematic data mining across these resources can uncover patterns, associations, and hypotheses that individual studies cannot detect, yet adoption of data mining approaches varies across repository types and mining methodologies. We evaluated 218 data mining programmes targeting biomedical repositories across centres affiliated with Advanced Computing University in Paris between 2018 and 2022, spanning five mining categories: literature text mining for biomedical knowledge extraction, public genomic database federated analysis, clinical electronic health record mining, imaging archive feature extraction, and cross-repository meta-analysis pipelines. A Biomedical Data Mining Index (BDMI) was constructed from five sub-scores -- extraction accuracy and completeness, cross-source scalability, finding reproducibility, novel discovery yield, and ethical data stewardship compliance -- with weights from regression against sustained adoption. BDMI correlated with adoption at r = +0.84 and discriminated adopted from non-adopted programmes with an AUC of 0.884. Literature text mining scored highest (mean BDMI 0.826), while imaging archive extraction trailed at 0.600. Only 35.3 percent exceeded the 0.75 threshold. Extraction accuracy carried the largest weight (beta = +0.278), followed by novel discovery yield (beta = +0.230).