Detecting Fraud-Associated Characteristics in the Medical AI Literature: A Multi-Signal NLP Framework Reveals Distinct Paper Mill Subtypes

Paper mills increasingly compromise the integrity of the medical artificial intelligence (AI) literature. We developed a pre-registered, multi-signal natural language processing pipeline combining seven feature categories -- tortured phrases, structural formulaicity, AI-generated text markers, citation anomalies, cross-document similarity, co-authorship networks, and geographic metadata -- and applied it to 2,478 medical AI papers (2018-2025) labelled using Retraction Watch data via the Crossref Labs API. A Random Forest-XGBoost ensemble classifier achieved average precision 0.858 and AUC-ROC 0.916 on 5-fold cross-validation, but the most important features were writing quality indicators (vocabulary diversity, reference count) rather than fraud-specific signals, reflecting the dominance of the 2021-2022 Hindawi mass retractions. Retraction subtype analysis revealed distinct fingerprints across fraud types: AI-generated content papers had twice the boilerplate density of other subtypes, while fake peer review papers had the highest co-authorship network density. Unsupervised clustering identified an "author pool" cluster (n=133, 47% retracted) with extreme co-author reuse (0.75 vs 0.11 corpus mean). Among unlabelled papers, 9.1% fell in high or very high risk tiers. Prevalence was broadly uniform across WHO regions (15-20%) and robust to corpus definition. Four pre-registered sensitivity analyses confirmed robustness. The heterogeneity of paper mill operations -- synonym-substitution mills, AI content generators, and peer-review manipulation rings -- demands subtype-aware detection strategies. Code: https://doi.org/10.5281/zenodo.19488868. Pre-registration: https://doi.org/10.17605/OSF.IO/JB4T6.