The language of top-tier medical journal abstracts has grown substantially more complex over the twenty-first century, driven by reporting-guideline reform, biomedical specialisation, and — most recently — the rapid diffusion of large language models (LLMs). Yet longitudinal evidence spanning the evidence-based medicine era through the LLM period, at annual resolution, remains absent. This study presents a corpus analysis of 759 abstracts from four flagship biomedical journals (New England Journal of Medicine, Nature, The Lancet, JAMA; 2000–2026) retrieved via the PubMed E-utilities API. Sixteen textual features were computed: readability indices (FKGL, FRE, SMOG), lexical measures (lexical density, mean word length, MATTR-50), and syntactic measures (average sentence length, passive-voice rate, statistical density). Over 26 years, FKGL rose from 14.10 to 17.63 (2025 peak) and FRE fell from 27.34 to 11.34 — a 16-point decline. Passive-voice rate decreased from ~30% to 19.3%. Statistical density followed an inverted-U trajectory, peaking at 6.75 per 100 words (2016) before falling to 1.79 (2025). In the LLM diffusion period (2023–2026), four theoretically motivated hypotheses were tested: all four were empirically supported — mean word length, lexical density, and MATTR-50 reached sequence highs while statistical density reached its nadir, constituting a coherent multi-feature LLM fingerprint. Each major reporting-guideline update produced a detectable step-change with a 1–3 year transmission lag. These findings carry direct implications for EAP curriculum design — where norms calibrated to pre-2010 corpora are demonstrably outdated — for journal editorial policy on readability governance, and for understanding the stylistic consequences of AI-assisted scientific writing.