Retrospective comparison of ChatGPT-4 treatment recommendations with real-world physician management in newly diagnosed hypertension: a single-centre study

Background – Despite the availability of effective first-line therapies, hypertension remains underdiagnosed and undertreated worldwide. Large language models (LLMs), such as ChatGPT-4, may support guideline-adherent prescribing, yet their role in chronic disease management has not been adequately studied. Objectives – To assess the concordance between ChatGPT-4’s antihypertensive treatment recommendations and real-world physician prescriptions, and to evaluate the association between treatment concordance and short-term blood pressure (BP) control. Methods – This retrospective, single-centre study included 100 newly diagnosed, treatment-naïve hypertensive adults aged 18–65 years. For each patient, anonymized clinical data were retrospectively entered into ChatGPT-4 after the clinical encounter using a standardized prompt based on ESC/ESH and ACC/AHA guidelines. These recommendations were generated retrospectively after the clinical encounter and were not available to the treating physicians, who were therefore unaware of the ChatGPT-4 outputs.ChatGPT-4 recommendations were generated for study comparison only and did not influence physician prescribing or patient management. The primary outcome was concordance between AI and physician treatment decisions. Secondary outcomes included BP control at 30 ± 5 days and the distribution of therapy classes among uncontrolled cases. Statistical analyses included Cohen’s κ, χ² testing, and multivariable logistic regression. Results – ChatGPT-4 and physician treatment choices showed substantial agreement (κ = 0.67). Guideline-recommended dual therapy (e.g. ACEi + thiazide) was selected in 74% of ChatGPT-4 and 67% of physician cases. BP control was achieved in 73% of concordant cases versus 40% of discordant cases. Concordance was independently associated with BP control (adjusted OR = 3.8; 95% CI: 1.5–9.5). Among uncontrolled cases, ChatGPT-4 more frequently recommended fixed-dose combination therapy. Conclusion – ChatGPT-4 demonstrated strong concordance with expert prescribing patterns, and concordant cases had higher rates of early BP control. Because ChatGPT-4 recommendations were generated retrospectively and did not guide treatment, these findings should be interpreted as showing an association rather than causation, and they warrant prospective validation.