A Survey around GenAI-Assisted Decision Trees on the UCI Heart Disease Datase

Heart disease prediction studies using the UCI Heart Disease dataset commonlyreport high classification precision while leaving decision structure and robustnesslargely unexamined. This survey analyzes the literature with emphasis on explicitdecision representation and the stability of predictive mechanisms. Published stud-ies are organized into four analytical dimensions: modeling paradigm, decisionrepresentation, role of generative AI, and clinical alignment. Most reported sys-tems implement statistical models, classical machine learning classifiers, ensemblemethods, or deep neural networks evaluated through aggregate performance met-rics and operating as implicit decision systems. Explicit reasoning structures, suchas interpretable decision trees and robustness evaluation under controlled pertur-bations, appear infrequently.This paper extends the survey with an empirical hybrid pipeline in which alarge language model (Gemini 2.5 Flash) participates in three structural decisions:feature prioritization guided by clinical domain knowledge, hyperparameter selec-tion constrained by clinical interpretability requirements, and post-hoc validationof trained tree rules against established cardiovascular medicine. Experimental re-sults demonstrate that the LLM-guided tree achieves comparable predictive accu-racy to an unconstrained CART tree while producing a substantially simpler deci-sion structure. The clinical validation experiment further reveals that standard pre-processing pipelines introduce threshold opacity that renders decision rules clin-ically uninterpretable—a limitation undetectable through accuracy metrics alone.The analysis defines a framework in which generative models participate in theconstruction and validation of interpretable decision trees, treating decision struc-ture, transparency, and robustness as measurable properties of predictive systems.