Typical Angina, Reversed: A Chest-Pain Coding Error in the UCI Heart Disease Data That Has Persisted Since 1990
Year: 2026
Type: preprint
Abstract: The UCI Heart Disease data are among the most widely used benchmark datasets in machine learning. We show that their chest-pain variable is stored in reverse of its documentation. At all four sites, the code documented as "typical angina" (1) marks patients without chest pain, and the code documented as "asymptomatic" (4) marks patients with typical angina. Two independent checks establish this: (i) the chest-pain table and regression coefficients of the original 1986 Cleveland study are reproduced only under the reversed reading; (ii) in the Budapest, Swiss and Long Beach files, 93–96% of patients coded 4 have all three classical angina features. A 1990 word-labeled Cleveland file already uses the reversed reading, and Statlog (Heart), a Cleveland subset, inherits it. So do OpenML, PMLB and the most-downloaded Kaggle copies (about 676,000 downloads for the two largest). In a pre-specified audit of a random sample of 150 papers (124 readable), 32 state a direction for a named chest-pain category. In 25 of them (78%; 95% CI 61–89%) the direction is clinically backwards, nine explain it clinically, and none identifies the reversal. Conventional models are essentially unaffected, which is why the error went unnoticed. Their explanations are reversed, and language-model classifiers that read the labels as text are misled. With the documented labels, three open-weight models lose 0.12–0.33 AUC given age, sex and chest-pain type, and two of them lose 0.016–0.077 AUC given all 13 predictors. We give the corrected mapping, the evidence and the code, and recommend correcting the documentation and its copies.
Author: Pattabhi, Tejas Tovinkere
Institution: University of the Cumberlands
Language: English
Subfield: Health Information Management
Field: Health Professions
Domain: Health Sciences
Open Access status: green
OpenAlex ID: W7236668055
