Diabetes dataset details
Supplemental historical source, zero placeholder and binary label details for dataset.diabetes().
For loading and examples, start with Well known datasets. This page documents the Pima Indians Diabetes classification dataset returned by dataset.load("diabetes") and dataset.diabetes().
Source and attribution
Smith, J. W., Everhart, J. E., Dickson, W. C., Knowler, W. C., and Johannes, R. S. (1988). Using the ADAP learning algorithm to forecast the onset of diabetes mellitus. Proceedings of the Symposium on Computer Applications and Medical Care, 261 to 265. Original owners: National Institute of Diabetes and Digestive and Kidney Diseases. Original donor: Vincent Sigillito, Johns Hopkins University Applied Physics Laboratory.
Aner pins the OpenML dataset 37, version 1 copy, downloaded from diabetes.arff on October 13, 2026. The unmodified ARFF retains the original description and citation. OpenML metadata declares the license string Public; Aner records it as OpenML-Public, without claiming an SPDX or Creative Commons license. The exact metadata response and source status notice are retained.
The original data were withdrawn from UCI and mlbench over reported informed consent concerns, according to the current mlbench source documentation. This is a historical benchmark copy from OpenML, not an endorsement by the original repository or evidence that withdrawal has been reversed. The synthetic replacement is a different dataset and is not silently substituted here.
| Property | Value |
|---|---|
| Variant | Original Pima classification values from OpenML 37 v1 |
| Observations / features | 768 / 8 |
| Classes | tested_negative = 0 (500), tested_positive = 1 (268) |
Upstream ARFF SHA-256 | 4eddd5b2b64679e8888348e306520a393d6a28e1ddc9643cfb76fc5d912d6d40 |
| Runtime input | diabetes.data, headerless predictors followed by target |
Runtime input SHA-256 | 33e704cdafa8769a75728e4658dcce5bc1da1ce36174603687fe545f46e39394 |
| Version | openml-pima-diabetes-sha256-33e704cdafa8 |
Feature order and missing value conventions
| Index | Feature | Source field | Unit / meaning |
|---|---|---|---|
| 0 | pregnancies | preg | Number of pregnancies |
| 1 | glucose | plas | Plasma glucose concentration in a two hour oral glucose tolerance test |
| 2 | blood_pressure | pres | Diastolic blood pressure, mm Hg |
| 3 | skin_thickness | skin | Triceps skin fold thickness, mm |
| 4 | insulin | insu | Two hour serum insulin, micro units/ml |
| 5 | bmi | mass | Body mass index, kg/m² |
| 6 | diabetes_pedigree | pedi | Diabetes pedigree function |
| 7 | age | age | Age, years |
Aner removes only ARFF syntax and maps the two class strings to the indices above. All feature values and source row order remain unchanged. The loader does not standardize, impute, drop records, or interpret zeros automatically.
Zero values occur in glucose (5), blood pressure (35), skin thickness (227), insulin (374), and BMI (11). These are commonly treated as missing measurement placeholders; zero pregnancies remains valid. If an experiment replaces or imputes such values, make that policy explicit and estimate any imputation parameters only from the training partition. The bundled numeric Tensor contains the original zeros, not nullable cells or inferred clinical measurements.
The source population comprises women aged at least 21 of Pima heritage. This historical benchmark is not a representative general population sample. It has no official train/test assignment. Results support teaching and implementation comparisons, not clinical use.