Pima Indians Diabetes Dataset: A Deep Dive

The Pima Indians Diabetes Dataset is a collection of 768 medical records from female members of the Akimel O’odham (Pima) community in Arizona, each described by eight clinical measurements and a binary diabetes outcome label. Originally published in 1988 as a test bed for a neural network called ADAP, it has since become one of the most widely used benchmark datasets in machine learning, appearing in thousands of research papers, student projects, and Kaggle competitions. Its small size and simple structure make it easy to load and experiment with, but the story behind the numbers, and the problems baked into them, rarely get the attention they deserve.

Where the Data Came From

The dataset traces back to a massive longitudinal study run by the National Institute of Diabetes and Digestive and Kidney Diseases (NIDDK). Starting in 1965 and continuing until 2007, NIDDK’s Phoenix Epidemiology and Clinical Research Branch invited Pima community members aged five and older to attend research examinations every two years, regardless of whether they were healthy or sick. Each visit included an oral glucose tolerance test, blood and tissue sample collection, and a general medical exam.¹ The particular slice that became “the Pima dataset” was drawn from the NIAMD Pittsburgh database and filtered to include only women who were at least 21 years old at the time of examination.²

The ADAP paper that introduced this subset to the machine learning community reported that its neural network achieved about 76% classification accuracy on a held-out test set of 192 instances, compared with 58% for logistic regression at the time.² That gap was impressive enough in 1988 to draw attention, and the dataset was subsequently deposited in the UCI Machine Learning Repository, where it became freely available to anyone. It has lived there ever since, accumulating citations and spawning an entire cottage industry of papers claiming to beat the previous best accuracy.

Why the Pima Community Was Studied So Intensively

Researchers did not choose the Pima community at random. The Akimel O’odham living in the Gila River Indian Community of southern Arizona have one of the highest documented rates of type 2 diabetes of any population in the world. A comparative study found that the age- and sex-adjusted prevalence of type 2 diabetes among U.S. Pima Indians was about 38%, compared with roughly 6.9% among genetically related Pima Indians living in the Sierra Madre region of Mexico and 2.6% among non-Pima Mexicans in the same area.³ That gap, between two groups sharing much of the same genetic ancestry but living under very different dietary and lifestyle conditions, became a cornerstone of diabetes research.

The leading explanation is the “thrifty genotype” hypothesis, first proposed by James Neel in 1962. The idea is that populations which historically experienced cycles of feast and famine evolved a metabolism that was exceptionally efficient at storing fat. When a steady, calorie-dense food supply replaced that cycle, the same metabolic efficiency became a liability, promoting obesity, insulin resistance, and eventually diabetes. Research in the Pima community supported a specific version of this: insulin in these individuals appeared capable of maintaining fat stores even while the body resisted insulin’s role in clearing glucose from the blood.⁴ That selective resistance helps explain why diabetes rates climbed so sharply after the community’s traditional food systems were disrupted and replaced with a Western diet.

Youth-onset type 2 diabetes, which is now recognized as a growing global problem, was first documented in the Pima community back in the 1960s.¹ In many ways, the Pima experience served as an early warning signal for trends that are now appearing in Indigenous, Pacific Islander, and other communities worldwide.

What the Eight Features Actually Measure

The dataset’s simplicity is part of its appeal: eight numeric input columns and one binary outcome. But understanding what each column represents, and why it matters for diabetes risk, helps explain both why the dataset works as a benchmark and why it has serious limitations.

Pregnancies: Number of times the patient has been pregnant. Gestational diabetes raises the risk of later type 2 diabetes, and repeated pregnancies can amplify that effect.
Glucose: Plasma glucose concentration measured two hours into an oral glucose tolerance test. This is one of the standard diagnostic markers for diabetes, and in the dataset it tends to be the single strongest predictor.
Blood pressure: Diastolic blood pressure in mm Hg. Hypertension and diabetes frequently co-occur, though diastolic pressure alone is a relatively weak predictor.
Skin thickness: Triceps skinfold thickness in millimeters, used as a proxy for body fat. Research has confirmed that triceps skinfold measurement estimates insulin resistance in a way that correlates well with more expensive imaging methods like DXA scans.⁵
Insulin: Two-hour serum insulin level in μU/mL. Higher fasting and post-load insulin levels have been linked to the development of type 2 diabetes over time.⁶
BMI: Body mass index, calculated from weight and height. A blunt but widely used indicator of overweight and obesity.
Diabetes pedigree function: A numeric score reflecting the likelihood of diabetes based on family history. This is the most unusual feature in the dataset, and researchers have found that it correlates with age, insulin level, and triceps skinfold thickness, among other variables.⁷
Age: Age in years, with a minimum of 21 in this dataset.

The outcome variable is binary: 1 for a positive diabetes diagnosis, 0 for negative. Roughly 35% of the 768 records are labeled positive, creating a moderate class imbalance that affects how machine learning models perform.

The Glucose Column Deserves Special Attention

Of the eight features, the two-hour plasma glucose level does the heaviest lifting in most predictive models, and for good reason. It is the core diagnostic measurement for diabetes itself. A fasting plasma glucose level of 126 mg/dL or a two-hour post-load glucose of 200 mg/dL or above has long been used as the clinical threshold for diagnosis, and studies have confirmed that these markers correlate strongly with each other and with HbA1c.⁸ Beyond diagnosis, the two-hour glucose level also has clinical significance as an independent risk factor for cardiovascular damage, even in people whose glucose levels are considered “normal.” Research has shown that higher two-hour glucose is associated with thickening of the carotid artery wall, an early marker of atherosclerosis.⁹

This creates an interesting quirk for machine learning practitioners. Because the glucose feature is so directly related to the diagnosis itself, a model that learns to rely heavily on glucose is, in a sense, learning a clinical threshold rather than discovering a novel pattern. Removing glucose from the feature set and seeing how well a model predicts diabetes with the remaining seven variables is a much harder and arguably more informative exercise, but most benchmark papers do not bother.

Known Data Quality Problems

Anyone who has loaded this dataset into a Jupyter notebook has likely noticed something strange: hundreds of records contain zero values for features like glucose, blood pressure, skin thickness, insulin, and BMI. A blood pressure of zero is not a data point; it is a dead patient or, more likely, a missing measurement that was coded as zero instead of being left blank or flagged as missing. The insulin column is the worst offender, with roughly half the records showing a zero. Skin thickness is similarly sparse.

This is not a minor bookkeeping issue. How you handle those zeros changes your model’s performance substantially. The standard preprocessing steps include replacing zeros with column medians or means, using more sophisticated imputation methods like polynomial regression or k-nearest-neighbors imputation, or simply dropping incomplete rows (which can cost you a large share of the already small dataset). One study specifically noted that preprocessing for the Pima dataset involves feature selection, normalization, treatment of missing or zero values, and removal of duplicate records, all of which must happen before any model training can produce meaningful results.¹⁰

The class imbalance issue, where positive cases make up about 35% of the data, also matters. Models trained naively on imbalanced data tend to over-predict the majority class (non-diabetic) and miss actual cases. Researchers have applied techniques like SMOTE (Synthetic Minority Oversampling Technique) and GANs (Generative Adversarial Networks) to generate synthetic positive samples and balance the training set, reporting improved performance afterward.¹¹ Other work has combined GAN and SMOTE-based augmentation specifically on the Pima dataset to produce a more balanced training set for deep learning classifiers.¹²

The Benchmark Arms Race and Why the Numbers Vary So Widely

If you survey papers that report accuracy on the Pima dataset, the numbers span a suspiciously wide range. At the modest end, a stacking ensemble model achieved about 77% accuracy using cross-validation.¹³ At the other extreme, one deep learning study reported 98% accuracy,¹⁴ another claimed 98.16% with a deep neural network,¹⁵ and yet another reported over 99% using an attention-based deep neural network with Kendall’s correlation coefficient for feature selection.¹⁶

Those high-end numbers should be treated with skepticism. With only 768 samples, overfitting is a real danger, and small methodological choices, like which rows get dropped during preprocessing, how missing values are imputed, and whether the test set is truly held out, can swing accuracy by several percentage points. One study using ensemble methods with SMOTE-family techniques on the Pima dataset reported perfect 1.0 scores across accuracy, F1, specificity, and ROC-AUC, which in a dataset this size almost certainly reflects memorization rather than generalization.¹¹ When a model scores perfectly on 768 records, the right reaction is suspicion, not celebration.

Comparisons between algorithms also get muddied by inconsistent evaluation protocols. Some papers use a simple train-test split (often 80/20), others use 10-fold cross-validation, and others report results on the training set itself. A study comparing Naïve Bayes, random forest, and J48 decision tree models found that Naïve Bayes worked better with careful feature selection while random forest handled a larger feature set more gracefully, a finding that says more about the importance of methodology than about any single algorithm’s superiority.¹⁷ The honest summary is that well-tuned models land somewhere around 75 to 80% accuracy under rigorous cross-validation conditions, and claims far above that range warrant careful scrutiny of the evaluation setup.

Why Results on This Dataset Do Not Generalize Easily

Even if you build a model that performs well on the Pima dataset, it tells you less about diabetes prediction in the real world than you might hope. The most fundamental limitation is that the dataset contains only women, all from a single ethnic community with unusually high genetic susceptibility to type 2 diabetes.¹⁸ A model trained on this population has no exposure to male patients, no exposure to other ethnic groups with different risk profiles, and no exposure to the kinds of clinical measurements (like HbA1c, lipid panels, or waist circumference) that modern diabetes screening guidelines emphasize.

Recent work has tried to bridge this gap. One study benchmarked machine learning models and the FINDRISC questionnaire (a standard clinical diabetes risk tool) on a large prospective cohort and then tested external validity on both U.S. national data and the Pima dataset. The best models achieved an internal ROC AUC of up to 0.87, compared with 0.70 for FINDRISC, but external validation on the Pima population was part of a broader evaluation rather than proof that Pima-trained models work elsewhere.¹⁹ The takeaway is that the Pima dataset works as a test case for algorithm development, but deploying a Pima-trained model in a multiethnic clinic would be irresponsible without retraining and validation on the target population.

Data Sovereignty and the Ethics of a Famous Dataset

There is an uncomfortable dimension to the Pima dataset that the machine learning community has only recently started to grapple with. The data was collected from a specific Indigenous community over decades, under conditions that would not pass modern ethical review without substantial community involvement. Today, the dataset is freely downloadable, stripped of identifying information but permanently linked to the Pima name, and used by thousands of researchers and students who have no connection to the community and no obligation to share findings back with it.

Scholars working on Indigenous data sovereignty have pointed to the Pima dataset as a textbook case of “deterritorialization,” where data is extracted from its community of origin and circulated in ways that the community cannot control or benefit from. One framework analysis identified the Pima Indian Diabetes Dataset as exemplifying systematic deterritorialization and its associated harms, contrasting it with emerging practices in other Indigenous communities that maintain territorial control over research data.²⁰

The principles of Indigenous data sovereignty, articulated by groups like the Global Indigenous Data Alliance, hold that Indigenous peoples should govern the collection, ownership, and application of data about their communities. By that standard, the continued use of the Pima dataset without community oversight or benefit-sharing is problematic, regardless of how anonymized the records are. Some researchers have started noting this tension in their papers, though it remains far from standard practice to discuss it. The dataset’s very convenience, its availability on every tutorial site and in every textbook, makes it harder to retire or restrict.

What the Dataset Has Contributed to Diabetes Science

Despite its limitations, the broader Pima research program (of which the 768-record dataset is only a tiny sliver) has genuinely advanced the understanding of type 2 diabetes. The decades-long NIDDK study in the Gila River community produced foundational insights into the relationship between insulin resistance, obesity, and diabetes onset. It validated the thrifty genotype framework, demonstrated that environmental factors could massively amplify genetic susceptibility, and provided some of the earliest evidence that type 2 diabetes could appear in children and teenagers. The comparison between U.S. and Mexican Pima populations remains one of the cleanest natural experiments showing that lifestyle and diet trump genetics when both are present.³

For machine learning specifically, the dataset’s role has been more pedagogical than scientific. It is small enough to fit in memory on any laptop, has a clear binary target, and presents just enough messiness (the zero-coded missing values, the class imbalance, the correlated features) to force students to think about preprocessing. It is unlikely that any clinical tool will ever be deployed based solely on a model trained on these 768 records. But as a sandbox for learning how to handle real-world data problems, it has been remarkably durable. The question going forward is whether the community deserves more say in how its data continues to circulate, and whether the field can find benchmark datasets that carry less historical baggage.

1
PubMed Central. Diagnostic Criteria and Etiopathogenesis of Type 2 Diabetes and Its Complications: Lessons from the Pima Indians
2
PubMed Central. Using the ADAP Learning Algorithm to Forecast the Onset of Diabetes Mellitus
3
PubMed. Effects of traditional and western environments on prevalence of type 2 diabetes in Pima Indians in Mexico and the U.S.
4
PubMed. Diabetes mellitus in the Pima Indians: genetic and evolutionary considerations
5
PubMed. Is skinfold thickness as good as DXA when measuring adiposity contributions to insulin resistance in adolescents?
6
BMJ Open Diabetes Research & Care. Clinical correlates of plasma insulin levels over the life course and association with incident type 2 diabetes: the Framingham Heart Study
7
Internal Medicine. Determinants of Gestational Diabetes Pedigree Function for Pima Indian Females
8
Diabetes Research and Clinical Practice. Correlation among fasting plasma glucose, two-hour plasma glucose levels in OGTT and HbA1c
9
Diabetic Medicine. Two-hour post-load plasma glucose levels are associated with carotid intima-media thickness in subjects with normal glucose tolerance
10
ITM Web of Conferences. Diabetes Prediction using Deep Learning: An Analysis of the Pima Indian Dataset
11
Frontiers in Artificial Intelligence. Robust predictive framework for diabetes classification using optimized machine learning on imbalanced datasets
12
Academia. ENHANCING DIABETES PREDICTION THROUGH GAN AND SMOTE-BASED DATA AUGMENTATION
13
Heliyon. Improving diabetes disease patients classification using stacking ensemble method with PIMA and local healthcare data
14
PubMed Central. Deep learning approach for diabetes prediction using PIMA Indian dataset
15
International Journal of Cognitive Computing in Engineering. Type 2: Diabetes mellitus prediction using Deep Neural Networks classifier
16
PubMed Central. A deep neural network prediction method for diabetes based on Kendall’s correlation coefficient and attention mechanism
17
PubMed Central. Pima Indians diabetes mellitus classification based on machine learning (ML) algorithms
18
medRxiv. Enhanced Diabetes Prediction Using Novel Additive-Multiplicative Neural Networks: A Comprehensive Machine Learning Analysis of the PIMA Indians Dataset
19
PubMed Central. Internal and External Validation of Machine Learning Algorithms Versus FINDRISC for Incident Type 2 Diabetes: A Transparent, Explainable Benchmark Using SHAP
20
Graduate Student Works. Classify, Communicate, Enforce: Territorializing Data for Indigenous Data Sovereignty in Research Data Repositories