---
license: mit
library_name: sklearn
tags:
  - sklearn
  - tabular-classification
  - healthcare
  - logistic-regression
  - skops
pipeline_tag: tabular-classification
model-index:
  - name: diabetes-risk-prediction
    results:
      - task:
          type: tabular-classification
          name: Diabetes risk classification
        dataset:
          name: Pima Indians Diabetes Dataset
          type: tabular
        metrics:
          - type: accuracy
            value: 0.7208
          - type: precision
            value: 0.6
          - type: recall
            value: 0.6111
          - type: f1
            value: 0.6055
          - type: roc_auc
            value: 0.8104

---

# Diabetes Risk Prediction

A **logistic_regression** pipeline (median imputation → standard scaling →
classifier, `class_weight="balanced"`) that estimates the probability
that a patient has diabetes from 8 routine clinical measurements. Model
selection, hyperparameters, and the decision threshold were all chosen
by 5-fold cross-validation on the training set — see the search results
below.

**This is an educational prototype, not a diagnostic or clinical
decision-making tool.** Do not use it for real patient care.

## Intended use

- **Intended**: a worked example of an end-to-end tabular ML workflow
  (cleaning → training → cross-validated tuning → evaluation) on a
  small, well-known dataset.
- **Not intended**: diagnosis, screening, or any clinical
  decision-making. The training data is small (768 patients), from a
  single population (Pima Indian women, a study from the 1990s), and is
  not representative of a general patient population.

## How to use

> The "Use this model" button above may show a snippet based on
> `skops.hub_utils.download`, which was removed in skops >= 0.12 (this
> model was published with skops 0.14) and will raise `ModuleNotFoundError`
> if run as-is. Use the snippet below instead — it's verified against
> this exact repo.

```python
from huggingface_hub import hf_hub_download
import skops.io as sio
import pandas as pd

path = hf_hub_download("RavshanjonEminov/diabetes-risk-prediction", "model.skops")
# numpy.dtype is skops's only flagged type here -- it comes from the
# scaler/imputer's stored statistics arrays, not from executable code.
artifact = sio.load(path, trusted=["numpy.dtype"])
pipeline, threshold = artifact["pipeline"], artifact["threshold"]

patient = pd.DataFrame([{
    "Pregnancies": 2, "Glucose": 130, "BloodPressure": 70,
    "SkinThickness": 25, "Insulin": 90, "BMI": 28.5,
    "DiabetesPedigreeFunction": 0.35, "Age": 45,
}])
probability = pipeline.predict_proba(patient)[0, 1]
prediction = "Diabetes" if probability >= threshold else "No Diabetes"
print(prediction, probability)
```

## Training data

- Pima Indians Diabetes Dataset (National Institute of Diabetes and
  Digestive and Kidney Diseases), 768 patients, 8 features, 1 binary
  outcome.
- Features: `Pregnancies`, `Glucose`, `BloodPressure`, `SkinThickness`,
  `Insulin`, `BMI`, `DiabetesPedigreeFunction`, `Age`.
- `Glucose`, `BloodPressure`, `SkinThickness`, `Insulin`, and `BMI` use
  `0` to encode missing values in the source data; these are treated as
  missing and median-imputed (fit on the training split only).
- 80/20 train/test split, stratified by outcome, `random_state=42`.
- Class balance: ~65% no diabetes, ~35% diabetes.
- Sources: [Kaggle mirror](https://www.kaggle.com/datasets/uciml/pima-indians-diabetes-database),
  [UCI ML Repository listing](https://archive.ics.uci.edu/ml/datasets/pima+indians+diabetes).

## Model selection & tuning

A 5-fold cross-validated grid search (training set only, scored by
ROC-AUC) compared 3 model families, each with `class_weight="balanced"`
to address low recall on the diabetes class:

| Candidate | CV ROC-AUC | Best hyperparameters |
|---|---|---|
| logistic_regression | 0.845 | `{'classifier__C': 0.1}` |
| random_forest | 0.8386 | `{'classifier__max_depth': 5, 'classifier__min_samples_leaf': 1, 'classifier__n_estimators': 200}` |
| hist_gradient_boosting | 0.829 | `{'classifier__learning_rate': 0.03, 'classifier__max_depth': 3, 'classifier__max_iter': 100}` |

The decision threshold (0.58) was then chosen from out-of-fold
predictions to maximize F1, instead of leaving it at the default 0.5.

## Evaluation

**Test set** (n=154, threshold=0.58):

| Metric | Value |
|---|---|
| Accuracy | 0.7208 |
| Precision | 0.6 |
| Recall | 0.6111 |
| F1 | 0.6055 |
| ROC-AUC | 0.8104 |

**5-fold cross-validation** (training set, mean ± std):

| Metric | Score |
|---|---|
| accuracy | 0.7671 ± 0.0081 |
| precision | 0.6515 ± 0.0203 |
| recall | 0.7198 ± 0.05 |
| f1 | 0.6823 ± 0.0168 |
| roc_auc | 0.845 ± 0.013 |

**Sanity checks**: model ROC-AUC 0.8104 vs.
a class-balance-only baseline of 0.5452.
Bootstrap 95% CI for test-set ROC-AUC:
[0.7321, 0.8759]
(2000 resamples) — a reminder that a single
154-patient test set carries real uncertainty.

## Limitations

- Small, single-population dataset; generalization to other populations
  is unverified.
- Median imputation for missing values is a simplification.
- No external validation cohort — metrics reflect a random 20% split of
  the same source data.
- Not calibrated or validated for clinical use.

## Ethical considerations

Diabetes risk prediction is sensitive: false negatives could delay
appropriate care, and false positives could cause unnecessary worry.
Real-world use would require clinical validation, regulatory review, and
ongoing monitoring — well beyond the scope of this prototype.

## Links

- Source code: see the project's GitHub repository (search
  `RavshanjonEminov/diabetes-risk-prediction`).
- License: MIT (code); see dataset sources above for data terms.