Upload tuned diabetes risk prediction model

d365c2f verified 4 days ago

5.93 kB

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.

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, UCI ML Repository listing.

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.

RavshanjonEminov
/

diabetes-risk-prediction