Diabetes Prediction (Stacking)

A stacking ensemble for early diabetes prediction on the PIMA dataset: six hyperparameter-tuned base classifiers feed a Random Forest meta-learner. Co-authored and published at IEEE AIMV 2021.

ROLE: Researcher
PERIOD: 2021
DOMAIN: Machine Learning
STATUS: Published
CODE: GitHub ↗

OVERVIEW

A stacking ensemble for early diabetes prediction on the PIMA Indians dataset (768 patients, 8 clinical features), co-authored and published at IEEE AIMV 2021. Six base classifiers, Gaussian Naive Bayes, Random Forest, Decision Tree, SVM, an ANN, and Logistic Regression, are each tuned with randomized search, then combined as the level-0 estimators of a scikit-learn StackingClassifier (cv=4) whose Random Forest meta-learner makes the final prediction. The committed notebook reaches 74.46% accuracy on a 30% held-out test set (87.9% on training), and the whole pipeline, data through evaluation, lives in one reproducible notebook.

ARRIVED AS

Single classifiers on the small PIMA Indians Diabetes dataset trade off against each other: one is better on some patients, another on others. The question this project asked was whether stacking, letting a meta-model learn how to combine several tuned base classifiers, would predict diabetes more reliably than any one model on its own.

This was a research project that became a co-authored IEEE paper (AIMV 2021) with four collaborators. The dataset is the well-known PIMA Indians Diabetes set: 768 patients, 8 clinical features (glucose, blood pressure, BMI, age, and so on), and a binary diabetes outcome. It is small and noisy, which is exactly the setting where ensembling is supposed to help, so the project tunes several classifiers and then stacks them under a meta-learner to see whether the combination is steadier than any single model.

WHAT I BUILT

01Six base classifiers, Gaussian Naive Bayes, Random Forest, Decision Tree, SVM, an ANN (MLP), and Logistic Regression, each hyperparameter-tuned with randomized search and cross-validation.
02A stacking ensemble where the six base models' predictions become inputs to a Random Forest meta-learner (scikit-learn StackingClassifier, cv=4).
03Standard preprocessing on the PIMA Indians dataset (768 patients, 8 clinical features) with a 70/30 train/test split.

WHAT CHANGED

Co-authored and published at the 2021 International Conference on Artificial Intelligence and Machine Vision (AIMV 2021) on IEEE Xplore (DOI 10.1109/AIMV53313.2021.9670920).
The committed notebook reaches 74.46% accuracy on the 30% held-out test set, against 87.9% on the training set, a gap that is itself informative on a dataset this small.
A single reproducible notebook: data, preprocessing, six tuned base models, the stacking ensemble, and per-model evaluation end to end.

Data flow

click a stage

Read the PIMA dataset (768 rows, 8 clinical features) and prepare the feature matrix and binary outcome.

COMPONENT

Preprocessing + split

Loads the PIMA dataset, assembles the 8 feature columns and the outcome, and makes a reproducible 70/30 train/test split.

Decisions, with the cost of each.

A decision without its trade-off is marketing. Each row says what was chosen, why, and what it gave up.

DECISIONBECAUSEAT THE COST OF

Stacking rather than a single tuned model or simple voting

On a small, noisy dataset different classifiers make different mistakes. Stacking lets a meta-learner learn how to combine them from data, which is more flexible than hard or soft voting and was the hypothesis the paper set out to test.

One well-tuned model (simpler, but leaves the complementary strengths of others unused); majority voting (combines models but cannot learn how to weight them).

Tune every base model before stacking

A stack is only as good as its members. Running RandomizedSearchCV on each base classifier first means the ensemble combines reasonable models rather than carrying weak defaults, and randomized search keeps tuning cheap across six models.

Stack untuned defaults (weaker base learners); exhaustive GridSearchCV (far more expensive across six models for little gain here).

Evaluate on a fixed held-out split and report it honestly

With 768 rows the result is sensitive to the split, so fixing the random_state keeps it reproducible. The committed notebook's 74.46% test accuracy (against 87.9% train) is reported as-is, including the train/test gap, rather than rounded up.

Report only the best-looking number (misleading on small data); skip a held-out set (no honest generalization estimate).

The part that mattered.

The numbers behind the work, and the code that produced them.

model ensemble: 6 + 1; 6 tuned base models + RF meta-learner
test accuracy: 74.46%; 30% held-out (87.9% train)
PIMA dataset: 768 x 8; patients x clinical features
co-authored paper: IEEE 2021; AIMV, 5 authors

Six tuned base models stacked under a meta-learnerpython

from sklearn.ensemble import StackingClassifier, RandomForestClassifier

level0 = [
    ('nb',  nb_model),
    ('rfc', rf_grid_search),
    ('dt',  dt_grid_search),
    ('lr',  lr_grid_search),
    ('svm', svm_grid_search),
    ('ann', ann_grid_search),
]
level1 = RandomForestClassifier()   # meta-learner

model = StackingClassifier(estimators=level0, final_estimator=level1, cv=4)
model.fit(X_train, y_train.ravel())

The six tuned base classifiers are the level-0 estimators; the StackingClassifier runs them with 4-fold cross-validation and feeds their predictions to a Random Forest level-1 meta-learner that makes the final call. (The committed notebook uses a Random Forest as the meta-learner.)

A fixed split, so the number is reproduciblepython

feature_col_names = ['Pregnancies', 'Glucose', 'BloodPressure',
                     'SkinThickness', 'Insulin', 'BMI',
                     'DiabetesPedigreeFunction', 'Age']

X = data_frame[feature_col_names].values
y = data_frame[['Outcome']].values

X_train, X_test, y_train, y_test = train_test_split(
    X, y, test_size=0.3, random_state=42)

Eight clinical features predict the binary outcome. With only 768 patients the score depends on the split, so random_state is fixed: the 30% held-out set is 231 patients, and that is the set every model, base and stacked, is scored on.

✓ LEARNED

Stacking helps most when the base models are individually decent and make different errors, so tuning each one before combining them matters more than the choice of meta-learner.
On a small dataset the train/test gap (87.9% vs 74.46%) is part of the result, not a footnote: it is a reminder of how little headroom 768 rows leaves before overfitting.
Fixing the split and reporting the held-out number as-is is the honest way to present a result that a different random seed would move.