[Paper Review] An Ensemble Classifier for Predicting the Onset of Type II Diabetes

John Semerdjian, Spencer Frank|arXiv (Cornell University)|Aug 24, 2017

Artificial Intelligence in Healthcare4 references3 citations

TL;DR

This study proposes an ensemble classifier combining five machine learning models to predict type II diabetes onset using 16 lifestyle and demographic features from NHANES survey data. Despite the ensemble approach, the gradient boosting classifier alone achieved the highest AUC of 0.84, outperforming the ensemble’s 0.834, suggesting limitations in model weighting and hyperparameter tuning under computational constraints.

ABSTRACT

Prediction of disease onset from patient survey and lifestyle data is quickly becoming an important tool for diagnosing a disease before it progresses. In this study, data from the National Health and Nutrition Examination Survey (NHANES) questionnaire is used to predict the onset of type II diabetes. An ensemble model using the output of five classification algorithms was developed to predict the onset on diabetes based on 16 features. The ensemble model had an AUC of 0.834 indicating high performance.

Motivation & Objective

• To improve prediction accuracy of type II diabetes onset using easily obtainable survey and lifestyle data from the NHANES dataset.
• To develop an ensemble model that combines multiple classifiers to enhance predictive performance beyond individual models.
• To interpret model outputs for real-world clinical and public health applications, particularly in early disease detection and preventative care.
• To evaluate the impact of data imputation and hyperparameter tuning on model performance.
• To determine whether ensemble methods consistently outperform individual models on this dataset, given the complexity of tuning and feature weighting.

Proposed method

• The study uses 5,515 NHANES survey samples (1999–2004) with 16 selected features, including age, BMI, waist circumference, glucose levels, and lifestyle factors.
• Labels were assigned using two criteria: self-reported diabetes diagnosis and plasma glucose levels >126 mg/dL, with 4600 samples labeled as diabetic via the latter.
• A 20% test set was reserved, and 10-fold cross-validation with grid-search was used for hyperparameter tuning on the training set.
• Five base classifiers (SVM, Random Forest, Gradient Boosting, KNN, Logistic Regression) were trained independently, and their predicted probabilities were averaged in an ensemble model with a tunable decision threshold.
• Missing values (>25% across features) were imputed using mean (numerical) and mode (categorical) values from the training set.
• The ensemble model used unweighted averaging of individual model probabilities, with the final decision threshold tuned to achieve a target recall of 75% for diabetics.

Experimental results

Research questions

• RQ1 Can an ensemble of multiple classifiers improve prediction accuracy for type II diabetes onset compared to individual models using NHANES survey data?
• RQ2 What is the optimal decision threshold for balancing sensitivity (recall) and specificity in a clinical screening context?
• RQ3 How do data imputation strategies and hyperparameter tuning affect model performance on this dataset?
• RQ4 Why did the ensemble model underperform compared to the standalone gradient boosting classifier despite theoretical advantages?
• RQ5 To what extent can survey-based models achieve high AUC, and is 0.83–0.84 a practical upper bound for this data modality?

Key findings

• The ensemble classifier achieved an AUC of 0.834, indicating strong discriminative performance in predicting type II diabetes onset from survey data.
• The standalone gradient boosting classifier outperformed the ensemble, achieving an AUC of 0.84, the highest among all models tested.
• The top five most important features, as identified by random forest, were age, waist circumference, family history of diabetes, height, and total cholesterol.
• A decision threshold of T=0.78 was selected to achieve 75% recall for diabetic patients, correctly identifying 75 out of 100 actual diabetics.
• This recall level required notifying 2,165 patients as potentially diabetic, including 55% of non-diabetics, highlighting a trade-off between sensitivity and specificity.
• The study suggests that current performance limits (AUC ~0.83–0.84) may be near the upper bound for simple classifiers on this dataset, with data imputation being a key future improvement path.