Identification of Potential Type II Diabetes in a Large-Scale Chinese Population Using a Systematic Machine Learning Framework

Mingyue Xue, Yinxia Su, Chen Li, Shuxia Wang, Hua YaoJournal of Diabetes Research 2020 · DOI

TL;DR

This paper introduces a novel machine learning framework for identifying potential Type II Diabetes (T2DM) in a large Chinese population using easily obtainable, non-invasive variables. By comparing four tree-based machine learning algorithms, the study highlights XGBoost as the most effective, achieving high accuracy and providing valuable insights into the key risk and protective factors for T2DM.

What Problem Does This Paper Address?

Type II Diabetes Mellitus (T2DM) is a growing global health crisis, placing a significant burden on healthcare systems. Early detection is crucial for effective management and prevention of its severe complications. Traditional screening methods often rely on expensive blood tests and require significant manpower, making them challenging, especially in remote areas. This paper addresses the need for a more accessible, cost-effective, and accurate screening method that utilizes readily available patient data. The study aims to identify individuals at risk of T2DM using a systematic machine learning approach applied to a large-scale Chinese population.

Methods

The study employed a systematic machine learning framework. Initially, Logistic Regression (LR) was used for feature selection, identifying 14 statistically significant variables associated with T2DM from a larger set. These selected variables were then used to train and evaluate four tree-based machine learning algorithms: Decision Tree (DT), Random Forest (RF), AdaBoost, and Extreme Gradient Boosting (XGBoost).

Key methodological aspects include:

  • Data Cleaning and Preprocessing: Handling missing values and outliers was a critical first step.
  • Feature Selection: LR was used to identify significant risk factors based on p-values and odds ratios.
  • Data Imbalance Processing: The SMOTE (Synthetic Minority Over-sampling Technique) algorithm was employed to address the class imbalance problem, where the number of non-diabetic subjects significantly outnumbered diabetic subjects. This is a standard technique for imbalanced datasets.
  • Model Training and Validation: A 70/30 split was used for training and testing datasets, with 5-fold cross-validation for hyperparameter tuning.
  • Classifier Comparison: The performance of the four ML models was rigorously compared using metrics like accuracy, precision, recall, F-1 score, and AUC (Area Under the ROC Curve).
  • Feature Importance Ranking: The study utilized the feature importance scores from the best-performing model (XGBoost) to rank the influence of each variable on T2DM prediction.

The novelty lies in the systematic comparison of these four tree-based ML algorithms for T2DM screening in this specific population, using a combination of LR for initial feature selection and then advanced tree-based methods for classification and risk factor identification.

Datasets

The study utilized data from the National Physical Examination (NPE) conducted in Urumqi, China, in 2018.

  • Original Dataset: 643,439 adult subjects participated, with 584,168 included after exclusions (pregnancy, Type I diabetes, age < 20).
  • T2DM Incidence: The final incidence of T2DM in the study population was 12.4%.
  • SMOTE-processed Dataset: After applying SMOTE for data imbalance, the dataset was expanded to 1,020,822 samples (714,575 for training, 306,247 for validation).

This dataset is notable for its large scale and its inclusion of diverse ethnic groups within China, contributing to the generalizability of the findings.

Key Insights and Ideas

  • XGBoost Dominance: XGBoost emerged as the superior classifier, outperforming DT, RF, and AdaBoost in accurately identifying potential T2DM cases. This underscores its power in handling complex, high-dimensional data.
  • Non-invasive Variables are Sufficient: The study successfully demonstrated that a combination of easily obtainable, non-invasive variables (physical measurements and questionnaire data) can be highly effective for T2DM screening, reducing the reliance on costly and invasive tests.
  • BMI as the Primary Risk Factor: Body Mass Index (BMI) was identified as the most significant predictor of T2DM, followed closely by age and waist circumference. This reinforces established knowledge but provides quantitative evidence within this large cohort.
  • Protective Factors Identified: Beyond risk factors, the study also highlighted protective factors, such as physical activity and a balanced diet (meat balanced), suggesting actionable lifestyle interventions.
  • Dose-Response Relationship: The study revealed a nuanced dose-response relationship for smoking and drinking, indicating that while complete cessation is ideal, even reducing consumption can mitigate risk.

Key Results

  • XGBoost Performance: The best-performing XGBoost model achieved an accuracy of 0.906, precision of 0.910, recall of 0.902, F-1 score of 0.906, and an AUC of 0.968.
  • Top Risk Factors: The top influencing factors for T2DM were:
  • BMI (F score: 577,984)
    1. Age (F score: 301,525)
    2. Waist circumference (F score: 289,592)
    3. Systolic pressure (F score: 264,367)
    4. Ethnicity (F score: 44,073)
  • Protective Factors: Variables with an Odds Ratio (OR) < 1, indicating a protective effect, included ethnicity (Kazak and Mongolian), physical activity, and a meat-balanced diet.
  • LR-XGBoost Classifier: The proposed classifier, combining LR for feature selection and XGBoost for classification, uses 14 easily obtained, non-invasive variables.

Limitations and Future Work

  • Cross-sectional Study: The study's cross-sectional design prevents the establishment of causal relationships between T2DM and other comorbidities.
  • Generalizability: While large, the data is from China, which may limit extrapolation to other populations due to potential differences in T2DM pathophysiology and lifestyle.
  • Missing Data: Education and family history, known determinants of diabetes, were not available in the dataset.
  • Parameter Tuning: Only the "maximum depth" parameter was optimized for the classification trees. Further tuning of multiple parameters could potentially improve model performance.
  • Subjective Indicators: Some indicators, like eating habits, rely on self-reporting and lack objective, unified evaluation criteria, potentially affecting prediction accuracy.
  • Future Work: The authors suggest that future research could involve follow-up studies to assess the causal relationships and incorporate more objective evaluation criteria for lifestyle factors. Further optimization of ML models by tuning multiple parameters is also recommended.

Summary updated

Abstract

We proposed a classifier based on LR-XGBoost which used fourteen variables of patients which are easily obtained and noninvasive as predictor variables to identify potential incidents of T2DM. The classifier can accurately screen the risk of diabetes in the early phrase, and the degree of variables' importance scores gives a clue to prevent diabetes occurrence.
BibTeX
@article{xue2020identification,
  author = {Mingyue Xue and Yinxia Su and Chen Li and Shuxia Wang and Hua Yao},
  title = {{Identification of Potential Type II Diabetes in a Large-Scale Chinese Population Using a Systematic Machine Learning Framework}},
  journal = {Journal of Diabetes Research},
  year = {2020},
  doi = {10.1155/2020/6873891},
}
Figure