The repository tree of this project is given below:
The graphical abstract of this research is as follows:
Diabetes is a kind of metabolic disease that forms by lack of insulin due to the malfunctioning of the pancreas. Diabetes can push a person into pathological destruction of pancreatic beta cells, coma, cardiovascular dysfunction, renal and retinal failure, joint failure, sexual dysfunction, pathogenic effects on immunity, weight loss, and peripheral vascular diseases. So, for the early detection of diabetes, a robust framework was proposed, where outlier rejection, filling the missing values, data standardization, K-fold validation, and different Machine Learning (ML) classifiers (k-NN, decision trees (DT), random forest (RF), AdaBoost (AB), naive Bayes (NB), and XGBoost (XB)) were used. To improve the result, the weighted ensembling of different ML models also proposed here. The corresponding Area Under ROC Curve (AUC) of the ML model as the performance metric estimated these weights. Using the grid search technique, the AUC is then maximized during hyperparameter tuning. All experiments were conducted under the same experimental conditions on publicly available Pima Indian population near Phoenix, Arizona of 768 female diabetic patients, where there are 268 diabetic patients (positive) and 500 non-diabetic patients (negative) with eight different attributes. The proposed framework is shown in the figure below.
After having Pima Dataset, we preprocessed the data like outlier rejection, filling missing values, data standardization, and dimensionality reduction of the attribute. After preprocessing, the 5-fold Cross-Validation technique is used for model selection and error estimation of classifiers. In our proposal, 4 folds and grid search algorithms were used to train and fine-tune the hyper-parameters in the inner loop, whereas the remaining fold is used for testing the model. After these processing, k-NN, DT, AB, RF, NB, and XB were implemented. Again, the ensembling of the ML model using a group of classifiers is used to improve the precision of the prediction.
From all the models in the experiment, XGBoost performed better with sensitivity, specificity, false omission rate, diagnostic odds ratio, and AUC as 0.768, 0.943, 0.100, 71.369, and 0.946 respectively. The proposed ensembling model outperforms XGBoost and state of the art results by 0.40 % and 2.00 % respectively in AUC.
All the results reported in the literature were produced using the following version Python and Python API:
- python 3.6.5
- numpy 1.18.1
- pandas 1.0.0
- matplotlib 3.1.2
- seaborn 0.10.0
- scikit-learn 0.22.1
- PyXGBoost 1.0.9
- xgboost 0.90
- scipy 1.4.1
The more details of the proposed framework are available in the following Journal-
https://ieeexplore.ieee.org/document/9076634
Md. Kamrul Hasan
Erasmus Scholar on Medical Imaging and Application (MAIA) [2017-2019] [http://maiamaster.udg.edu/]
Assistant Professor
Department of EEE, KUET, Khulna-9203, Bangladesh
For more details write me at kamruleeekuet@gmail.com