Clinically adaptable machine learning model to identify early appreciable features of diabetes

Objective Diabetes mellitus is a serious disease where the body of affected patients are failed to produce enough insulin that causes an abnormality of blood sugar. This disease happens for a number of reasons including modern lifestyle, lethargic attitude, unhealthy food consumption, family history, age, overweight, etc. The aim of this study was to propose a machine learning based prediction model that detected diabetes at the beginning.

Methods In this work, we collected 520 patients records from the University of California, Irvine (UCI) machine learning repository of Sylhet Diabetes Hospital, Sylhet. Then, a similar questionnaire of that hospital was followed and assembled 558 patients records from all over Bangladesh through this questionnaire. However, we accumulated patient records of these two datasets. In the next step, these datasets were cleaned and applied thirty five state-of-arts classifiers such as logistic regression (LR), K nearest neighbors (KNN), support vector classifier (SVC), Nave Byes (NB), decision tree (DT), random forest (RF), stochastic gradient descent (SGD), Perceptron, AdaBoost, XGBoost, passive aggressive classifier (PAC), ridge classifier (RC), Nu-support vector classifier (Nu-SVC), linear support vector classifier (LSVC), calibrated classifier CV (CCCV), nearest centroid (NC), Gaussian process classifier (GPC), multinomial NB (MNB), complement NB, Bernoulli NB (BNB), categorical NB, Bagging, extra tree(ET), gradiant boosting classifier (GBC), Hist gradiant boosting classifier (HGBC), one vs rest classifier (OVsRC), multi-layer perceptron (MLP), label propagation (LP), label spreading (LS), stacking, ridge classifier CV (RCCV), logistic regression CV (LRCV), linear discriminant analysis (LDA), quadratic discriminant analysis (QDA), and light gradient boosting machine (LGBM) to explore best stable predictive model. The performance of the classifiers has been measured using five metrics such as accuracy, precision, recall, F1-score, and area under the receiver operating characteristic. Finally, these outcomes were interpreted using Shapley additive explanations methods and identified relevant features for happening diabetes.

Results In this work, different classifiers were shown their performance where ET outperformed any other classifiers with 97.11% accuracy for the Sylhet Diabetes Hospital dataset (SDHD) and MLP shows the best accuracy (96.42%) for the collected dataset. Subsequently, HGBC and LGBM provide the highest 94.90% accuracy for the combined datasets individually.

Conclusion LGBM, stacking, HGBC, RF, ET, bagging, and GBC might represent more stable prediction results for each dataset.