Design of intelligent diabetes mellitus detection system using hybrid feature selection based XGBoost classifier

• Published in: Computers in biology and medicine Vol. 136; p. 104664
• Main Authors: Prabha, Anju, Yadav, Jyoti, Rani, Asha, Singh, Vijander
• Format: Journal Article
• Language: English
• Published: Oxford Elsevier Ltd 01.09.2021; Elsevier Limited
• Subjects: Accuracy; Age; Algorithms; Artificial intelligence; Classification; Classifiers; Computer applications; Datasets; Decision trees; Deep learning; Diabetes; Diabetes detection; Diabetes mellitus; Diabetic retinopathy; Electrocardiography; Feature selection; Glucose; Hybrid systems; Insulin; Internal Medicine; Laboratories; Learning algorithms; Machine learning; MFCC; Neural networks; Optimization techniques; Other; Physiology; Polypharmacy; PPG; Smartphones; Support vector machines; XGBoost; Feature selection; MFCC; PPG; XGBoost; Diabetes detection
• ISSN: 0010-4825; 1879-0534; 1879-0534
• DOI: 10.1016/j.compbiomed.2021.104664

In this work, a non-invasive diabetes mellitus detection system is proposed based on the wristband photoplethysmography (PPG) signal and basic physiological parameters (PhyP) to enable easy detection of diabetes mellitus (DM). A dataset of 217 participants with diabetes, prediabetes and normal conditions is used to develop the system. The Mel frequency cepstral coefficients (MFCC) extracted from 5s PPG signal segments and the PhyP are used as input for the machine learning algorithms. The K-nearest neighbors, support vector machine, random forest and extreme gradient boost (XGBoost) classifiers are used for classification. In addition, a hybrid feature selection method (Hybrid FS) is proposed to reduce the size of the input data. The Hybrid FS-based XGBoost system achieves a high accuracy of 99.93 % for non-invasive diabetes detection with fewer features and less computational effort. The analysis suggests that the PPG signal from a wearable sensor is a good alternative for simple non-invasive blood glucose measurements in routine applications. [Display omitted] •An intelligent diabetes mellitus detection system based on the photoplethysmography (PPG) signal and physiological parameters is proposed.•Mel frequency cepstral coefficients (MFCC) features of the PPG signal are used.•XGBoost classifier with a hybrid feature selection technique is suggested for the classification.•A high accuracy of 99.93 % is achieved with fewer features and less computational effort.