Identification of Potential Type II Diabetes in a Chinese Population with a Sensitive Decision Tree Approach

Abstract

Background Diabetes mellitus is a chronic disease with a steadfast increase in prevalence. Due to the chronic course of the disease combining with devastating complications, this disorder could easily carry a financial burden. The early diagnosis of diabetes remains as one of the major challenges medical providers are facing, and the satisfactory screening tools or methods are still required, especially a population- or community-based tool. Methods This is a retrospective cross-sectional study involving 15,323 subjects who underwent the annual check-up in the Department of Family Medicine of Shengjing Hospital of China Medical University from January 2017 to June 2017. With a strict data filtration, 10,436 records from the eligible participants were utilized to develop a prediction model using the J48 decision tree algorithm. Nine variables, including age, gender, body mass index (BMI), hypertension, history of cardiovascular disease or stroke, family history of diabetes, physical activity, work-related stress, and salty food preference, were considered. Results The accuracy, precision, recall, and area under the receiver operating characteristic curve (AUC) value for identifying potential diabetes were 94.2%, 94.0%, 94.2%, and 94.8%, respectively. The structure of the decision tree shows that age is the most significant feature. The decision tree demonstrated that among those participants with age ≤ 49, 5497 participants (97%) of the individuals were identified as nondiabetic, while age > 49, 771 participants (50%) of the individuals were identified as nondiabetic. In the subgroup where people were 34 < age ≤ 49 and BMI ≥ 25, when with positive family history of diabetes, 89 (92%) out of 97 individuals were identified as diabetic and, when without family history of diabetes, 576 (58%) of the individuals were identified as nondiabetic. Work-related stress was identified as being associated with diabetes. In individuals with 34 < age ≤ 49 and BMI ≥ 25 and without family history of diabetes, 22 (51%) of the individuals with high work-related stress were identified as nondiabetic while 349 (88%) of the individuals with low or moderate work-related stress were identified as not having diabetes. Conclusions We proposed a classifier based on a decision tree which used nine features of patients which are easily obtained and noninvasive as predictor variables to identify potential incidents of diabetes. The classifier indicates that a decision tree analysis can be successfully applied to screen diabetes, which will support clinical practitioners for rapid diabetes identification. The model provides a means to target the prevention of diabetes which could reduce the burden on the health system through effective case management.

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

“…The order of feature importance ( Figure 3 ) showed that age, BMI, and waist circumference were the top three influencing factors of diabetes, which was consistent with Pei et al's T2MD screening model based on a j48 decision tree [ 35 ]. The variables whose OR > 1 are risk factors for the disease, including age, BMI, waist circumference, systolic pressure, hypertension, ethnicity (Hui), daily smoking amount (cigarettes), fatty liver, weekly drinking amount ≥ 170 g, smoking status, and diet habit (oil loving).…”

“…In this study, we wanted to establish a simple model that can predict the risk of T2DM without blood sampling. We selected 18 variables from the questionnaire and physical examination based on the previous studies [ 35 – 37 ] ( Table 1 ).…”

Systematic literature review of machine learning methods used in the analysis of real-world data for patient-provider decision making

“…Other approaches included latent growth mixture modeling [ 45 ], support vector machine classifiers [ 46 ], LASSO regression [ 47 ], boosting methods [ 23 ], and a novel Bayesian approach [ 26 , 40 , 48 ]. Within the analytical approaches to support machine learning, a variety of methods were used to evaluate model fit, such as Akaike Information Criterion, Bayesian Information Criterion, and the Lo-Mendel-Rubin likelihood ratio test [ 22 , 45 , 47 ], and while most studies included the area under the curve (AUC) of receiver-operator characteristic (ROC) curves (Table 3 ), analyses also included sensitivity/specificity [ 16 , 19 , 24 , 30 , 41 – 43 ], positive predictive value [ 21 , 26 , 32 , 38 , 40 – 43 ], and a variety of less common approaches such as the geometric mean [ 16 ], use of the Matthews correlation coefficient (ranges from -1.0, completely erroneous information, to + 1.0, perfect prediction) [ 46 ], defining true/false negatives/positives by means of a confusion matrix [ 17 ], calculating the root mean square error of the predicted versus original outcome profiles [ 37 ], or identifying the model with the best average performance training and performance cross validation [ 36 ].…”

“…As noted above, one study using decision tree analysis used Quinlan’s C5.0 decision tree algorithm [ 15 ] while a second used an earlier version of this program (C4.5) [ 20 ]. Other decision tree analyses utilized various versions of R [ 18 , 19 , 22 , 24 , 27 , 47 ], International Business Machines (IBM) Statistical Package for the Social Sciences (SPSS) [ 16 , 17 , 33 , 47 ], the Azure Machine Learning Platform [ 30 ], or programmed the model using Python [ 23 , 25 , 46 ]. Artificial neural network analyses used Neural Designer [ 34 ] or Statistica V10 [ 35 ].…”

“…Seven studies predefined the success criteria for model performance [ 20 , 21 , 23 , 35 , 36 , 46 , 47 ], and five studies discussed the generalizability of the model [ 20 , 23 , 34 , 45 , 48 ]. Six studies [ 17 , 18 , 21 , 22 , 35 , 36 ] discussed the balance between model accuracy and model simplicity or interpretability, which was also a criterion of quality publication in the Luo scale [ 14 ]. The items on the checklist that were least frequently met are presented in Fig.…”