Application of machine learning algorithms to identify risk factors for depression in type 2 diabetes mellitus patients: A Taiwan diabetes registry study

Figures

• 

  Fig. 1. Stepwise logistic regression model ROC. The stepwise logistic regression model selects variables that contribute most to the model. The accuracy of this method was 86.11%, which is slightly higher than the results of the unselected logistic regression (accuracy of 85.73%). The AUC is 0.806, indicating good discriminatory power. AUC = area under the curve; ROC = receiver operating characteristic.

• 

  Fig. 2. Variable importance according to the random forest model. (A) Using all variables. This figure illustrates the mean decrease in accuracy for each variable in the random forest model. The two variables that contribute most to accuracy reduction are “QoL-S11: I feel more worried and anxious than most people I know” and “QoL-S10: I am afraid that I may not take responsibility for the family in the future.” (B) After removing QoL variables. This figure shows the importance of variables in the random forest model after excluding QoL-related variables. The factors with the highest importance are bloating, autoimmune disease, eGFR, BMI, age, and city of residence. BMI = body mass index; CAD = coronary artery disease; DM = diabetes mellitus; eGFR = estimated glomerular filtration rate; HbA1c = hemoglobin A1c; PVD = peripheral vascular disease; QoL = quality of life; QoL-1 = “I think my DM is well controlled”; QoL-S2 = “I often feel afraid that my disease may get worse”; QoL-S4 = “I feel that DM interferes with what I wanted to do”; QoL-S5 = “I am worried about the risk of hypoglycemia”; QoL-S6 = “Dealing with DM is more difficult than before”; QoL-S7 = “Dealing with DM makes me…

• 

  Fig. 3. Decision tree model. (A) Using all variables. This figure shows the importance of QoL questions in the decision tree model. The question “QoL-S11: I feel more worried and anxious than most people I know” contributes most to predicting depression, followed by “QoL-S4: I feel diabetes is hindering me from doing what I really wanted to do.” QoL prompts were rated as follows: 1 = strongly disagree; 2 = disagree; 3 = agree; 4 = strongly agree; 5 = cannot or refused to answer. QoL-1 = “I think my DM is well controlled”; QoL-S4 = “I feel that diabetes mellitus interferes with what I wanted to do”; QoL-S11 = “I feel more nervous and anxious compared to most people I know”; QoL-S13 = “I worry that I may bring trouble to my family and children”. (B) Excluding quality of life questions. This figure presents the decision tree model after removing QoL variables. The factors that contribute most to the prediction model are bloating, autoimmune disease, city of residence, and PDR of the right eye. The accuracy of this model, excluding QoL questions, is 82.45%. HbA1c = hemoglobin A1c; PDR = proliferative diabetic retinopathy; QoL = quality of life; TC = total cholesterol.