Comparison of classification models for the stroke among elderly: A case study of Somdech Phra Pinklao Hospital

The objective of this paper is to compare classification modes and study factors associated for stroke prediction in the elderly in Somdech Phra Pinklao Hospital, Bangkok, Thailand. The personal medical records of elderly patients who are over 60-year-old and visit the hospital in 2018, total 28,928 patients have been collected and preprocessed. Because of imbalance data, over-sampling technique is used to increase smaller group size. Then they have been partitioned into two groups. The former (80%) is used to construct models, which are the stepwise binary logistic regression (glm) and decision tree models (ID3, CART, J48, CTREE and C5.0) with Bootstrap Aggregating (Bagging). While the latter (20%) is used to evaluate the accuracy of the model.  The result shows that the prevalence rate of stroke patients is 5.50% (95% CI 5.24% -5.76%). The most effective model is the C5.0 decision tree model with the accuracy of 95.31 percent, sensitivity of 94.48 percent, specificity of 96.12 percent, the positive prediction value of 95.93 percent and the negative prediction value of 94.73 percent. Using the C5.0 decision tree model, the important risk factors effecting on the stroke of the elderly by order are Transient ischemic attack, Age, Anemia, Epilepsy, Smoking, Clotting disorder and bleeding, Head injury, Heart disease, Cancer, Drinking alcohol, Kidney disease, The presence of implants and implants for the heart and blood vessels, Sex, Hypertension, Diabetes, Body mass index, Disorders of arteries, arterioles and capillaries, and Pulmonary embolism. While Overweight, obesity & hypernutrition and Metabolic disorder are not included the model to classify the stroke of the elderly.