Author + information
To establish a prediction model for Coronary Heart Disease (CHD) instead of Coronary Angiography (CAG), and to establish a more simple and noninvasive method for the diagnosis and treatment of a large number of patients with CHD. The model could provide low-cost diagnostic treatment regimens and health interventions.
(1) Data description. The study enrolled 599 cases of CHD patients and 398 healthy people over 60 years old. 85 health indicators such as height, weight, body mass index (BMI), smoking, drinking, angina pectoris, abnormal heart sounds and 74 health intervention methods, were used to build the data platform and prepare for the next step.
(2) Dimension reduction. K-means cluster analysis and main component analysis were used for dimension reduction and screening independent variables strongly associated with CHD. For examples, the former was used for regional division, fault correlation and filter of health data indicators, and the later obtained variables closely related to CHD.
(3) Prediction. A type of Back Propagation (BP) neural network, using L-M learning algorithm, was established, where 25 kinds of independent variables were as input, and the type of CHD and intervention methods were as output. Then The neural network was trained and evaluated through 448 cases (nearly 50%) as training samples, 224 cases (nearly 25%) as validated samples and 225 cases (nearly 25%) as test samples.
(1) The results using above methods (K-means cluster analysis and main component analysis) were basically same. Total 25 variables such as basic information (e.g. age, sex, BMI, heart rate, carotid murmur, lung rales, foot cyanosis, lower extremity edema, highest systolic /diastolic blood pressure, heart lingering, high density lipoprotein cholesterol (HDL), triglycerides), symptoms (e.g. angina pectoris, lifestyle e.g. moderate intensity exercise, drinking) and comorbidities (e.g. stroke, diabetes, hyperhomocysteinemia, hyperuricemia, diabetes, stroke, peripheral vascular disease, hypertension, tumor) were selected.
(2) In the sample, the ratio of the control group and the experimental group was 1: 1.5, where unstable angina accounted for 32.2%, myocardial infarction accounted for 14.9%, stable angina accounted for 29.8%, ischemic cardiomyopathy accounted for 7.5% and asymptomatic myocardial ischemia accounted for 15.6%. The accuracy rate of prediction reached to 93.4% through the BP neural networks.
(3) According to different types of CHD, nitrates (e.g. calcium antagonists, trimexazine, nicotilil, aspirin, clopidogrel, ACEI, ARB, statins, beta blockers), hypoglycemic (e.g. metformin, sulfonylureas, glibenium class, thiazolidinediones, alpha glucosidase inhibitor), antihypertensive (e.g. diuretics, calcium antagonists), lipid-lowering drugs (e.g. statins, bite class, niacin), exercise or percutaneous coronary intervention (PCI), CABG and other recommendations were given.
Based on the big data platform and data mining, the BP neural network can predict the 5 types of CHD and propose 8 main types of health intervention solutions, which could promote accuracy of prediction.