作者: Kalyan Guin , Sourangshu Bhattacharya , Asis Roy
DOI:
关键词:
摘要: Tests for Esophageal cancer can be expensive, uncomfortable and have side effects. For many patients, we predict non-existence of disease with 100% certainty, just using demographics, lifestyle, medical history information. Our objective is to devise a general methodology customizing tests user preferences so that expensive or avoided. We propose use classifiers trained from electronic health records (EHR) selection tests. The key idea design false normal rates, possibly at the cost higher abnormals. compare Naive Bayes classification (NB), Random Forests (RF), Support Vector Machines (SVM) Logistic Regression (LR), find kernel regression most suitable task. an algorithm finding best probability threshold LR, based on test set accuracy. Using proposed algorithm, describe schemes selecting tests, which appear as features in automatic costs discomfort users. our EHRs collected more than 3000 part project carried out by reputed hospital Mumbai, India. Kernel SVM LR polynomial degree 3, yields accuracy 99.8% sensitivity 100%, without MP features, i.e. only clinical demonstrate two case studies, one other "discomfort" values compute sets corresponding lowest abnormals each criterion described above, exhaustive enumeration 15 turn different, substantiating claim customize preferences.