A weighted logistic regression based on similarity learning for prediction of readmission event in hospitals

The federal government announced that hospitals with higher readmission rates than expected will be penalized and get less money from payers like Medicare, Medicaid or insurance companies. The number of patients who experience readmission to a hospital after a previous hospital stay is used to evaluate the quality of hospital care. And high readmission rates are considered as the wasteful spending. There are a variety of statistical methods can be used for predicting the probability of a specific event. By predicting the probability of the readmission with the patients' history data before discharging, hospitals can change the schedule to discharge the patients and treat them something more eventually to reduce the readmission rates. The assumption of the experiment in this paper is that if we consider the similarity between test and training data when we fit the model, the discriminatory power will be better and more accurate. For this purpose, Gaussian kernel logistic regression was used. Gaussian kernel function measures similarity between a point of interest and one of N covariate vectors with kernel trick. Kernel logistic regression (KLR) is a promising technique in forecasting and other applications for big databases, non-linearity classification or in addition to many predictors (Elbashir and Wang, 2015). To compare this Gaussian kernel to the other similarity method, Jaccard similarity & Pearson correlation methods were used. After calculating the similarity with these two similarity methods, fit the weighted logistic regression to the data. And these two similarity-based approaches were compared with the normal logistic regression to evaluate the classifier performances. The experimental results show that weighted logistic regression using Jaccard & Pearson Correlation achieved slightly better prediction performance than others.