期刊名称：International Journal of Advanced Computer Science and Applications(IJACSA)
印刷版ISSN：2158-107X
电子版ISSN：2156-5570
出版年度：2019
卷号：10
期号：2
页码：268-275
DOI：10.14569/IJACSA.2019.0100236
出版社：Science and Information Society (SAI)
摘要：Hospital readmission is considered a key metric in order to assess health center performances. Indeed, readmissions involve different consequences such as the patient’s health condition, hospital operational efficiency but also cost burden from a wider perspective. Prediction of 30-day readmission for diabetes patients is therefore of prime importance. The existing models are characterized by their limited prediction power, generalizability and pre-processing. For instance, the benchmarked LACE (Length of stay, Acuity of admission, Charlson comorbidity index and Emergency visits) index traded prediction performance against ease of use for the end user. As such, this study propose a comprehensive pre-processing framework in order to improve the model’s performance while exploring and selecting a prominent feature for 30-day unplanned readmission among diabetes patients. In order to deal with readmission prediction, this study will also propose a Multilayer Perceptron (MLP) model on data collected from 130 US hospitals. More specifically, the pre-processing technique includes comprehensive data cleaning, data reduction, and transformation. Random Forest algorithm for feature selection and SMOTE algorithm for data balancing are some example of methods used in the proposed pre-processing framework. The proposed combination of data engineering and MLP abilities was found to outperform existing research when implemented and tested on health center data. The performance of the designed model was found, in this regard, particularly balanced across different metrics of interest with accuracy and Area under the Curve (AUC) of 95% and close to the optimal recall of 99%.
