Electronic health records (EHR) quality control and temporal data analysis for clinical decision support

Data mining;Electronic health records

Venugopalan, Janani ; Wang, May Biomedical Engineering (Joint GT/Emory Department) Butera, Robert Maher, Kevin Chanani, Nikhil Qiu, Peng Braunstein, Mark ; Wang, May

University:Georgia Institute of Technology

Department:Biomedical Engineering (Joint GT/Emory Department)

关键词: Data mining;    Electronic health records;   

Others  :  https://smartech.gatech.edu/bitstream/1853/62224/1/VENUGOPALAN-DISSERTATION-2018.pdf

美国|英语

来源: SMARTech Repository

【 摘 要 】

US healthcare is undergoing major reforms towards evidence based and precision medicine with an emphasis on data driven models. There is a strong impetus for improving the quality of healthcare while decreasing the cost incurred. The objective of this research is to develop methodologies for clinical-decision support which target acute and chronic care using EHR data. We address current data mining challenges such as 1) missing data 2) the sequential nature of records in the ICU and 3) integration of heterogenous data for analysis. In this thesis, we developed novel strategies to solve these issues and contribute to this field of computer-aided diagnosis, with the following three specific aims: 1) To improve predictive performance by developing data imputation techniques for missing data in EHR, 2) To develop predictive models and personalized temporal risk profiling for temporal EHR data, and 3) To integrate EHR data using deep learning based predictive models at multiple temporal resolutions and modalities. In the first aim we focus on the data issues and is aimed at solving challenges such as missing data. We divide the missing data into multiple types on the basis of the statistical properties of the data and develop novel methodologies to impute each missing data type. In the second aim, we perform temporal analysis of quality-controlled data and compare with conventional non-temporal analysis. We also integrate survival analysis into temporal predictive modeling to visualize personalized patient risk profiles across time. In the third aim we develop deep architectures for data integration across multiple temporal scales and modalities. We also develop techniques for interpreting the features and results from deep models. We demonstrate our results using data from the intensive care units and Alzheimer’s disease populations, for end-points such as the prediction of ICU readmission, mortality, length of stay, and Alzheimer’s stage detection.