Cloud-based Predictive Modeling System and its Application to Asthma Readmission Prediction

Abstract

The predictive modeling process is time consuming and requires clinical researchers to handle complex electronic health record (EHR) data in restricted computational environments. To address this problem, we implemented a cloud-based predictive modeling system via a hybrid setup combining a secure private server with the Amazon Web Services (AWS) Elastic MapReduce platform. EHR data is preprocessed on a private server and the resulting de-identified event sequences are hosted on AWS. Based on user-specified modeling configurations, an on-demand web service launches a cluster of Elastic Compute 2 (EC2) instances on AWS to perform feature selection and classification algorithms in a distributed fashion. Afterwards, the secure private server aggregates results and displays them via interactive visualization. We tested the system on a pediatric asthma readmission task on a de-identified EHR dataset of 2,967 patients. We conduct a larger scale experiment on the CMS Linkable 2008-2010 Medicare Data Entrepreneurs' Synthetic Public Use File dataset of 2 million patients, which achieves over 25-fold speedup compared to sequential execution.

Figure 1.

An overview of our predictive modeling system. We extract EHR information from the database from the hospital and store the information in these event sequence files through persistent web services running on a private dedicated server. These event sequence files are uploaded to the web service on the cloud.

Figure 2.

Screenshots of the cohort construction module (A), feature construction module (B), predictive modeling module (C) and the performance analysis module (D). The cohort construction module allows users to specify criteria for selection of cases and for identifying matching controls. In the predictive modeling module, the user may specify parameters for particular feature selection and classification algorithms. The performance analysis module allows users to visualize key performance metrics as well as top predictive features selected in feature selection.

Figure 3:

Timeline of modules run and elapsed time. The data-splitting, training and testing times refer to the run times for each respective step of cross validation. Times are shown in seconds (s).