Predictive Modeling of Health Care Associated Infections (HAIs) Using National Linked Individual Health Care Data in Scotland

• Activity Number: 407 - Data Science Applications in Epidemiology
• Type: Contributed
• Date/Time: Tuesday, August 1, 2017 : 2:00 PM to 3:50 PM
• Sponsor: Section on Statistics in Epidemiology
• Author(s): Kimberley Kavanagh* and Jiafeng Pan and Chris Robertson and Marion Bennie and Charis Marwick and Colin McCowan
• Companies: University of Strathclyde and University of Strathclyde and University of Strathclyde and University of Strathclyde and University of Dundee and University of Glasgow
• Keywords: Predictive modelling ; regression ; machine learning ; analytics ; healthcare

Abstract

The use of patient history data to support patient management could be delivered by embedding predictive analytics into existing linked data held by the National Health Service. We consider creation of models for the risk of acquiring a HAI, here Clostridium difficile (CDI), which could aid clinical decision making. Using linked national individual data on community prescribing, hospitalisations, infections and death records we created a population matched case-control study to examine the impact of prior healthcare contact, previous prescriptions, comorbidities, and cumulative antibiotic exposure, to the risk of CDI acquisition. Predictive models are built using both regression and machine learning and their ability to predict CDI infection assessed using cross-validation. Regression modelling highlights previous cumulative prescribing as a predictive factor (1-7 days exposure OR=3.8 rising to OR=17.9 for 29+ days) and machine learning highlights comorbidities being of dominating importance. Both approaches achieve high predictability (AUC~0.8) and sensitivity and specificity values in the range 0.7-0.8. Possible model stratification to improve prediction will also be considered.