Authors:
Anima Singh
and
John Guttag
Affiliation:
Massachusetts Institute of Technology, United States
Keyword(s):
Machine Learning, Collaborative Filtering, Latent Factor Models, Intensive Care Units, Prescription Omissions.
Related
Ontology
Subjects/Areas/Topics:
Artificial Intelligence
;
Biomedical Engineering
;
Business Analytics
;
Clinical Problems and Applications
;
Data Engineering
;
Data Mining
;
Databases and Information Systems Integration
;
Datamining
;
Decision Support Systems
;
Enterprise Information Systems
;
Health Information Systems
;
Pattern Recognition and Machine Learning
;
Sensor Networks
;
Signal Processing
;
Soft Computing
Abstract:
Medication errors in critical care are frequent and can lead to adverse consequences. One important category of errors is prescription omission, i.e., failure to prescribe a potentially useful medication. Studies have shown that failure to prescribe a medication can result in adverse consequences leading to patient morbidity or even patient mortality (Aspden et al., 2007; Olsen et al., 2007). In this paper, we present a machine learning based approach to building a system that can be used to provide physicians with an ordered list of possible omissions. We investigated three different collaborative filtering approaches as well as simple prevalence and co-occurrence methods. When evaluated on over 19,000 ICU admissions, each of the CF approaches outperformed both prevalence and co-occurrence based methods. This work highlights the importance of capturing a multi-scale view of the prescription data for the task of identifying omissions. Our results suggest that latent factor models and
neighborhood models are better at capturing different kinds of omissions. Latent factor
models demonstrated improved performance on identifying omission of rarely prescribed medications while neighborhood models were slightly better at identifying omissions of commonly prescribed medications.
(More)