It is increasingly recognized that optimal treatment is not the same for every patient - it depends on the individual patient's circumstances. One important factor that determines the optimal clinical management is the patient's life expectancy, which determines the temporal horizon within which medical decisions have to operate. For example, while adding an extra diabetes medication may save a 40-year-old individual from going blind or developing kidney failure 20 years later, it will not bring any benefits to an 80-yer old with a metastatic malignancy who is expected to live only a few months. Consequently, it is critical that when we measure quality of care, suggest treatment options to clinicians through clinical decision support or compare different treatment strategies, we take into account the patient's life expectancy. However, currently there are no methods available that can do this with sufficient accuracy. Most commonly used techniques to assess a patient's mortality risk draw primarily on administrative data and other structured data fields in electronic medical records. This approach leaves out a large amount of information that is only available in narrative documents such as provider notes, radiology reports, etc. In this project we propose to test the hypothesis that application of two novel technologies - artificial intelligence technique Dynamic Logic and natural language processing (NLP) - could leverage the information in narrative electronic documents to significantly improve the accuracy of identification of patients with low life expectancy. Dynamic Logic allows to circumvent the challenge of combinatorial complexity that limits the number of variables and their combinations that can be considered as predictors of an outcome by most currently used analytical methods. Dynamic Logic makes use of a limited number of iterative approximations to reduce the complexity of a problem with multiple predictor variables from exponential to approximately linear. Utilization of Dynamic Logic will allow us to greatly increase the richness of the models for identification of patients with low life expectancy and ultimately improve their accuracy. Information such as the patient's functional status that is usually only found in narrative documents may be critical to improving accuracy of identifying frail patients at high mortality risk. Modern NLP techniques can effectively identify key concepts in medical text but until now analytical methods allowed consideration of only a few of pre-selected concepts in prediction models. Combining NLP with Dynamic Logic will allow us to greatly expand the number of concepts from narrative text that could be included in the life expectancy prediction model, likely leading to a considerable improvement in accuracy. In the proposed translational multidisciplinary project our team that will include experts on artificial intelligence, natural language processing, analysis of data in electronic medical records, and geriatrics will test whether a combination of Dynamic Logic and NLP improves identification of patients at high risk of death.

Public Health Relevance

Identification of patients with limited life expectancy is important for accurate measurement of quality of care delivered by clinicians, clinical decision support in electronic medical records and research that compares different treatment options for patients. However, currently available methods suffer from low accuracy or use information that is not widely available. In this project we propose to combine two advanced computational technologies - artificial intelligence technique Dynamic Logic and natural language processing - to improve accuracy of identification of patients with low life expectancy.

Agency
National Institute of Health (NIH)
Institute
Agency for Healthcare Research and Quality (AHRQ)
Type
Research Project (R01)
Project #
5R01HS024090-04
Application #
9481256
Study Section
Healthcare Information Technology Research (HITR)
Program Officer
Nourjah, Parivash
Project Start
2015-08-01
Project End
2019-05-31
Budget Start
2018-06-01
Budget End
2019-05-31
Support Year
4
Fiscal Year
2018
Total Cost
Indirect Cost
Name
Brigham and Women's Hospital
Department
Type
DUNS #
030811269
City
Boston
State
MA
Country
United States
Zip Code