1:50 - 2:10 pmSunday, September 18
LK 120
Not your average guideline: clinical pathway learning using patient-centered evidence
LK 120
Not your average guideline: clinical pathway learning using patient-centered evidence
Carnegie Mellon University
Background Clinical pathways adapt evidence-based recommendations from clinical practice guidelines (CPGs) to local practice workflows and are used by more than 80% of all US hospitals. They aim to reduce... Read more

Description

Background 
Clinical pathways adapt evidence-based recommendations from clinical practice guidelines (CPGs) to local practice workflows and are used by more than 80% of all US hospitals. They aim to reduce variations in treatments and support clinical decision making when faced with multiple or ambiguous care options, thus improving the quality of care and controlling costs. This research investigates data-driven clinical pathway development by analyzing patients’ demographic and detailed treatment data captured in electronic health records (EHR) as part of routine care delivery.

Method
We derive clinical pathways from data by modeling the longitudinal treatment process, including office and hospital visits and laboratory results, as a dynamic Bayesian network. A novel representation of EHR data allows us to consolidate and temporally organize clinical activities of multiple categories in a succinct model. Sequential clustering is applied to the consolidated representation to identify latent patient subgroups. Furthermore, our model considers the differing durations for patients’ conditions and treatments to evolve, and allows identification of pathways that may improve outcomes. Variable selection incorporates domain knowledge and analytic techniques. The clinical pathway learning algorithm is generalizable to many domains and decision making settings, and is not limited by the number of variables or clinical categories.

Results
The model and algorithm are evaluated using EHR data from 2009 to 2013 on patients with chronic kidney disease and associated comorbidities and complications. Results indicate that the algorithm learns the most probable course and length of treatment for patients, given their particular demographic and clinical conditions, including visit types, diagnoses, procedures, medications and duration between visits. In addition, preliminary results show that the algorithm can predict the profile of patients’ health conditions and expected treatments in the near future, up to 65% and 75%, with and without inter-visit duration, respectively, and false negative rate as low as 0% in test instances using leave-one-out cross validation. We expect further improvements with larger sample sizes and parameter tuning.

Conclusion
Significant diversity in patients’ EHR data poses a challenge for efficient pathway learning and prediction. While partly due to the nature of the disease and variability in patients’ conditions, it also suggests a review of practices, both for treatment variations and inconsistent EHR usage, for improved data capture. Future steps include evaluation of the derived pathways by expert clinicians and inclusion of medical costs and critical clinical outcomes into the pathway model. Data-driven clinical pathways have the potential to help healthcare providers review their current practices against CPGs, standardize them, and customize them into personalized, patient-centered pathways of care delivery.

Contact Us

We're not around right now. But you can send us an email and we'll get back to you, asap.

Not readable? Change text. captcha txt

Start typing and press Enter to search