Hostname: page-component-848d4c4894-tn8tq Total loading time: 0 Render date: 2024-06-30T01:55:03.510Z Has data issue: false hasContentIssue false

266 Inpatient Quality Indicators Risk-Adjustment Using Interactions Selected by Machine Learning Methods

Published online by Cambridge University Press:  19 April 2022

Monika Ray
Affiliation:
General Internal Medicine, University of California Davis Health and Center for Healthcare Policy and Research
Patrick S. Romano
Affiliation:
General Internal Medicine, University of California Davis Health and Center for Healthcare Policy and Research
Rights & Permissions [Opens in a new window]

Abstract

Core share and HTML view are not available for this content. However, as you have access to this content, a full PDF is available via the ‘Save PDF’ action button.

OBJECTIVES/GOALS: Predictive models for health outcomes often have poor calibration potentially due to interactions that are ignored by standard methods. Using AHRQ models for Inpatient Quality Indicator (IQI) 11 Abdominal Aortic Aneurysm Repair and IQI 09 Pancreatic Resection mortality, we hypothesize that identifying interactions may improve model calibration. METHODS/STUDY POPULATION: We used adult discharge data from 16 states obtained from AHRQ Healthcare Cost and Utilization Project (State Inpatient Database), California Department of Health Care Access and Information, and New York State Department of Health. We used AHRQ’s v2021-1 Clinical Classifications Software Refined (CCSR) with present on admission flags to create features for risk-adjustment. We compared the performance of Least absolute shrinkage and selection operator (LASSO) model and first-order interaction models estimated using Hierarchical Group Lasso Regression (HGLR), after splitting the data into training and test sets. C-statistics, area under the precision-recall curve and Hosmer-Lemeshow calibration plots are reported. Finally, logistic regression models with selected CCSRs were evaluated on the test set. RESULTS/ANTICIPATED RESULTS: IQI 11 has four strata: open and endovascular repair of ruptured aneurysms (39% and 21% mortality, respectively); open and endovascular repair of unruptured aneurysms (6% and 0.8% mortality, respectively). IQI 09 has two strata: with and without pancreatic cancer (2% and 2.5% mortality, respectively). Comparing the HGLR model (with interaction effects) with Lasso models (without interactions), we noticed meaningful improvements in discrimination and calibration. However, for IQI 09, the extremely low mortality rate did not result in good HGLR or LASSO models. Interactions involving CCSRs could be identified using the novel HGLR method, which improved model performance given a heterogeneous population in IQI 11 with a mix of high and low event rates, unlike the more homogeneous patient population in IQI 09. DISCUSSION/SIGNIFICANCE: Standard implementations of regression models fail to address critical issues that arise in healthcare data – (a) quadratic explosion of potential interactions that cannot be manually identified, and (b) categorical variables with multiple levels or values (e.g., age categories). We propose innovative use of HGLR to robustly address these issues.

Type
Valued Approaches
Creative Commons
Creative Common License - CCCreative Common License - BYCreative Common License - NCCreative Common License - ND
This is an Open Access article, distributed under the terms of the Creative Commons Attribution-NonCommercial-NoDerivatives licence (https://creativecommons.org/licenses/by-nc-nd/4.0/), which permits non-commercial re-use, distribution, and reproduction in any medium, provided the original work is unaltered and is properly cited. The written permission of Cambridge University Press must be obtained for commercial re-use or in order to create a derivative work.
Copyright
© The Author(s), 2022. The Association for Clinical and Translational Science