Researchers Find More Accurate Model for Estimating Patient Costs

August 21, 2019 - Using single diagnostic codes and leveraging “present on admission” designations improved Medicare payment models, predicting total patients costs within 30 days of hospitalization better than the current grouped diagnostic code method, a study published in JAMA Open Network found.

Changing the variables currently used by CMS in patient cost prediction models could have serious implications for research, benchmarking public reporting, and calculations of population-based payments in programs like Medicare Advantage, researchers said.

“With a rapidly increasing number of beneficiaries enrolled in Medicare Advantage, the importance of risk models to determine payment is growing,” they wrote in the study. “Better models could also have relevance to planning at the health system level as the focus on population health increases and the need to identify people with the highest risk of incurring costs increases.”

Predicting total patient costs for particular conditions is becoming increasingly important as the industry transitions to value-based reimbursement and population health management. Payers rely on cost estimates to calculate population-based payment rates, as well as benchmarks used to determine if providers will be penalized or rewarded for cost performance.

Medicare is one of those payers. In 2015, CMS started to publicly report payments for hospitalizations related to acute myocardial infarction (AMI), heart failure, and pneumonia. The agency plans to use the data as part of the Hospital Value-Based Purchasing Program in 2021, and it employs a similar model to calculate Medicare Advantage payments.

But researchers, led by Harlan Krumholz at the Department of Internal Medicine at the Yale School of Medicine, found that CMS’ current methodology for estimating patient costs is not the most accurate.

Currently, CMS predicts patient costs by grouping ICD-9-CM diagnosis codes into clinically coherent categories. By grouping the codes, the model reduces the number of candidate variables but may reduce the performance of the models by combining different codes with different relationships to outcomes, researchers explained.

Using single codes rather than condition categories more accurately predicts patient costs, Krumholz and colleagues found using data from Medicare fee-for-service hospitalizations for AMI, heart failure, and pneumonia at acute care hospitals from July 1, 2013, through September 30, 2015.

“The individual-codes models produced better discrimination, goodness of fit, and predictive range, with similar calibration,” they wrote.

Using present on admission codes, which have been purposely excluded from patient cost prediction models due to their inability to distinguish some conditions as present on admission or as occurring during the hospitalization, could also improve patient cost prediction models, researchers stated.

“With POA codes, it is possible to know that the condition was not a complication during hospitalization, allowing the use of more diagnoses from the index with the possibility of model improvement,” they wrote.

A model using both single and present on admission codes improved the accuracy of identifying patients at the upper and lower ends of cost, and the model did so without adding to clinician data entry, the study uncovered. For example, the upper cost prediction range for patients with AMI increased from $55,387 to $113,446, while the estimate on the lower end declined from $10,993 to $8,204.

The difference in patient cost estimates could have significant consequences for hospitals.

“Many hospitals that the models classified as better or worse than the national mean payment changed categories, even as the overall percentage that the models identified as significantly different from the mean changed only slightly,” the study stated. “The presumption was that the better performing model was more appropriately identifying outliers by better accounting for the case mix of the hospitals.”

The finding represents an opportunity to improve patient cost prediction models. But stakeholders will need to take some steps to incorporate single code use and present on admission designations for improved cost estimates, including testing the new model using ICD-10-CM codes, researchers said.

“The findings open the possibility of improving research, performance assessment, and payment determinations by improving characterizations of case mix as well as improving population health by being better able to identify individuals at high risk of incurring high costs,” the study concluded.