Study overview
As part of our planned participation in the major bowel bundled payment program, CMS provided data on a cohort of patients at our institution who were treated between 2014 and 2016 and who would have met the criteria for inclusion in the BPCI-A program. This time period represented our “baseline” performance and was used by CMS to calculate price targets. Based upon these historical costs, the negotiated reimbursement amount from CMS for one care episode in the BPCI-A Major Bowel Bundle Program was set at $60,000. CMS would allocate $60,000 for all of the care required for a single patient’s 90-day care episode, but UCSF would be responsible for the cost of any care delivered over that amount. Any patient with a care episode exceeding $60,000 was defined as “high-cost”. We utilized the itemized claims data to identify areas for improvement prior to program initiation. The BPCI-A program for major bowel includes patients undergoing intra-abdominal operations that involve some element of a bowel resection and fall into the qualifying Medicare Severity Diagnosis Related Groups (MS-DRGs 329, 330, and 331). As such, the patients included in our study underwent either elective or emergent small bowel, colon, and/or rectal procedures for acute conditions (e.g., ischemia, obstruction, or perforation), colorectal cancer, benign disease, and inflammatory bowel disease. All included patients had Medicare as their primary insurer and were enrolled in Medicare Part A & B. Patients who were enrolled in Medicare Advantage were excluded from the CMS program.
CMS provided itemized claims data that included all costs within the 90-day care episode, which began when each patient was admitted for a procedure that fell within a designated DRG. Data included every hospitalization within the 90-day episode, including the “Anchor Stay” (the initial acute care hospitalization) and any readmissions following the Anchor Stay. All claims related to professional fees, outpatient clinic visits, durable medical goods, hospice, pharmacy, rehabilitation or skilled nursing service, and home service (including home health, physical therapy, and ostomy) claims after the index hospitalization were provided. For analysis, contributions from less frequent cost categories were aggregated into an “Other” category and included inpatient psychiatry, inpatient rehabilitation, transfer, physical therapy, and hospice costs.
The preoperative and procedure-related information was abstracted from the electronic health record for each patient and was entered into the ACS NSQIP SRC to determine the predicted risk of readmission. The ACS NSQIP SRC is a well-validated tool used in hospitals around the USA, which utilizes 20 patient variable inputs as well as the planned procedure to predict 30-day outcomes in patients (including readmission risk) following surgery.
This study was deemed exempt by the Institutional Review Board at UCSF.
Statistical analysis
Itemized claims data were aggregated for the cohort by calculating the mean across each claim category for all patients. The variability of cost within each category was analyzed and compared by calculating the coefficient of variation (cv)—a standardized measure of dispersion—for each cost category. The average relative contribution of each cost category was also calculated as a percentage of the average total care episode cost. A cost category was targeted for improvement based on the combination of a high coefficient of variation with a high relative contribution to the total care episode cost. A relative risk was calculated to evaluate whether readmission status affected the likelihood of having a high-cost care episode (defined as >$60,000). A logistic regression model was used to assess the predictive performance of the ACS NSQIP SCR’s anticipated readmission risk with the actual 30-day readmission rate realized in the data. All participants were included in the logistic regression model. The area under the receiver operator curve (AUROC), also known as the c-statistic, was calculated to assess the discrimination ability of the model. All hypothesis tests were two-sided, and the significance threshold was set to 0.05. The statistical analyses were performed using R Studio (RStudio Team, 2019)—an open-access programming platform for statistical computing.