UM CRMC RecuR Score Pilot

1. BACKGROUND Hospital readmission is an adverse health outcome that incurs significant cost to the healthcare ecosystem. While undesired, unplanned hospital readmission within 30 days of discharge is not uncommon. To improve care quality and reduce unnecessary healthcare costs, in 2013, the Center for Medicare and Medicaid Services (CMS) launched the Hospital Readmissions Reduction Program (HRRP) as part of the Value Based Purchasing (VBP) program to encourage better discharge care coordination. Under the HRRP program, hospitals with high readmission rate incur a payment reduction of up to 3 percent. Since the launch of the HRRP program, reducing hospital readmissions has elevated to a strategic priority of hospitals. Best practices to effectively reduce hospital readmissions while maintaining a healthy operating margin are sought after by hospitals across the country. Studies on most optimal intervention structure and intensity have yet to identify a single effective strategy, and the effectiveness and external validity of interventions in the literature remains uncertain. Across all patients at University of Maryland Charles Regional Medical Center (UM CRMC) between January 2019 and January 2022, the unplanned hospital readmission rate was 11% and this value is as high as 30% across certain highest risk groups. UM CRMC has implemented a Transitional Care Program with Nurse Navigators since 2011 that focuses on patients that are typically known to have a higher rate of readmission (patients whose primary reason for admission is Diabetes, Congestive Heart Failure \[CHF\], Chronic Obstructive Pulmonary Disease \[COPD\] and Hypertension). Despite genuine efforts to manage these patients and provide additional support to these patients prior to discharge and post-discharge, the readmission rate at UM CRMC has remained relatively unchanged over the past five years between 7-15% (mean 11% ± 1.5%) with no sustained year-over-year improvement. At the University of Maryland Medical System (UMMS), we have developed an artificial intelligence (AI)-powered risk score called the RecuR Score (Readmission Risk Score). The RecuR Score estimates the risk of 30-day unplanned readmission for patients both in-house and during the 30 days after inpatient discharge. Patients are grouped in one of five score levels (1-5), where a RecuR Score of 1 indicates the lowest risk of readmission and a RecuR Score of 5 indicates the highest risk of readmission. This risk score is retrained monthly using data from patients with encounters at UMMS hospitals. The target population is inpatients, currently in-house non-inpatients (Emergency Department, Observation Unit) who might become inpatients, and previous inpatients within 30 days of discharge. The score uses data from the UMMS electronic health record system (EHR), CRISP (Chesapeake Regional Information System for our Patients - the state-designated Health Information Exchange for Maryland), commercial and non-commercial claims, and the U.S. Census Bureau. A comparison of the performance of the RecuR Score compared to LACE and HOSPITAL on the same patients showed that the Area Under the Receiver Operating Characteristic Curve, sometimes known as the Area Under the Curve (AUC), of the RecuR Score significantly outperforms the other two metrics, even prior to discharge. While LACE is only available at discharge, HOSPITAL is described as most accurate at discharge, and even then, it is outperformed by the RecuR Score at 48 hours post-arrival when the RecuR Score is not at its best The literature review shows that efforts to reduce readmission rates are not consistently effective, and it has been difficult to extract a set of interventions that reliably reduces readmissions. Our team theorizes that efforts to reduce readmission rates are not effective because the patients are not adequately stratified into risk categories resulting in interventions not being used on the patients who will benefit the most from the interventions. This pilot addresses this issue by identifying patients at higher risk of readmission using the RecuR Score. The RecuR Score accurately identifies patients at high risk of readmission with an area under the ROC curve of 0.83. This higher risk population (limited to selected principal diagnoses and other inclusion and exclusion criteria) has a higher readmission rate (19.4%) than the hospital's overall readmission rate (11%), which results in a greater opportunity to reduce the readmission rate for the target population. To address the issue of identifying the most effective interventions, our team also theorizes that the interventions used are not robust, meaning that the impact of the intervention is insufficient. For example, most readmission intervention programs focus on phone calls post-discharge without considering a more complete view of the patient's situation. To address this issue, this pilot is implementing more complex interventions such as additional educational materials, a focus on the patient's readmission risk during interdisciplinary medical team huddles/care transition rounds, and multiple home healthcare programs that cover a broad spectrum of potential interventions. The expectation is that by accurately identifying the higher risk patients and having a broader view of the patients' situation with multiple interventions, we can reduce the 30-day unplanned readmission rate. 2. STUDY OBJECTIVES Offering more complex interventions that are higher intensity than those currently universally provided at UM CRMC. By targeting patients who are at a higher risk of readmission using a novel AI-based risk score, the RecuR Score, resources can be best allocated to those who need them most. These more intense interventions include additional educational materials, emphasis on a patient's readmission risk during their multidisciplinary team huddle, and home health services. For the study, we will only be targeting patients with a high readmission risk (based on the patient's RecuR Score), to test the efficacy of the standardized use of these resources. The first primary hypothesis for this study is that using a novel UMMS algorithm (RecuR Score) to identify patients at a higher risk of unplanned hospital readmission combined with enhanced pre-discharge and follow-up care interventions including, additional educational material about their health, a focus on their readmission risk during interdisciplinary team huddles, and home health care, can reduce 30-day unplanned hospital readmissions in this high-risk group by 30%. The second primary hypothesis in this fallback design trial, is that using the aforementioned enhanced interventions, there will be a 30% reduction of unplanned hospital readmission risk in patients determined to be a medium-high risk of readmission (RecuR Score level 2 or 3). 3. METHODS This is a parallel-group, two-arm, prospective, randomized, non-blinded, fallback design, controlled superiority study of the impact of a new transitional care model for patients determined to be at higher risk of 30-day unplanned hospital readmission conducted at UM CRMC. UM CRMC is an approximately 100-bed community hospital located in Charles County, Maryland and is part of the University of Maryland Medical System. Patients will be 1:1 (equally) randomized to Arm 1 or Arm 2 using a stratified randomization method with stratification by RecuR Score. Participants assigned to Arm 1 will receive the activities in Intervention "A" and the participants assigned to Arm 2 will receive the activities in Intervention "A" AND the activities in Intervention "B." As there is an apparent difference between Arm 1 and Arm 2, neither the patients nor providers will be blinded to the study assignment. Informed consent will be obtained and documented for this study. The target patient population is admitted patients (inpatients) at UM CRMC with a high readmission risk (RecuR Score) and/or specific admission diagnoses that are amenable to peri- and post-discharge interventions, where the goal is to reduce the readmission rate of this high-risk population. This study will use a fallback design with two primary endpoints. The first sequential primary endpoint is 30-day post-discharge unplanned hospital readmission in the overall study population. The second sequential primary endpoint is 30-day post-discharge unplanned hospital readmission in the medium-high risk population (RecuR Score 2 \& 3). The significance level (0.05) will be divided evenly between the two primary endpoints, for a significance level of 0.025 for the first primary endpoint and a reserved 0.025 significance level for the second primary endpoint. 4. DATA COLLECTION The patient's Electronic Health Record (EHR) from Epic will be the main source of information about the patient's demographics, admission diagnoses, length of stay, and outcomes data like subsequent hospital encounters in the 90 days post-discharge. The CRISP ("Chesapeake Regional Information System for our Patients") Admission Discharge Transfer (ADT) tool is the source that provides what hospital encounters the enrolled patients have post-discharge if they do not occur at UMMS facilities. 5. DATA ANALYSIS As this is a fallback method design, we will first test the first primary endpoint against a 0.025 significance level. If the first test shows statistical significance, we will be able to add the unused alpha level from the first test to the reserved alpha level, for significance level of 0.05 to test the second primary endpoint. If this test fails to show statistical significance, we will proceed and test the second primary endpoint at the reserved 0.025 significance level. Patient characteristics will be summarized and compared between Arm 1 and Arm 2 for both the overall and RecuR Score 2 and 3 populations. Most results will be compared using a test of difference for two proportions. If the data are multi-categoric, the Mantel-Haenszel method will be used for comparisons. If an event is rare (\<5%), the Fisher exact test will be used. Any continuous data will be compared using a t-test or Wilcoxon Rank Sum statistic. If data are greatly skewed, additional methods might be needed. These analyses will be used to present baseline and background characteristics, demonstrating that the analysis is balanced and representative of the general population, or pointing to areas where primary analyses might need to be adjusted to account for imbalances. All analyses of results for primary and secondary endpoints will be based on the Intent-to-Treat principle, where results are computed based on the Arm that the patient was randomized to rather than which interventions they actually received. All comparisons will be done for both the overall and medium-high risk study populations, assuming the study continues to the second primary endpoint. Comparison of hospital readmission will be performed using either z-tests, or Fischer's Exact test, depending on the rarity of the outcome of interest. If imbalances are noted in the baseline characteristics, adjusted regression models will be tested where indicated. Subgroup analysis will be performed, including a specific focus on differing readmission rates between the various RecuR Score risk levels.