Implementation of a Machine Learning Algorithm in an Urban Hospital Setting to Reduce Vascular Surgery Readmission Rates
Anahita Dua, MD MS MBA1, Gilbert R. Upchurch, Jr., MD2, Sapan S. Desai, MD PhD MBA3.
1Stanford Hospital, Palo Alto, CA, USA, 2University of Florida, Gainsville, FL, USA, 3Northwest Community Hospital, Arlington Heights, IL, USA.
OBJECTIVES: Readmission rates in vascular surgery patients are higher than other surgical specialties. We developed a machine learning algorithm that predicts which patients are likely to be readmitted within 30 days following a vascular procedure, and implemented the model in an urban hospital setting. METHODS: 1,088 patients who underwent peripheral angiograms, aortic bypass, peripheral bypass, major amputations, or minor amputations were analyzed to create a machine learning algorithm that predicted which patients were likely to be readmitted. This predictive model had a sensitivity of 53.4%, specificity of 89.5%, positive predictive value of 60.0% and negative predictive value of 86.8%. Overall accuracy of the model was 81.3%. A performance improvement program was initiated to mitigate the risk of readmission by targeting patients most likely to be readmitted using our machine learning model. Any patient predicted by this algorithm to be readmitted had a social work consultation, physical therapy consultation, occupational therapy consultation, and a consultation with the hospitalist specifically to help manage their medical comorbidities. Financial analysis of additional supportive costs was reported.
RESULTS: Patients required an additional 0.8 +/- 1.2 days in the hospital as a result of additional support once flagged by the predictive model. Specific efforts to reduce transfer to SNF were also made through the increased use of home with home health and closer follow up in clinic. Our baseline readmission rate decreased from 22.7% to 9.7% (57.3% decrease). For the 268 patients who underwent procedures in the test group following the PI rollout, an estimated 35 readmissions to the hospital were avoided. A financial analysis was completed to determine potential cost savings. Of the 268 patients in the test group, our algorithm flagged 81 patients for additional supportive services. This led to an estimated $129,600 increase in costs due to the extended length of stay and any additional testing (total cost of care increase). Avoiding 35 readmissions led to an estimated $420,000 decrease in costs, leading to an overall $290,400 decrease in total cost of care. Improved reimbursement with pay for performance and HCAHPS are an additional savings. CONCLUSIONS: This is the first instance of using machine learning to develop an accurate model of vascular readmission leading to a subsequent performance improvement program to successfully reduce readmission rate.
Back to 2019 Abstracts