How to use AI to circumvent hospital readmissions and improve patient and business outcomes
Patient readmission presents a substantial financial burden to healthcare systems. As value-based reimbursement models replace fee-for-service payments, especially Medicare, indicators of patient readmission (e.g. low-quality care) significantly impact hospital business operations and the bottom line.
Hospital readmissions cost Medicare roughly $26 billion annually, with about $17 billion spent on avoidable hospital trips after discharge. To crack down on hospitals with high readmission rates and improve patient health outcomes, the Centers for Medicare and Medicaid Services apply payment penalties (a downward payment adjustment up to three percent across all services provided to Medicare beneficiaries) to hospitals with higher than expected readmission rates. According to a study by CMS Patient readmission, 2,599 hospitals or 82 percent of participating organizations will receive reduced reimbursement under the Hospital Readmissions Reduction Program (HRRP) in fiscal year 2019.
One of the biggest challenges to preventing readmissions is identifying patients at risk of readmission. Early identification enables hospital staff to provide requisite and necessary care during the initial inpatient admission. This empowers them to plan for more effective care transition interventions before and after the patient is discharged. Technology plays a crucial role in achieving this goal: Data analytics and predictive risk stratification tools enable hospitals to pinpoint which patients are the most likely to end up back in the hospital after discharge.
Today, the LACE score is the most commonly used approach to identify patients at risk for readmission. The score takes into account four variables: the length of stay; whether the patient was admitted through the emergency department or came voluntarily; whether or not the patient has more than one disease or disorder; and the number of emergency department visits in the previous six months before admission. The sum of the scores for these variables predicts the rate of readmission or probability of death within 30 days of discharge. Patients with LACE scores of 10 or greater are considered at high risk of readmission.
We assessed the accuracy of the LACE score using 31,885 electronic patient records for Medicare patients from two independent hospital systems in the U.S. and discovered:
- 16,159 false positives (50.6 percent): Patients being flagged as probable readmission candidates when they were actually not readmitted within 30 days. False positives alert the hospital system to expend more resources during their stay. Over-allocation of hospital resources impacts hospital profitability under value-based reimbursement models.
- 743 false negatives (2.3 percent): Patients were flagged as not probable readmission candidates when they were actually readmitted within 30 days. A readmission not only results in the hospital failing to receive reimbursement for the cost incurred, but it could also be subject to a downward payment adjustment – especially with Medicare cases.
To more accurately predict patient admission and reduce hospital financial burden, we built a Predictive Readmission Calculator (PRC) using artificial intelligence (AI). It predicts the probability of patient readmission based on electronic patient health records data. The PRC yields highly accurate predictions through the use of an artificial neural network and ultimately recognizes the relationship between patient medical records and readmission.
Comparing predictive power of LACE and PRC
Of the 31,885 patient health records in our study, 3,929 patients were identified as being readmitted within 30 days of an initial discharge (re-admission status as ‘Yes’) and 27,956 patients were identified as not readmitted (re-admission status as ‘No’). Below is an Intersection Matrix to illustrate the predictive power of both models run on the same population data set.
Table 1. Intersection Matrix of readmission status predictions using the LACE score and PRC to show the concordance between the LACE score and the PRC.
|n=31,885||PRC Predicted Yes||PRC Predicted No|
|LACE Predicted Yes||2876||16469(a)|
|LACE Predicted No||583(b)||11957(a)|
As evidenced by the table above, hospitals using the PRC model instead of LACE would be able to:
- Reduce allocation of additional resources and costs towards care for 4,512 patients (difference in values) flagged unnecessarily as potential readmissions by LACE.
- Reduce the risk of reimbursement from potential readmissions for 583 patients that LACE was unable to predict.
Quantifying business outcomes
Hospitals penalized for high readmission rates have received low patient experience scores from the Hospital Consumer Assessment of Healthcare Providers and Systems (HCAPHS). In a similar vein, hospitals that scored high on HCAPHS have lower readmissions.
Readmission costs vary by medical condition, but we can calculate the cost savings potential of moving from LACE to PRC by using the average cost of an all-cause readmission which according to the Healthcare Cost Utilization Project is $11,200. By accurately predicting 583+ patients that were likely to be readmitted, PRC can potentially save these hospital systems $6.5 million.
The high price tag associated with LACE’s inaccurate predictive model for patient readmissions results in significant cost burdens for hospitals. The Predictive Readmission Calculator provides superior predictive power to the LACE score in terms of the number of false-negative and false-positive results. By utilizing an AI-driven predictive calculator like the PRC, hospitals can start down a pathway to exceptional patient care and reduced patient readmissions costs.
Article originally appeared at Healthcare Business Today.
To download the whitepaper, click here.
ABOUT THE AUTHOR
Peter A. Noble is a Senior Data Analyst at PK, where he develops big data and analytics solutions that help healthcare companies realize competitive advantages. He is a former professor at the University of Washington and Alabama State University. Peter holds a doctorate in applied microbiology from the University of Saskatchewan and has held research associate and post-doctoral fellow roles at the University of Maryland Baltimore and University of California, Irvine.