The Centers for Medicare and Medicaid Services has proposed to financially penalize hospitals that have 30-day readmission rates above the national mean. As a result hospitals caring for disadvantaged populations with more needs might be penalized by current 30-day readmission models that do not include measures of social risk and functional status of the patients served. These are two important variable domains that directly impact a patient's ability to manage their disease. Social risk factors (e.g. living alone, social support, marginal housing, and alcohol abuse) and functional status (e.g. mobility, fall risk) are rarely present in administrative data, which is why so few readmission models include this data. Yet many of these variables are available in electronic health records (EHR) and the advancement of the field of informatics has made the extraction of these data feasible. These variables may improve the discriminative ability of 30-day readmission models which currently explain little of the variation in readmission rates among patients. We propose to improve 30-day readmission models by extracting measures of social risk and functional status from the EHR using the novel method of Natural Language Processing (NLP). We will combine administrative data (VA and Medicare) and data extracted from the national EHR in the VA for 6000 patients 65 and older in 2011 to improve upon currently available 30-day hospital readmission risk prediction models for congestive heart failure (CHF), acute myocardial infarction (AMI), pneumonia and stroke. We have chosen these conditions because hospital-level 30-day readmission rates for these conditions (CHF, AMI and pneumonia) are currently or will soon be (stroke) publicly reported. Our proposal has two goals: 1) to develop, test and evaluate automated NLP algorithms designed to extract measures of social risk and functional status from the EHR and 2) to understand the impact of these two novel domains on 30-day readmission across four conditions with fundamentally different post-discharge hospital course and disease trajectories. We propose a paradigm shift in the understanding and obtainment of factors predictive of 30-day readmission. Our overarching hypothesis is that social risk factors and functional status which directly influence a patient's self-management ability are critical factors predictive of 30-day readmission, can be extracted from the EHR, and should be included in risk prediction models. The development of better risk prediction models will allow the identification of patients at highest risk of readmission and facilitate post-discharge interventions in their care. In addition, if social risk factors and functional status are criticalin explaining variation in 30-day readmission rates, then hospitals that care for patients with a higher burden of social risk and functional needs may be inappropriately penalized by current risk predictions models that lack these measures. Also, as more hospitals adopt EHRs, we need to study more advanced technologies such as automated NLP as tools to efficiently extract information and to inform health systems about the characteristics of the patients they serve.
About a fifth of all Medicare patients are readmitted within 30 days of hospital discharge, potentially exposing the patients and their families to distress and unnecessary costs. An improved understanding of who is at highest risk of readmission will allow hospitals to efficiently intervene in the care of patients who may benefit the most from post-discharge interventions.
|Keyhani, Salomeh; Myers, Laura J; Cheng, Eric et al. (2014) Effect of clinical and social risk factors on hospital profiling for stroke readmission: a cohort study. Ann Intern Med 161:775-84|
|Blum, Alexander B; Egorova, Natalia N; Sosunov, Eugene A et al. (2014) Impact of socioeconomic status measures on hospital profiling in New York City. Circ Cardiovasc Qual Outcomes 7:391-7|