Improving prediction of patients at high-risk for readmission using data from physician notes

Tuesday, June 14, 2016: 8:50 AM
G65 (Huntsman Hall)

Author(s): Amol Navathe; Feiran Zhong; Victor J. Lei; Frank Chang; Li Zhou

Discussant: Anupam Jena

In this project, we leveraged free text in clinical notes to extract features in the data that provided complementary information to claims and structured clinical data. We examined whether using social factor information on ischemic heart disease patients in the EHR can improve classification of patients at high risk for readmission. We sought to answer three key questions. First, can using information in physician narrative notes increase the frequency of detecting eight social factors versus billing codes and structured data elements in the EHR? Second, will the social factor variables enhanced with information from physician narrative notes be significantly associated with readmission? Third, will the enhanced social factors improve classification of patients at high risk for readmission?

Our approach for modeling readmission followed the CMS definition for 30-day all-cause unplanned readmission used in the Hospital Readmissions Reduction Program (HRRP). We included detailed risk-adjustment, first starting with Elixhauser comorbidities and patient characteristics from administrative data then incrementally adding in diagnoses, additional patient characteristics, lab results, and medications. Identification of social factors was compared between those represented in ICD-9 codes, structured data such as problem lists in the EHR, and extracted features from EHR notes. The results demonstrated that social factors are much more prevalent then represented by administrative data. For example, smoking is four times more prevalent when using information in the EHR than codes alone. The other social factors showed prevalence increases of five times for alcohol abuse, two times for drug use, 1.5 times for mood disorders, four times for chronic wounds, 1.2 times for homelessness, ten times for fall risk, and 45 times for poor social support.

We also showed that these social factors enhanced with clinical data are highly associated with readmissions, even when risk-adjusting with administrative and clinical data variables. For example, patients with poor social supports are 10 percent more likely to readmitted that a patient without poor social supports, patients with chronic wounds are 9 percent more likely to be readmitted than those without, and homeless patients are 50 percent more likely to be readmitted. Finally, we examined whether these enhanced social factors would improve prediction and classification of patients, in a way that could inform the high-risk care management program. We found that 14% of patients readmitted would have been re-classified from low-risk to high-risk and 16% of patients not readmitted would have been re-classified from high-risk to low-risk. These results suggest that the enhanced predictive models could improve the care management program targeting of high-risk patients.