International Journal of General Medicine (Aug 2015)

Health behaviors and quality of life predictors for risk of hospitalization in an electronic health record-linked biobank

  • Takahashi PY,
  • Ryu E,
  • Olson JE,
  • Winkler EM,
  • Hathcock MA,
  • Gupta R,
  • Sloan JA,
  • Pathak J,
  • Bielinski SJ,
  • Cerhan JR

Journal volume & issue
Vol. 2015, no. default
pp. 247 – 254

Abstract

Read online

Paul Y Takahashi,1,2 Euijung Ryu,3 Janet E Olson,3 Erin M Winkler,4 Matthew A Hathcock,3 Ruchi Gupta,3 Jeff A Sloan,3 Jyotishman Pathak,3 Suzette J Bielinski,3 James R Cerhan3 1Division of Primary Care Internal Medicine, 2Department of Internal Medicine, 3Department of Health Sciences Research, 4Center for Individualized Medicine, Mayo Clinic, Rochester, MN, USA Background: Hospital risk stratification models using electronic health records (EHRs) often use age and comorbid health burden. Our primary aim was to determine if quality of life or health behaviors captured in an EHR-linked biobank can predict future risk of hospitalization. Methods: Participants in the Mayo Clinic Biobank completed self-administered questionnaires at enrollment that included quality of life and health behaviors. Participants enrolled as of December 31, 2010 were followed for one year to ascertain hospitalization. Data on comorbidities and hospitalization were derived from the Mayo Clinic EHR. Hazard ratios (HR) and 95% confidence interval (CI) were used, adjusted for age and sex. We used gradient boosting machines models to integrate multiple factors. Different models were compared using C-statistic. Results: Of the 8,927 eligible Mayo Clinic Biobank participants, 834 (9.3%) were hospitalized. Self-perceived health status and alcohol use had the strongest associations with risk of hospitalization. Compared to participants with excellent self-perceived health, those reporting poor/fair health had higher risk of hospitalization (HR =3.66, 95% CI 2.74–4.88). Alcohol use was inversely associated with hospitalization (HR =0.57 95% CI 0.45–0.72). The gradient boosting machines model estimated self-perceived health as the most influential factor (relative influence =16%). The predictive ability of the model based on comorbidities was slightly higher than the one based on the self-perceived health (C-statistic =0.67 vs 0.65). Conclusion: This study demonstrates that self-perceived health may be an important piece of information to add to the EHR. It may be another method to determine hospitalization risk. Keywords: alcohol, aging, multiple chronic conditions, EHR, health behavior, hospitalization, quality of life