Predicting Hospital Overuse by Combining Market Segmentation and Advanced Analytics with Clinical Insights

IntroductionUnnecessary and excessive treatment is a persistent problem inflating health care expenditures. Because traditional regression model-based approaches to examining excess hospital utilization are limited in their... [ view full abstract ]

Introduction

Unnecessary and excessive treatment is a persistent problem inflating health care expenditures. Because traditional regression model-based approaches to examining excess hospital utilization are limited in their ability to account for the plethora of co-occurring conditions affecting diverse patient populations or for differences in contributing factors among more homogenous subgroups within those populations, they often produce results that have minimal utility for informing targeted clinical treatment to help prevent hospital overuse. A combination of clinical insight and alternative, advanced analytics that emphasize intersecting clinical contributors to excess utilization represents a more nuanced approach that may offer hospitals greater promise for identifying areas where tailored clinical interventions can be applied.

Methods

Informed by hospital leader perspectives of need, we used segmentation—a popular approach in business analytics—to identify a target group of high utilizers, followed by random forest prediction with diagnosis codes to identify conditions driving excess utilization among that subgroup. Inpatient utilization data (2010-2013) from a rural Eastern NC hospital were analyzed.

Results

Segmentation identified a target subgroup of 1,475 patients age 60 or older with two or more inpatient visits in a given year prior to transitioning to a long-term care facility that same year. Random forest prediction identified five diagnosis codes most associated with this subgroup: 1) chronic ulcer of the skin (i.e., bed sores), 2) septicemia, 3) delirium/dementia/cognitive disorder; 4) urinary tract infections; and 5) fluid and electrolyte disorder. Almost all of these medical conditions are preventable, which highlights opportunities for targeted conditions-based interventions to prevent avoidable hospitalizations prior to a long-term care transition.

Discussion

Clinicians recognize the need to tailor care to best meet patients’ needs. Targeting unique and internally more homogeneous subgroups of utilizers and identifying conditions-based predictors strongly associated with those subgroups, may provide hospitals with more useful information upon which to tailor clinical interventions.