As Anne Fischer wrote in her last blog How a Data Scientist Thinks about Risk Stratification, in order to predict risk, we need to first determine what “risk” is being measured. One important risk is that of being hospitalized. Risk of hospitalization or admission models have become more targeted, more personal and seemingly more prevalent. Further, they are very much in line with the goals of population health and the “quadruple aim” (improving patient care, reducing costs, improving the health of populations and improving the provider’s experience). If a person has proper ambulatory or outpatient care for a chronic disease, acute inpatient admissions related to that disease should be rare.
At Truven Health, we have been developing and maintaining risk models for decades. Many risk models, such as risk of cost, risk of mortality and risk of complications, are best used in aggregate within a population subset. This evaluation is typically done at a service line or patient group level, for example, all cardiovascular patients or all patients with a specific MS-DRG.
In today’s world of increasing interest in care management and targeted outreach, individual level risk models hold much promise. These models evaluate and interpret risk for each individual patient. Several years ago, we started a journey to develop new targeted risk models – including risk of hospitalization – to meet this increasing business need. We focused on specific diseases with the intent to have high predictive accuracy overall but particularly at the “tails”, meaning we must perform well at identifying those who have a high likelihood of being admitted in the near future. Or, in data science terms, model performance was measured by its sensitivity, specificity and positive predictive value for all patients above a specific risk threshold, with an emphasis on high sensitivity (that is, reducing false negatives).
To start, we chose to focus on diabetes, congestive heart failure (CHF), asthma, and more recently chronic obstructive pulmonary disease (COPD). These conditions run the gamut for both prevalence and admission risk, with asthma being most prevalent and least likely to result in admission versus CHF, which is least common but most likely to result in admission. These are also chronic conditions that are typically managed, at some level, which also fit our criteria.
In addition to specific diseases, we also focused on identifying risk at one-, three- and six-month intervals so that a care manager can better understand the risk at hand and be able to prioritize accordingly. Further, we report risk across several categories including “all-cause admissions”, “potentially avoidable admissions” (largely defined by AHRQ) and risk of “related admissions” which represents conditions that are considered to be related to the main condition as defined by Truven Health. Finally, along with each model’s risk score, we provide the patient attributes that are driving the risk score, whether it be a recent hospitalization, level of disease severity or even age. We believe that this additional insight gives the care manager a bit more background on the patient, helping to explain why the patient may have an increased risk value.
These models have been notably successful in terms of their predictive accuracy and our work will continue as we expand the number of diseases and work with our clients to help make the information actionable. The general trend toward person specific risk versus risk in aggregate will only grow and will become more refined as we, and the industry in general, are able to obtain and incorporate more personalized information about people.
Senior Director, Advanced Analytics