Understanding the concentration and persistence of high health spending for a patient is critical to controlling overall health spending. A variety of methods attempt to capture this predictive modeling capability, but none is overwhelmingly successful. Researchers at Stanford have created some new models that appear to have improved performance. (Stanford Paper) The researchers developed their model by using comprehensive health use and other data from the population of West Denmark over the study period, about 2,150,000 people. The used some existing models for risk adjustment/cost prediction and created some new ones using statistical learning approaches. They tracked and tested performance against data from 2004 to 2011. In this time frame, 315,000 of the individuals had at least one year of high health spending (i.e., they were in the top 10%). The majority, 51%, had only one high-cost year, and among the remainder who had more than one, the years were often not consecutive. Having consecutive high-cost years was very predictive, however, of having another high-cost year.
In 2011, 155,795 high-cost patients accounted for 73% of total health spending. Half of these patients had not had a high-cost year in either 2009 or 2010. People tend not to have a lot of consecutive high-cost years, because, well, if you have that many health issues you likely die, and in fact 16% of persistent high-cost patients die within two years. The majority of high-cost patients were “cost bloomers” meaning they had not shown a tendency in a recent year to be high-cost. So are there factors that may help identify the potential cost bloomers. They were less likely than persistent high-cost patients to have had an inpatient hospitalization and they had fewer chronic diseases. They were more likely to have a congenital anomaly, a genitourinary disease or a nervous system disease. They were younger and had lower one or two-year mortality rates. The best performing model evaluated by the researchers identified 49% of costs ultimately attributed to cost bloomers in 2011, but most of the models performed comparably. This work is important, but realistically, I am not led to believe that the model would generate a lot of clinically useful data. To be useful for care management, a model would need to predict with great specificity who is likely to be high-cost in the coming year. Otherwise, people who are mis-identified as going to be high-cost have expensive care management efforts wasted on them, and people who are not picked up but are high-cost are a missed care management opportunity. One thing that would be useful is an analysis of whether people in say, the second decile in a year are more likely to move to the top decile in the next year. It is all a work in progress.