Efforts to manage care to improve quality and cost outcomes are greatly facilitated by accurate identification of likely high-cost patients, especially if that identification occurs ahead of time. A lot of tools have been built for this “predictive” modeling capability, but their utility is modest at best, as we explored in another recent post. Research published in Medical Care describes a slightly different approach, one which focuses on identifying “trajectories” of patient utilization and cost. (Med. Care Article) The researchers used full claims data from 2009 to 2011 on all Aetna fully-insured commercial (it appears) members. The study had two objectives; one was to describe and classify one-year spending patterns and the second was to identify patients with particularly high utilization who could then be targeted for care management efforts. A number of characteristics in addition to spending were collected for the patients, including demographic and clinical ones. Utilization in various categories, benefit plan design and certain drug use measures were also applied. The ACG risk adjustment methodology was applied.
Patients were identified as high-cost if they were in the top with percentile of spending or if they were in a high-cost group based on trajectory modeling which looks at patterns of utilization and spending over time. A number of models were created to evaluate the predictive ability of including certain subsets of characteristics and certain statistical techniques including R-squared methods which evaluate the proportion of variance explained, and C-statistic, which evaluates ability to discriminate how accurately patients were placed in a particular category. In general, models using trajectory modeling, were more accurate and had greater potential utility than those just classifying patients as high-spenders. Figure 1 of the study is particularly fascinating as it shows seven common trajectories that were identified by the modeling. About 31% of patients actually fell into the high-cost trajectory, which was patients who had pretty consistent spending of an average of around $1200 per month during the year. (Note that people who received no medical care, or at least didn’t submit a claim, aren’t included in the analysis and this could be a significant group of people, so the 31% is likely exaggerated.) That is a lot of people to potentially intervene on. It suggests that further subdividing is needed to target patients even within this group. A moderate spending group of around 26% had about $400 per month spending and a low-cost group, at 13% of the population, had almost no spending.
The other trajectories showed substantial unevenness in average monthly spending throughout the year, some declining, some growing. This appear to me to be largely influenced by the timing of the study, which forces a calendar year onto the timing. Looked at over a multiyear period, these patterns would likely converge and appear to be linked to acute care episodes with some expense–a surgery, a car accident, a cancer episode which resolves relatively quickly. Two major takeaways from the study are that most classifications, even ones based on relatively simply pharmacy claim indicators, can have meaningful predictive ability; and that the trajectory approach adds significantly to that predictive ability. It is also noteworthy that the study looked at “intervenable” characteristics as a factor; these include obesity, depression, etc., conditions that have known working interventions. This is also useful for guiding care management programs. This is a very useful study.