Because health care spending tends to be concentrated among a few patients, researchers have worked on how to identify those patients proactively and to identify the diseases or conditions which cause most of the spending. A study in Health Services Research examines a new method to categorize health spending. (HSR Article) A lot of the work on the contribution of medical conditions to overall spending has been in the risk adjustment arena, whereby plans are paid more for enrollees with a larger number of expensive medical conditions. The medical conditions are typically identified via diagnosis codes submitted by providers to payers. Payment adjustments are created by estimating the spending associated with various parameters, like age, sex and number of diagnoses or conditions. This author uses a machine-learning technique to try to improve on the relationship between medical condition identification and spending. A large claims database on commercial members was the basic data source. The exact technique is called nonparametric double robust targeted learning, which is a mouthful and if you can actually understand the math underlying it, God bless you. Basically, with a longitudinal database, the computer can “learn” what people with certain conditions in one year will end up spending in future years. The advantage is supposedly that it is more flexible in identifying associations and the end goal is to understand the average contribution of specific conditions to total spending.
74 medical conditions were used to analyze their role in spending. Of these 26 met a threshold of prevalence needed to ensure robust results. The largest prevalence was major depression and bipolar disease, followed by various cancers, but overall prevalence of these major conditions was low. The top five most expensive medical conditions were identified as multiple sclerosis (thank you, drug companies); congestive heart failure; lung, brain and other severe cancers; major depression and bipolar disease; and chronic hepatitis (thank you, drug companies). Having any of these conditions added at least an average of $10,000 a year in annual health spending, with multiple sclerosis adding $67,011. The list produced by this method differs, as does the size of the spending contribution, from that resulting from standard statistical procedures. Some of the differences were quite large. Aside from helping to understand what diseases are contributing most to health spending and spending growth, the paper points out the potential error in assuming that any method is highly accurate. Constant testing and refinement is needed, particularly since the results are used to determine how much someone gets paid. It is also notable that several of the highest cost categories are largely driven by drug expenses, just another example of the literal price we pay for excessively high charges for many branded drugs. And since this is a commercial population, one of the highest cost diseases–dementia–is very under-represented. Finding better ways to treat that disease is critical as the population ages.