A few people, often near the end-of-life, are well-known to account for the majority of health spending. Figuring out who those people are going to be in the near term has been a difficult task, but is necessary if spending is going to be significantly reduced. (Stanford Paper) The researchers attempted to create an algorithm which did a better job than current methods of predicting who was likely to die, in order for health organizations to manage the care of those people to avoid unnecessary services and reduce spending. Currently, as much as 60% of deaths still occur in hospitals, where intensive and aggressive care is common, even though most people say they want to die at home. While many hospitals claim to offer palliative care, few patients who are appropriate for it actually end up using it. Some of this may be patient preference; some may be financially driven, but some may just be uncertainty on the part of patients and physicians about prognosis and therefore the appropriate course of treatment.
Prior research suggests that physicians are too optimistic in making prognostic judgments. There are some existing scales and tools designed to improve these judgments but they typically aren’t useful for advance prediction and they take clinician time just to provide inputs. These researchers wanted to try to identify which patients are most likely to die in the next year. They took advantage of the ubiquity of EHRs and the state of machine learning or artificial intelligence techniques to try to create a method that would avoid requiring clinician time and would basically use large amounts of historical data to identify correlations that could then be tested for prognostic value, both of patients in general, but also more specifically of patients currently admitted to the hospital. From their data base they ended up with around 220,000 patients, each of which was assigned a prediction year. About 7% died in the year used for the study period, and of the subset of patients who were admitted to a hospital at the time of the start of the year, 11% died. Over 13,500 separate data points ended up being evaluated. One finding was that even patients who did not die within the year start date could be identified as having a terminal illness and therefore benefiting from palliative care. The algorithm was able to identify a fair percent of the patients who were likely to die and therefore be candidates for palliative care. This is an interesting and potentially valuable tool for helping with managing care for expensive patients. It needs further real-life testing, but seems like a good approach.