Decedents accounted for 13% of yearly healthcare expenditure at age 65 and above, but only 2.8% was spent on those who, according to our machine learning model, had a likelihood of dying of more than 50%. While the mean healthcare expenditure per day alive on a decedent was ten times that of a survivor, when comparing to an equally frail population of survivors, the mean expenditure per day alive on a decedent was only 2.5 times higher. The main strength of the study is the availability of data for an entire population, with rich health care and sociodemographic predictor data and registry coverage of 97% of all healthcare expenditure12, as well as the inclusion of communal care in addition to treatment. As healthcare expenditure in Denmark is tax-funded, differences will not be artefacts of differential insurance coverage and rates. Individual-level expenditure data, however, may be misestimated to some extent: Hospital costs are DRG rates which are averages and may not entirely correspond to the actual cost of treatment, and the computation of individual-level expenditure on nursing home and home care involve some amount of estimation and imputation. The study deals only with expected mortality at baseline, which may arguably be a limited indicator of cost-efficiency of healthcare spending, and other measurements such as quality-adjusted life years could have been taken into account.
The distribution of predicted mortalities resemble that estimated2 for American Medicare enrollees. The inclusion of a wider array of personal characteristics has not improved prediction materially, as our AUC is essentially the same as that of the Medicare study—a result that compares reasonably well to what other studies have achieved6,7,13,14,15,16, particularly considering the relatively wide time horizon of prediction for our study. The very low proportion with high predicted mortality might be due to essential randomness in mortality, the accrual of health-impacting events after start of follow-up, or due to shortcomings of the data available. But while we absolutely might point to health indicators that were not available for the study there are indications10,17,18 that these may not improve mortality prediction that much.
The mass of treatment costs is concentrated at low predicted mortalities in a pattern resembling that of Einav et al.2. Care-related costs, conversely, are concentrated at higher mortalities and increase more markedly with increasing mortality, whereas the costs of treatment among decedents actually decrease up to a predicted mortality of about 30%. This is not surprising—predicted mortality is a proxy for frailty and thus for the need for communal care, and the need for care is likely to change less as the result of health-impacting events over the course of follow-up. It is interesting that we see a decline with predicted mortality in treatment-related expenditure per day alive for decedents. This was not observed for the American population and may reflect different medical culture in Denmark and the US, but the different prediction algorithms might also be part of the explanation—treatment-related expenditure decreases with age in Danish decedents11, and if a high predicted mortality is more reflective of age and frailty in our algorithm than for the American data, that might explain the difference.
At similar predicted mortalities, there is little difference between the care-related expenditure per day alive of decedents and survivors. The treatment-related expenditure of decedents is much higher than that of survivors, although the differences are lower at higher mortalities. This pattern may in part be explained by the passage of time—by the time a person dies, their health has likely deteriorated since their status at entry, and it seems likely that a person who dies at low predicted mortality will have experienced some dramatic health event requiring treatment, while death at higher predicted mortality might be a more direct continuation of patterns already established by the time of entry. Also, a person with low predicted mortality might be a better candidate for treatment, being less frail. But to the extent that the difference between survivors and decedents at the same mortality is not due to curveball events, it might be seen as the “true” cost of dying.
Thus, nearly all healthcare expenditure occurs in situations where there is a reasonable expectation that the patient can survive, and so the concept of “the cost of dying” is confounded by frailty: We spend more on the frail, and the frail are more likely to die—but not certain to do so, at least within a relevant time frame. This underlying frailty, operationalized as high predicted one-year mortality accounted for 39% healthcare expenditure in the last year of life in Denmark, an estimate in line with that in American Medicare enrollees2. The idea of a potential for reductions in health care expenditure at the end of life is enticing, and it seems possible to find groups that could benefit from a switch to a palliative course of treatment. Still, our results, along with those of our model paper, add to a list of arguments for why it might be illusory to reduce healthcare expenditure much by cutting the cost of dying. The proportion of spending occurring at the end of life is lower than has previously been reported1, decedents make up a relatively small share of high-cost individuals3, rising levels of demand drive increasing health care costs in ageing populations at least as much as the cost of dying19, and high end-of-life costs seem driven more by multimorbidity than last-ditch lifesaving efforts1,11,20. Our study design does not touch upon the question of individual treatment effects—whether specific treatments improve survival for specific individuals—and it may be that better methods than ours can detect high-mortality subgroups, but it seems unlikely for such subgroups to be large enough that costs reductions there could matter on the scale of a national budget.
Source link
