Longitudinal temporal and probabilistic prediction of survival in a cohort of patients with advanced cancer.

Context: Survival prognostication is important during the end of life. The accuracy of clinician prediction of survival (CPS) over time has not been well characterized.

Objectives: The aims of the study were to examine changes in prognostication accuracy during the last 14 days of life in a cohort of patients with advanced cancer admitted to two acute palliative care units and to compare the accuracy between the temporal and probabilistic approaches.

Methods: Physicians and nurses prognosticated survival daily for cancer patients in two hospitals until death/discharge using two prognostic approaches: temporal and probabilistic. We assessed accuracy for each method daily during the last 14 days of life comparing accuracy at Day -14 (baseline) with accuracy at each time point using a test of proportions.

Results: A total of 6718 temporal and 6621 probabilistic estimations were provided by physicians and nurses for 311 patients, respectively. Median (interquartile range) survival was 8 days (4-20 days). Temporal CPS had low accuracy (10%-40%) and did not change over time. In contrast, probabilistic CPS was significantly more accurate (P < .05 at each time point) but decreased close to death.

Conclusion: Probabilistic CPS was consistently more accurate than temporal CPS over the last 14 days of life; however, its accuracy decreased as patients approached death. Our findings suggest that better tools to predict impending death are necessary.