How different are accuracies of clinicians' prediction of survival by assessment methods?

Background: No standardized method has been established for evaluating the accuracy of a clinicians' prediction of survival (CPS). Till now, no study has compared the accuracy of CPS according to the evaluation methods using the same dataset. We aimed to examine the accuracy of CPS by different statistical approaches in patients with far-advanced cancer.

Methods: The current study was a secondary analysis of an international multicenter prospective cohort study. Newly admitted patients with advanced cancer were enrolled in palliative care units (PCUs) in Japan, Korea, and Taiwan. We obtained the temporal CPS at enrollment. The patients were classified into groups of days (≤7 days) and weeks (≤30 days) based on CPS and actual survival (AS). We evaluated the accuracy of CPS by the distribution, area under the receiver operating characteristics curve (AUROCs), and an estimate ±33% of AS.

Results: A total of 2,571 patients were assessed and admitted in 37 PCUs between January 2017 and September 2018. As for the "days" category, the distribution of AS is larger than that of CPS, however, the results are reversed in the "weeks" category. The AUROCs showed over 80% discrimination for both the "days" and "weeks" categories. Accurate CPS within ±33% of AS was approximately 30% in both "days" and "weeks" categories.

Conclusions: We showed a discrepancy of approximately 30-80% in the accuracy of CPS among three different analysis methods: distribution, AUROC, and AS comparison. Considering the low accuracy of AS comparisons, clinicians should provide a wide range of survival time. CPS was able to effectively discriminate and may be useful for risk stratification.