Likelihood Ratios Offer Clinicians a Sharper Read on Prognostic Tools

Prognostication sits at the heart of clinical decision-making in palliative medicine, yet most clinicians report inadequate training in it and rarely use validated prognostic tools in practice. A new review in Supportive Care in Cancer argues that a critical barrier is limited understanding of statistical performance metrics, particularly likelihood ratios, which quantify how well a tool discriminates and how useful it is for an individual patient. The authors, Mellar P. Davis, Arunkumar Krishnan, Anthony Caprio and Spencer Strobel, set out to give clinicians a practical framework for critically assessing and interpreting published prognostic models.

The review’s objective was twofold: to explain the statistical methodology underlying prognostic tools, with emphasis on likelihood ratios, and to translate that methodology into a usable interpretive framework. Rather than simply cataloguing tools, the authors examined how each tool’s reported operating characteristics translate into probabilities that matter at the bedside, where a clinician must decide how much weight to give a score alongside clinical judgment.

Methodologically, the team went beyond narrative summary. They reviewed the primary publications contributing to their comparison table to verify, for each operating point, the study population, prediction horizon, model or threshold, outcome direction, sensitivity and specificity. The quantitative analysis included 27 model-threshold-horizon combinations with verifiable sensitivity and specificity. From these they calculated positive likelihood ratios, negative likelihood ratios, the Weissler negative likelihood ratio, diagnostic odds ratios and log10 transformations. Likelihood ratio isopleths were plotted in receiver operating characteristic space, a technique that makes explicit how discrimination varies across thresholds rather than collapsing it into a single summary value.

Not every study lent itself to this analysis. Investigations reporting only area under the curve, hazard ratios or survival by risk group were summarized descriptively rather than included in the ROC-space analyses, a distinction the authors maintain to avoid mixing incomparable metrics.

The results revealed substantial heterogeneity. Across the 27 operating points, likelihood ratios varied widely by tool, clinical setting, prediction horizon and threshold choice. Nine operating points met both a positive likelihood ratio of at least 2 and a negative likelihood ratio of at most 0.5; eight met the positive criterion alone, nine met the negative criterion alone, and one met neither. In other words, only a minority of reported operating points cleared both conventional benchmarks for clinically meaningful rule-in and rule-out value.

Which tools performed best depended on the question asked. Selected thresholds of the Palliative Performance Scale, the Palliative Prognostic Score and its D-PaP variant provided the strongest short-term mortality discrimination. By contrast, the Surprise Question and NECPAL showed high sensitivity but lower specificity, a profile consistent with screening rather than definitive mortality classification. That asymmetry matters: a highly sensitive tool is good at flagging patients who might be approaching the end of life, but it will also flag many who are not, so a positive screen requires follow-up rather than certainty.

The diagnostic odds ratio, which combines sensitivity and specificity into a single measure, told a similar story of uneven performance. It was strong for one operating point, good for six, moderate for nineteen and minimal for one. The authors characterize the overall pattern as heterogeneous and often asymmetric, with selected short-horizon PPI, PaP and D-PaP thresholds offering stronger mortality discrimination while the Surprise Question and NECPAL primarily support screening.

The review’s central conclusion is that likelihood ratios bridge population-level discrimination and patient-specific prognostic assessment, making them the metric clinicians should attend to when judging whether a published tool will actually help an individual patient. But the authors are careful about scope: prognostic tools should supplement rather than replace clinical judgment. Published as a review article in Supportive Care in Cancer, the work offers no new clinical recommendations beyond this interpretive guidance, and its quantitative comparisons are limited to the 27 operating points whose sensitivity and specificity could be independently verified from primary sources.

Subject of Research: Statistical evaluation of prognostic tools in palliative care using likelihood ratios

Article Title: Prognostication and likelihood ratios: methodology, discrimination, and interpretation – a clinician’s guide to prognostic tools

Article References: Davis, M. P., Krishnan, A., Caprio, A., & Strobel, S. (2026). Prognostication and likelihood ratios: methodology, discrimination, and interpretation – a clinician’s guide to prognostic tools. Supportive Care in Cancer, 34(11), Article 1142. https://doi.org/10.1007/s00520-026-11327-0

Image Credits: AI Generated

DOI: 10.1007/s00520-026-11327-0

Keywords: palliative care, prognostication, likelihood ratios, Palliative Performance Scale, Surprise Question, NECPAL, sensitivity, specificity, diagnostic odds ratio, ROC analysis, clinical decision-making, likelihood