Prognostication and likelihood ratios: methodology, discrimination, and interpretation - a clinician's guide to prognostic tools.

backgroundPrognostication is fundamental to clinical decision-making in palliative medicine, yet most clinicians report inadequate training and rarely use validated prognostic tools in practice. A critical barrier is limited understanding of statistical performance metrics, particularly likelihood ratios (LRs), which quantify a tool's discriminative ability and clinical utility for individual patients.

objectiveTo review the statistical methodology underlying prognostic tools, with emphasis on likelihood ratios, and to provide clinicians with a framework for critically assessing and interpreting published prognostic models.

methodsThe primary publications contributing to Table 1 were reviewed to verify the study population, prediction horizon, model or threshold, outcome direction, sensitivity, and specificity for each operating point. The quantitative analysis included 27 model-threshold-horizon combinations with verifiable sensitivity and specificity. Positive likelihood ratio (PLR), negative likelihood ratio (NLR), Weissler negative likelihood ratio (wNLR), diagnostic odds ratio (DOR), and log10 transformations were calculated. Likelihood ratio isopleths were plotted in receiver operating characteristic (ROC) space. Studies reporting only area under the curve (AUC), hazard ratios, or risk-group survival were summarized descriptively rather than included in the ROC space analyses.

resultsThe 27 operating points showed substantial variation in likelihood ratios across tools, settings, prediction horizons, and thresholds. Nine operating points met both PLR ≥  2 and NLR ≤ 0.5, eight met the PLR criterion alone, nine met the NLR criterion alone, and one met neither criterion. Selected PPI, PaP, and D-PaP thresholds provided the strongest short-term mortality discrimination. The Surprise Question and NECPAL showed high sensitivity but lower specificity, consistent with screening rather than definitive mortality classification. DOR was strong for 1 operating point, good for 6, moderate for 19, and minimal for 1.

conclusionsLikelihood ratios bridge population-level discrimination and patient-specific prognostic assessment. Prognostic performance was heterogeneous and often asymmetric: selected short-horizon PPI, PaP, and D-PaP thresholds provided stronger mortality discrimination, whereas the Surprise Question and NECPAL primarily supported screening. Prognostic tools should supplement rather than replace clinical judgment.