Background The Multinational Association for Supportive Care in Cancer (MASCC) Risk Index is the most widely used tool for stratifying febrile neutropenic cancer patients as low or high risk for serious complications. Studies evaluating its discriminative performance have relied predominantly on sensitivity, specificity, receiver operating characteristic (ROC) curves, and c-statistics—metrics that are mathematically elegant but have limited direct utility in clinical decision-making at the bedside. Methods Sensitivity and specificity data from Rivest et al. [<CitationRef CitationID="CR13">13</CitationRef>], the original Klastersky et al. derivation study [<CitationRef CitationID="CR3">3</CitationRef>], and the Zheng et al. meta-analysis [<CitationRef CitationID="CR10">10</CitationRef>] were used to calculate positive and negative likelihood ratios. Posttest probabilities were derived using Bayesian updating (pretest odds × LR = posttest odds), with a cohort pretest probability of complications of 35%. Likelihood ratios for procalcitonin (Ahn et al.) were applied sequentially to demonstrate the compounding of independent predictors. Results Applying Rivest et al. data in the inverted framing (i.e., predicting complication presence), a MASCC score > 21 yields a posttest probability of neutropenic complications of approximately 22%, and MASCC < 21 yields 65%. When procalcitonin ≥ 0.5 ng/mL is added to a high-risk MASCC score, the posttest probability rises to 85%; when procalcitonin is negative in a low-risk patient, it falls to 11%. Likelihood ratios varied meaningfully across Klastersky, Rivest, and Zheng (LR + range: 1.98–4.00; LR − range: 0.29–0.51), reflecting heterogeneity in populations, outcome definitions, and study settings. Conclusions Unlike sensitivity/specificity and predictive values, LRs are relatively independent of outcome prevalence, making them more transportable across clinical settings, though they are not immune to population and spectrum effects, stopping at sensitivity, specificity, ROC curves, and mathematical completeness but clinical incompleteness. Clinicians, who are the end-users of these tools, deserve the metric that speaks their language: the updated probability that this patient, in front of them, will experience harm.

MASCC score and neutropenic complications: what is the likelihood?

期刊论文
发表日期：

2026年5月16日

查看全文

Mellar P. Davis

摘要

Background

The Multinational Association for Supportive Care in Cancer (MASCC) Risk Index is the most widely used tool for stratifying febrile neutropenic cancer patients as low or high risk for serious complications. Studies evaluating its discriminative performance have relied predominantly on sensitivity, specificity, receiver operating characteristic (ROC) curves, and c-statistics—metrics that are mathematically elegant but have limited direct utility in clinical decision-making at the bedside.

Methods

Sensitivity and specificity data from Rivest et al. [13], the original Klastersky et al. derivation study [3], and the Zheng et al. meta-analysis [10] were used to calculate positive and negative likelihood ratios. Posttest probabilities were derived using Bayesian updating (pretest odds × LR = posttest odds), with a cohort pretest probability of complications of 35%. Likelihood ratios for procalcitonin (Ahn et al.) were applied sequentially to demonstrate the compounding of independent predictors.

Results

Applying Rivest et al. data in the inverted framing (i.e., predicting complication presence), a MASCC score > 21 yields a posttest probability of neutropenic complications of approximately 22%, and MASCC < 21 yields 65%. When procalcitonin ≥ 0.5 ng/mL is added to a high-risk MASCC score, the posttest probability rises to 85%; when procalcitonin is negative in a low-risk patient, it falls to 11%. Likelihood ratios varied meaningfully across Klastersky, Rivest, and Zheng (LR + range: 1.98–4.00; LR − range: 0.29–0.51), reflecting heterogeneity in populations, outcome definitions, and study settings.

Conclusions

Unlike sensitivity/specificity and predictive values, LRs are relatively independent of outcome prevalence, making them more transportable across clinical settings, though they are not immune to population and spectrum effects, stopping at sensitivity, specificity, ROC curves, and mathematical completeness but clinical incompleteness. Clinicians, who are the end-users of these tools, deserve the metric that speaks their language: the updated probability that this patient, in front of them, will experience harm.