Bayesian diagnostic test accuracy meta-analysis of ¹⁸F-FDG PET/CT for detecting Richter's transformation in chronic lymphocytic leukemia

Abstract

Abstract Introduction: Richter's transformation (RT) occurs in approximately 8–10% of patients with chronic lymphocytic leukemia (CLL) and is associated with a poor prognosis, with median overall survival of only 4.7 months following diagnosis. ¹⁸F-fluorodeoxyglucose positron emission tomography/computed tomography (¹⁸F-FDG PET/CT) has shown potential in identifying RT by detecting regions of elevated glycolytic activity consistent with high-grade clonal proliferation. However, the diagnostic accuracy of this imaging modality remains uncertain. We conducted a systematic review and Bayesian diagnostic test accuracy meta-analysis to evaluate the performance of ¹⁸F-FDG PET/CT in detecting RT among patients with CLL. Methodology: This PRISMA-DTA–adherent meta-analysis included studies evaluating ¹⁸F-FDG PET for detecting Richter's transformation in CLL. A systematic search was conducted across PubMed, Embase, Scopus, Web of Science, and Cochrane Library. Eligible studies reported sufficient data to construct 2×2 tables (true positives [TP], false positives [FP], true negatives [TN], false negatives [FN]). Bayesian hierarchical models were fitted in R (v4.5.1) via R-Studio. Logit-transformed sensitivities and specificities were assigned normal priors (mean = 0, SD = 1). Between-study heterogeneity was modeled using truncated normal priors (mean = 0, SD = 1), and an LKJ(2) prior was used for the correlation between sensitivity and specificity. Results: Bayesian hierarchical models estimated a pooled sensitivity of 0.842 (95% CrI: 0.756–0.905) and specificity of 0.706 (0.547–0.829) for ¹⁸F-FDG PET in detecting Richter's transformation in CLL. Models were fitted using weakly informative priors: logit-sensitivity (median = 0.496; 95% prior interval: 0.048–0.949) and logit-specificity (median = 0.500; 95% prior interval: 0.048–0.950). Sensitivity was highest in the low SUV max tertile (≤5) at 0.883 (95% CrI: 0.814–0.933), but specificity was lowest at 0.469 (0.293–0.675). In contrast, the medium tertile (>5 to ≤10) showed a balanced diagnostic profile with sensitivity 0.761 and specificity 0.817, yielding the highest diagnostic odds ratio (DOR: 14.0; 5.9–34.3). The high SUV max tertile (>10) demonstrated lower sensitivity (0.710) and moderate specificity (0.684). The diagnostic odds ratio was 12.82 (5.25–34.12), with a positive likelihood ratio of 2.86 (1.82–4.97) and a negative likelihood ratio of 0.23 (0.13–0.38). Between-study heterogeneity was moderate: SD for logit-sensitivity = 0.794 (0.330–1.425), SD for logit-specificity = 1.469 (0.991–2.172), and between-study correlation ρ = 0.069 (–0.515 to 0.600). In a hypothetical cohort of 1,000 patients with 8-10% prevalence, this translates to 77 true positives (69–82), 14 false negatives (9–12), 642 true negatives (497–754), and 267 false positives (155–412). All models demonstrated satisfactory convergence (R-hat < 1.01). Conclusion: ¹⁸F-FDG PET demonstrates high sensitivity and moderate specificity for RT detection in CLL, with optimal performance observed in the medium SUV max range (5–10). This supports its role as a reliable, noninvasive diagnostic adjunct in clinical decision-making, particularly for excluding transformation in standard-prevalence settings.