Predictive performance of four biopsy-free scores for screening myeloid malignancies among Thai cytopenic patients

Myeloid malignancies predominantly affect older adults and are commonly diagnosed through invasive bone marrow biopsies. Necessity of developing noninvasive diagnostic approaches is increasing because of aging global population and increasing availability of next-generation sequencing. This study evaluated four biopsy-free scoring models to reduce unnecessary biopsies among cytopenic patients. Four Gradient Boosting Models from frameworks by Oster et al. (Model I: demographic and hematological parameters), Træden et al. (Model II: adding gene mutation data), Zheng et al. (Model III: mutation count and maximum variant allele frequency[VAF]), and DeZern et al. (Model IV: VAF of 17 highest influential genes) were developed for 285 cytopenic patients (210 with myeloid malignancies, 75 without). Models were evaluated using area under the precision-recall curve (AUCPR) and F-scores from 100-fold Monte Carlo cross-validation. All models showed acceptable-to-good discrimination (AUCPR: 0.89–0.94). Models I and II demonstrated 100% negative predictive value (NPV) bands, encompassing an average 16.48% (95% Confidence interval [CI], 14.82–18.21%) and 19.01% (95% CI, 17.27–20.69%) of patients without myeloid malignancy in repeated validation samples, respectively. Patients in these bands had relatively preserved hematologic profiles, with absence of ASXL1, TET2, and SRSF2 mutations for Model II. Models III and IV did not generate 100% NPV bands. Models I and II demonstrated potential utility for identifying cytopenic patients at sufficiently low predicted risk in whom bone marrow biopsy deferral may warrant further evaluation. Model I, using only routinely available parameters, may be particularly suited for broad clinical application. External validation remains necessary before clinical implementation.