- 在已由海德堡分类器 v12.8 成功分类的 195 例「常规」CNS 肿瘤中，Bethesda v3（Bv3）正确分类 189/195（97%）。
- 在海德堡无法分类的 321 例疑难队列中，Bv3 可分类 180/321（56%）；得分 ≥0.9 但与最终 WHO2021 诊断不符者 13/321（4%），其中半数标为节细胞胶质瘤。
- 胶质母细胞瘤中 79% 被正确识别（含 DNA 浓度与肿瘤纯度较低者）；支持 Bv3 作为常规神经病理诊断的互补机器学习工具。
摘要
基于 DNA 甲基化的肿瘤分类已成为神经病理诊断的重要工具，但一部分中枢神经系统（CNS）肿瘤用现有方法仍无法分类。机器学习分类器在不同病例复杂程度下的表现尚未充分界定。我们在「常规」与诊断困难的 CNS 肿瘤标本中评估了 Bethesda Classifier v3（Bv3）。
我们首先将 Bv3 应用于 195 例已由海德堡分类器 v12.8（Hv12.8；n=195）成功分类的 CNS 肿瘤样本；随后在一份回顾性、Hv12.8 无法分类的诊断困难队列（n=321）中比较 Bv3 分类结果。两队列均按 WHO CNS5（2021）重新评估整合组织分子诊断（下文称最终 WHO2021 诊断）。误分类定义为：Bv3 得分 ≥0.9，但与最终 WHO2021 诊断不符的病例。
在常规队列中，Bv3 正确分类 189/195 例（97%）。在困难队列中，Bv3 对既往无法分类的病例分类出 180/321 例（56%）。Bv3 误分类 13/321 例（4%），其中 7/13（54%）被标为节细胞胶质瘤。胶质母细胞瘤中 79% 被正确识别，包括 DNA 浓度较低与肿瘤纯度较低的肿瘤。Bv3 在 3 例以得分 ≥0.9 指定为 HGAP（伴毛细胞样特征的高级别星形细胞瘤），但与最终 WHO2021 诊断不符。
Bethesda Classifier v3 在常规与诊断困难的 CNS 肿瘤标本中均表现稳健，可作为常规神经病理诊断中有价值的互补机器学习工具。
Abstract
DNA methylation-based tumor classification has become an essential tool in neuropathological diagnostics, yet a subset of central nervous system (CNS) tumors remains unclassifiable using current approaches. The performance of machine learning-based classifiers across varying case complexity is not fully defined. We evaluated the Bethesda Classifier v3 (Bv3) in both straightforward and diagnostically challenging CNS tumor specimens. We first applied Bv3 to 195 CNS tumor samples that were successfully classified using the Heidelberg Classifier v12.8 (Hv12.8; n = 195). We then compared Bv3 classifications in a retrospective cohort of diagnostically challenging, unclassifiable cases by Hv12.8 (n = 321). Both cohorts have been re-evaluated to identify the integrated histomolecular diagnoses according to the WHO CNS5 (2021) classification, hereafter referred to as the final WHO2021 diagnosis. Misclassification was defined as cases with Bv3 scores ≥ 0.9 that did not match the final WHO2021 diagnosis. In the straightforward cohort, Bv3 correctly classified 189/195 cases (97%). In the challenging cohort, Bv3 classified 180/321 previously unclassifiable cases (56%). 13/321 (4%) cases were misclassified by Bv3, of which 7/13 (54%) were labeled as ganglioglioma. Among glioblastomas, 79% were correctly identified, including tumors with lower DNA concentration and lower tumor purity. Bv3 assigned HGAP (High-Grade Astrocytoma with Piloid features) with a score ≥ 0.9 in three cases, which were not concordant with the final WHO2021 diagnosis. The Bethesda Classifier v3 demonstrates robust performance across both straightforward and diagnostically challenging CNS tumor specimens and represents a valuable complementary machine learning tool in routine neuropathological diagnostics.