基于机器学习算法的H3K27M突变型弥漫中线胶质瘤患者预后预测模型构建及其验证

doi:10.3969/j.issn.1004-583X.2025.11.004

摘要/Abstract

摘要：

目的探究H3K27M突变型弥漫中线胶质瘤(diffuse midline glioma, DMG)生存预后的影响因素,构建并验证H3K27M突变型DMG预后不良的预测模型。方法回顾性分析2000-2019年监测、流行病学及预后数据库中H3K27M突变型DMG患者的临床资料,按照7:3的比例随机分为训练集(n=97)和验证集(n=41)。采用极端梯度提升、随机森林、最小绝对收缩和选择算子回归分析和决策树模型筛选变量,分析4种机器学习方法“重叠覆盖”的风险因素,采用多因素Cox回归方法验证结果的独立预测性,并在此基础上构建列线图预测H3K27M突变型DMG患者6、12、18个月生存率。使用受试者工作特征曲线下面积、校准曲线和临床决策曲线评估列线图模型的预测效能、准确性和临床适用性,Kaplan-Meier法绘制预后影响变量的生存曲线。结果 4种机器学习算法各自筛选出不同预后影响因素,经过“重叠覆盖分析”得到5种共同的H3K27M型DMG患者预后影响因素:年龄、肿瘤体积、世界卫生组织分级、肿瘤侧别和放疗。多因素Cox回归分析发现,>60岁(HR=3.018, 95%CI:1.15~7.92,P=0.025),肿瘤体积增大(HR=1.039, 95%CI:1.01~1.06,P=0.004),高级别胶质瘤(HR=2.057, 95%CI:1.21~3.49,P=0.008),中线结构(HR=2.101, 95%CI:1.32~3.34,P=0.002)是H3K27M突变型DMG预后的危险性因素,放疗(HR=0.410, 95%CI:0.23~0.75,P=0.004)是H3K27M突变型DMG的保护性因素。基于5个预后影响因素构建的列线图模型验证结果显示,训练集和验证集6、12、18个月的曲线下面积分别为0.647、0.746、0.625和0.632、0.725、0.725,表明模型预测性能良好,校准曲线显示预测值与理想曲线具有较好一致性,临床决策曲线分析曲线显示列线图模型有良好的临床效能。结论基于机器学习算法筛选出的预后影响因素构建的列线图模型有可靠的预测效能,可帮助临床医生识别H3K27M突变型DMG患者高危预后因素,制定个体化治疗方案。

关键词: H3K27M, 机器学习, 弥漫中线胶质瘤, 列线图, 预后因素

Abstract:

Objective To identify prognostic factors for H3K27 mmutant diffuse midline glioma (DMG) and to develop and validate a nomogram for predicting poor prognosis. Methods Patients with histologically confirmed H3K27Mmutant DMG recorded in the Surveillance Epidemiology and End Results database from 2000 to 2019 were retrospectively included. Cases were randomly split into a training set (n=97) and a validation set (n=41) at a 7:3 ratio. Candidate variables were screened by four machinelearning approaches-extreme gradient boosting random forest, least absolute shrinkage and selection operator regression, and decision tree. Multivariate Cox regression models were then used to develop predictive models. Risk factors identified by multiple methods were further tested by multivariate Cox regression to confirm independent prognostic value. A nomogram to predict 6, 12, and 18month overall survival was constructed from the independently significant predictors. Model discrimination, calibration, and clinical utility were evaluated using area under the receiver operating characteristic curve (AUC), calibration plots, and decision curve analysis (DCA). Survival differences were assessed by Kaplan-Meier analysis. Results The four machinelearning methods produced overlapping but not identical sets of candidate predictors; five variables overlapped across methods: Age, tumor size, WHO grade, laterality, and radiation. Multivariate Cox regression identified four independent adverse prognostic factors: Age >60 years (HR=3.018; 95%CI: 1.15-7.92; P=0.025), larger tumor size (per unit increase) (HR=1.039; 95%CI: 1.01-1.06; P=0.004), higher WHO grade (HR=2.057; 95%CI: 1.21-3.49; P=0.008), and midline location (HR=2.101; 95%CI: 1.32-3.34; P=0.002). Radiotherapy was independently associated with improved survival (protective effect: HR=0.410; 95%CI: 0.23-0.75; P=0.004). The nomogram incorporating these factors demonstrated AUCs for 6, 12, and 18 month OS of 0.647, 0.746, and 0.625 in the training set and 0.632, 0.725, and 0.725 in the validation set, respectively, indicating acceptable discrimination. Calibration plots showed good agreement between predicted and observed survival probabilities, and DCA indicated favorable clinical utility. Conclusion A nomogram developed from machine learning-selected predictors reliably estimates shortterm OS in patients with H3K27 mmutant DMG. This model may help clinicians identify highrisk patients and tailor individualized treatment strategies.

Key words: H3K27M, machine learning, diffuse midline glioma (DMG), nomogram, prognostic factors

中图分类号:

R739.4

王壮壮, 杨庆军, 任欢, 刘彦廷, 田春雷. 基于机器学习算法的H3K27M突变型弥漫中线胶质瘤患者预后预测模型构建及其验证[J]. 临床荟萃, 2025, 40(11): 988-998.

Wang Zhuangzhuang, Yang Qingjun, Ren Huan, Liu Yanting, Tian Chunlei. Development and validation of a machine learning-based prognostic model for H3K27 mmutant diffuse midline glioma[J]. Clinical Focus, 2025, 40(11): 988-998.

导出引用管理器 EndNote|Ris|BibTeX

链接本文: https://www.lchc.cn/CN/10.3969/j.issn.1004-583X.2025.11.004

https://www.lchc.cn/CN/Y2025/V40/I11/988

图/表 10

表1 两组基线特征比较

Tab.1 Comparison of baseline characteristics between the two groups

项目	训练集( $n$ =97)	验证集( $n$ =41)	χ²/ $Z$ 值	$P$ 值
年龄(岁)
<30	58(59.8)	24(58.5)
30~60	33(34.0)	12(29.3)	0.980	0.613
>60	6(6.2)	5(12.2)
性别[例(%)]
女性男性	52(53.6) 45(46.4)	19(46.3) 22(53.7)	0.620	0.431
种族[例(%)]
黑种人	7(7.2)	3(7.3)
白种人	83(85.6)	33(80.5)	0.720	0.698
黄种人	7(7.2)	5(12.2)
肿瘤侧别[例(%)]
左侧	8(8.2)	5(12.2)
右侧	12(12.4)	8(19.5)	1.850	0.396
中线结构	77(79.4)	28(68.3)
肿瘤分期[例(%)]
局部浸润	74(76.30)	28(68.30)
临近侵袭	12(12.40)	5(12.20)	1.930	0.381
远处转移	11(11.30)	8(19.50)
放疗[例(%)]
否是	16(16.5) 81(83.5)	8(19.5) 33(80.5)	0.180	0.674
化疗[例(%)]
否是	43(44.3) 54(55.7)	16(39.0) 25(61.0)	0.350	0.554
WHO分级[例(%)]
低(Ⅰ-Ⅱ) 高(Ⅲ-Ⅳ)	73(75.3) 24(24.7)	28(68.3) 13(31.7)	0.750	0.387
肿瘤数量[例(%)]
单发多发	1(1.0) 96(99.0)	0 41(100.0)	-	1.000
肿瘤体积(cm³)	41(33, 48)	42(30, 50)	-0.120	0.903

表1 两组基线特征比较

Tab.1 Comparison of baseline characteristics between the two groups

项目	训练集( $n$ =97)	验证集( $n$ =41)	χ²/ $Z$ 值	$P$ 值
年龄(岁)
<30	58(59.8)	24(58.5)
30~60	33(34.0)	12(29.3)	0.980	0.613
>60	6(6.2)	5(12.2)
性别[例(%)]
女性男性	52(53.6) 45(46.4)	19(46.3) 22(53.7)	0.620	0.431
种族[例(%)]
黑种人	7(7.2)	3(7.3)
白种人	83(85.6)	33(80.5)	0.720	0.698
黄种人	7(7.2)	5(12.2)
肿瘤侧别[例(%)]
左侧	8(8.2)	5(12.2)
右侧	12(12.4)	8(19.5)	1.850	0.396
中线结构	77(79.4)	28(68.3)
肿瘤分期[例(%)]
局部浸润	74(76.30)	28(68.30)
临近侵袭	12(12.40)	5(12.20)	1.930	0.381
远处转移	11(11.30)	8(19.50)
放疗[例(%)]
否是	16(16.5) 81(83.5)	8(19.5) 33(80.5)	0.180	0.674
化疗[例(%)]
否是	43(44.3) 54(55.7)	16(39.0) 25(61.0)	0.350	0.554
WHO分级[例(%)]
低(Ⅰ-Ⅱ) 高(Ⅲ-Ⅳ)	73(75.3) 24(24.7)	28(68.3) 13(31.7)	0.750	0.387
肿瘤数量[例(%)]
单发多发	1(1.0) 96(99.0)	0 41(100.0)	-	1.000
肿瘤体积(cm³)	41(33, 48)	42(30, 50)	-0.120	0.903

图1 变量相关热图

Fig. 1 Variable correlation heat map

图2 4种机器学习算法筛选H3K27M突变型DMG患者预后相关变量 a.XGBoost模型识别出来的预后影响变量的重要性得分;b.RF模型识别出来的预后影响变量的重要性得分;c.Lasso回归模型识别出来的预后影响变量的重要性得分;d.DT模型识别出来的预后影响变量的重要性得分

Fig.2 Prognostic variables for H3K27M mutant DMG patients screened using four machine learning algorithms a. The importance score of prognostic variables identified by XGBoost model; b. The importance score of the prognostic variables identified by the RF model; c. The importance score of the prognostic impact variables identified by the Lasso regression model; d. The importance score of prognostic variables identified by the DT model

表2 不同算法模型在训练集和验证集的准确性评估

Tab.2 Accuracy evaluation of different algorithm models in training set and validation set

模型	分类	AUC(95% $C I$ )	Brier值	灵敏度	特异度	阳性预测值	阴性预测值	平衡准确性
训练集	Lasso	0.82(0.76~0.87)	0.150	0.780	0.840	0.810	0.820	0.810
	XGBoost	0.91(0.87~0.94)	0.110	0.870	0.820	0.840	0.860	0.850
	RF	0.95(0.92~0.97)	0.080	0.910	0.880	0.890	0.900	0.900
	DT	0.88(0.83~0.92)	0.130	0.830	0.800	0.820	0.810	0.820
验证集	Lasso	0.76(0.68~0.83)	0.180	0.720	0.790	0.750	0.760	0.760
	XGBoost	0.80(0.73~0.86)	0.160	0.760	0.810	0.780	0.790	0.790
	RF	0.78(0.70~0.85)	0.170	0.740	0.770	0.760	0.750	0.760
	DT	0.72(0.64~0.79)	0.190	0.680	0.740	0.710	0.710	0.710

表2 不同算法模型在训练集和验证集的准确性评估

Tab.2 Accuracy evaluation of different algorithm models in training set and validation set

模型	分类	AUC(95% $C I$ )	Brier值	灵敏度	特异度	阳性预测值	阴性预测值	平衡准确性
训练集	Lasso	0.82(0.76~0.87)	0.150	0.780	0.840	0.810	0.820	0.810
	XGBoost	0.91(0.87~0.94)	0.110	0.870	0.820	0.840	0.860	0.850
	RF	0.95(0.92~0.97)	0.080	0.910	0.880	0.890	0.900	0.900
	DT	0.88(0.83~0.92)	0.130	0.830	0.800	0.820	0.810	0.820
验证集	Lasso	0.76(0.68~0.83)	0.180	0.720	0.790	0.750	0.760	0.760
	XGBoost	0.80(0.73~0.86)	0.160	0.760	0.810	0.780	0.790	0.790
	RF	0.78(0.70~0.85)	0.170	0.740	0.770	0.760	0.750	0.760
	DT	0.72(0.64~0.79)	0.190	0.680	0.740	0.710	0.710	0.710

表3 H3K27M突变型DMG患者预后的多因素Cox回归分析

Tab.3 Multivariate Cox regression analysis of prognosis in patients with H3K27M mutant DMG

项目	$H R$	95% $C I$	$P$ 值
年龄
30~60 vs <30	0.630	0.65~2.28	0.382
>60 vs <30	3.018	1.15~7.91	0.025
肿瘤体积	1.039	1.01~1.06	0.004
WHO分级	2.057	1.21~3.49	0.008
肿瘤侧别
右侧vs左侧	0.739	0.61~2.42	0.739
中线vs左侧	2.101	1.32~3.34	0.002
放疗	0.410	0.23~0.75	0.004

表3 H3K27M突变型DMG患者预后的多因素Cox回归分析

Tab.3 Multivariate Cox regression analysis of prognosis in patients with H3K27M mutant DMG

项目	$H R$	95% $C I$	$P$ 值
年龄
30~60 vs <30	0.630	0.65~2.28	0.382
>60 vs <30	3.018	1.15~7.91	0.025
肿瘤体积	1.039	1.01~1.06	0.004
WHO分级	2.057	1.21~3.49	0.008
肿瘤侧别
右侧vs左侧	0.739	0.61~2.42	0.739
中线vs左侧	2.101	1.32~3.34	0.002
放疗	0.410	0.23~0.75	0.004

图3 H3K27M突变型DMG患者预后的列线图预测模型

Fig.3 Prognostic nomogram prediction model of H3K27M mutant DMG patients

图4 训练集和验证集中6、12、18个月总生存率的ROC曲线 a.训练集;b.验证集

Fig.4 ROC curve of 6-, 12-, and 18-month OS in training set and validation set a. Training set; b. Validation set

图5 训练集和验证集中6、12、18个月总生存率的Calibration校准曲线 a.训练集6个月总生存率的Calibration校准曲线;b.训练集12个月总生存率的Calibration校准曲线;c.训练集18个月总生存率的Calibration校准曲线;d.验证集6个月总生存率的Calibration校准曲线;e.验证集12个月总生存率的Calibration校准曲线;f.验证集18个月总生存率的Calibration校准曲线

Fig.5 Calibration calibration curve of 6-, 12-, and 18-month OS in training set and validation set a. Calibration calibration curve of 6-month OS in training set; b. Calibration calibration curve of 12-month OS in training set; c. Calibration calibration curve of the 18-month OS of the training set; d. Calibration calibration curve of 6-month OS in validation set; e. Calibration calibration curve of 12-month OS in validation set; f. Calibration calibration curve for 18-month OS of the validation set

图6 训练集和验证集6、12、18个月总生存率的DCA校准曲线 a.训练集;b.验证集

Fig.6 DCA calibration curve of 6-, 12-and 18-month OS of training set and validation set a. Training set; b. Validation set

图7 H3K27M突变型DMG预后影响因素的Kaplan-Meier生存分析曲线 a.年龄;b.肿瘤侧别;c.放疗;d.WHO分级;e.肿瘤体积

Fig. 7 Kaplan-Meier survival analysis curves of prognostic factors in H3K27M mutant DMG a. Age; b. Tumor laterality; c. Radiotherapy; d. WHO grade; e. Tumor size

参考文献 31

[1]	胡鹏, 毛佳颖, 赵淦, 等. MAGED2、EMP2表达与H3K27M突变型弥漫性中线胶质瘤预后的关系研究[J]. 中华神经外科疾病研究杂志, 2025, 19(3): 52-57. doi:10.3760/cma.j.issn.0529-5807.2019.03.005.
[2]	杨沁怡, 李明娜, 陈天羽, 等. 成人H3K27M变异型弥漫性中线胶质瘤临床病理学分析[J]. 中华病理学杂志, 2023, 52(4): 376-383.doi:10.3760/cma.j.cn112151-20220926-00810.
[3]	杜娟, 李永晴, 崔文丽, 等. 弥漫性中线胶质瘤伴H3K27变异型的临床病理学及分子遗传学特征[J]. 临床与实验病理学杂志, 2025, 41(9): 1187-1193.doi:10.13315/j.cnki.cjcep.2025.09.011.
[4]	Sarnow K, Majercak E, Qurbonov Q, et al. Neuroimmune-competent human brain organoid model of diffuse midline glioma[J]. Neuro Oncol, 2025, 27(2): 369-382.doi: 10.1093/neuonc/noae245.
[5]	Van den BM, Saratsis AM, Geurts M, et al. H3 K27M-altered glioma and diffuse intrinsic pontine glioma: Semi-systematic review of treatment landscape and future directions[J]. Neuro Oncol, 2024, 26(2): S110-S124.doi: 10.1093/neuonc/noad220.
[6]	Daoud EV, Rajaram V, Cai C, et al. Adult brainstem gliomas with H3K27M mutation: Radiology, pathology, and prognosis[J]. J Neuropathol Exp Neurol, 2018, 77(4): 302-311.doi: 10.1093/jnen/nly006.
[7]	蒋会, 苏晓然, 吴玲, 等. 基于全域ADC直方图评估儿童高级别弥漫性中线胶质瘤H3K27M突变[J]. 临床放射学杂志, 2024, 43(2): 251-256.doi:10.13437/j.cnki.jcr.2024.02.020.
[8]	李海南, 李智, 赖名耀, 等. H3K27M突变型中线胶质瘤的病理学特征及其预后影响因素分析[J]. 中华神经外科杂志, 2022, 38(10): 1001-1006.doi:10.3760/cma.j.cn112050-20220226-00115.
[9]	Cheng L, Zhou M, Luo T, et al. H3F3B p.K27I-mutant diffuse midline glioma is a distinct subtype of H3K27-altered diffuse midline glioma[J]. Acta Neuropathol Commun, 2025, 13(1):183. doi: 10.1186/s40478-025-02101-0.
[10]	Li X, Sun W, Guo Z, et al. Identification of urinary metabolic biomarkers for histone 3 lysine27-to-methionine mutation diagnosis in brainstem gliomas[J]. Neuro Oncol, 2025, 27(6): 1536-1549.doi: 10.1093/neuonc/noaf038.
[11]	Lo Greco MC, Marano G, La RM, et al. Latest Advancements in the management of H3K27M-mutant diffuse intrinsic pontine glioma: A narrative review[J]. Cancers (Basel), 2025, 17(3): 420.doi: 10.3390/cancers17030420.
[12]	Sheikh SR, Recinos VMR, Thompson EM, et al. The role of brainstem biopsy and targeted therapies in pediatric diffuse midline glioma/diffuse intrinsic pontine glioma[J]. Front Oncol, 2024, 14((10): 1504440.doi: 10.3389/fonc.2024.1504440.
[13]	Angelico G, Mazzucchelli M, Attanasio G, et al. H3K27me3 loss in central nervous system tumors: Diagnostic, prognostic, and therapeutic implications[J]. Cancers (Basel), 2024, 16(20): 3451. doi: 10.3390/cancers16203451.
[14]	Chaturvedi A, Sadashiva N, Shukla D, et al. Thalamic H3K27M altered diffuse midline gliomas: Clinicopathological and outcome analysis[J]. Clin Neurol Neurosurg, 2024, 244(12):108449.doi: 10.1016/j.clineuro.2024.108449.
[15]	潘昱辰, 黄国栋, 张协军, 等. H3K27M突变型弥漫性中线胶质瘤的诊治分析[J]. 中国临床神经外科杂志, 2021, 26(10):761-763.doi:10.13798/j.issn.1009-153X.2021.10.006.
[16]	周江芬, 赖名耀, 邓官华, 等. 成人脑干胶质瘤临床特点及预后:单中心10年81例患者回顾性分析[J]. 临床神经外科杂志, 2022, 19(1):70-73.
[17]	梁博, 梁庭毓, 王芳, 等. 弥漫性中线胶质瘤伴H3K27M突变的临床诊疗及预后分析[J]. 中国微侵袭神经外科杂志, 2019, 24(7): 299-302.
[18]	Nonnenbroich LF, Bouchal SM, Millesi E, et al. H3K27-altered diffuse midline glioma of the brainstem: From molecular mechanisms to targeted interventions[J]. Cells, 2024, 13(13):1122. doi: 10.3390/cells13131122.
[19]	Liu C, Kuang S, Huang T, et al. Radiotherapy plus temozolomide with or without anlotinib in H3K27M-mutant diffuse midline glioma: A retrospective cohort study[J]. CNS Neurosci Ther, 2024, 30(4): e14730. doi: 10.1111/cns.14730.
[20]	Zang D, Dong Z, Liu Y, et al. Single-cell RNA sequencing of anaplastic ependymoma and H3K27M-mutant diffuse midline glioma[J]. BMC Neurol, 2024, 24(1):74.doi: 10.1186/s12883-024-03558-7. pmid: 38383423
[21]	Yang Z, Sun L, Chen H, et al. New progress in the treatment of diffuse midline glioma with H3K27M alteration[J]. Heliyon, 2024, 10(2): e24877.doi: 10.1016/j.heliyon.2024.e24877.
[22]	Huang B, Chen T, Zhang Y, et al. Deep learning for the prediction of the survival of midline diffuse glioma with an H3K27M alteration[J]. Brain Sci, 2023, 13(10):1483.doi: 10.3390/brainsci13101483.
[23]	Adhikari S, Bhutada AS, Ladner L, et al. Prognostic indicators for H3K27M-mutant diffuse midline glioma: A population-based retrospective surveillance, epidemiology, and end results database analysis[J]. World Neurosurg, 2023, 178(14): e113-e121. doi: 10.1016/j.wneu.2023.07.001.
[24]	罗敏捷, 陈嘉志, 张旺明. 儿童H3 K27M突变型弥漫性中线胶质瘤治疗进展[J]. 中国现代神经疾病杂志, 2020, 20(4):270-276. doi: 10.3969/j.issn.1672-6731.2020.04.005
[25]	王佳, 许朗宁, 葛明, 等. 伴H3 K27M突变的儿童弥漫性中线胶质瘤的临床特点、治疗及预后[J]. 中华神经外科杂志, 2021, 37(2):118-122.doi:10.3760/cma.j.cn112050-20200717-00408.
[26]	Gong X, Kuang S, Deng D, et al. Differences in survival prognosticators between children and adults with H3K27M-mutant diffuse midline glioma[J]. CNS Neurosci Ther, 2023, 29(12): 3863-3875.doi: 10.1111/cns.14307. pmid: 37311690
[27]	Xu K, Sun Z, Wang L, et al. Bilateral thalamic gliomas harboring alterations of EGFR and H3K27M: An integrated clinicopathological characteristics case series[J]. World Neurosurg, 2022, 168(21): e442-e450. doi: 10.1016/j.wneu.2022.09.118.
[28]	Bin-Alamer O, Jimenez AE, Azad TD, et al. H3K27M-altered diffuse midline gliomas among adult patients: A systematic review of clinical features and survival analysis[J]. World Neurosurg, 2022, 165(14): e251-e264.doi: 10.1016/j.wneu.2022.06.020.
[29]	Vuong HG, Ngo TNM, Le HT, et al. The prognostic significance of HIST1H3B/C and H3F3A K27M mutations in diffuse midline gliomas is influenced by patient age[J]. J Neurooncol, 2022, 158(3): 405-412.doi: 10.1007/s11060-022-04027-2.
[30]	Grimaldi S, Harlay V, Appay R, et al. Adult H3K27M mutated thalamic glioma patients display a better prognosis than unmutated patients[J]. J Neurooncol, 2022, 156(3): 615-623.doi: 10.1007/s11060-022-03943-7.
[31]	Gu Q, Huang Y, Zhang H, et al. Case report: Five adult cases of H3K27-altered diffuse midline glioma in the spinal cord[J]. Front Oncol, 2021, 11(10): 701113.doi: 10.3389/fonc.2021.701113.