Mohammadzadeh Ibrahim, Niroomand Behnaz, Hajikarimloo Bardia, Habibi Mohammad Amin, Mortezaei Ali, Behjati Jina, Albakr Abdulrahman, Borghei-Razavi Hamid
Skull Base Research Center, Loghman-Hakim Hospital, Shahid Beheshti University of Medical Sciences, Tehran, Iran.
Department of Neurological Surgery, University of Virginia, Charlottesville, VA, USA.
Clin Neurol Neurosurg. 2025 Feb;249:108762. doi: 10.1016/j.clineuro.2025.108762. Epub 2025 Jan 25.
Early prediction of recurrence in high-grade glioma (HGG) is critical due to its aggressive nature and poor prognosis. Distinguishing true recurrence from treatment-related changes, such as radionecrosis, is a major diagnostic challenge. Machine learning (ML) offers a novel approach, leveraging advanced algorithms to analyze complex imaging data with high precision. A comprehensive search of PubMed, Embase, Scopus, Web of Science, and Google Scholar identified eligible studies. The sensitivity, specificity, accuracy, precision, F1 score, and the (area under the curve) AUC items were extracted from the included studies. After screening 1077 records, seven studies met the eligibility criteria for the systematic review, of which five were included in the meta-analysis. ML algorithm, particularly Support Vector Machines (SVM), demonstrated promising performance. A meta-analysis of five studies revealed a pooled sensitivity of 0.95 (95% CI: 0.84-0.99) and specificity of 0.80 (95% CI: 0.69-0.88). Additionally, the positive diagnostic likelihood ratio (DLR) was 4.75 (95% CI: 2.91-7.76), the negative DLR was 0.06 (95% CI: 0.02-0.21), and the diagnostic odds ratio was 80.97 (95% CI: 17.5-374.61). The diagnostic score was calculated as 4.39 (95% CI: 2.86-5.93), and the AUC was 0.86 (95% CI: 0.83-0.89). Subgroup analyses showed SVM-based models with higher sensitivity (0.98 vs. 0.87) and specificity (0.82 vs. 0.77) than non-SVM (p = 0.13). Comparing glioblastoma and Grade 3 tumors, sensitivities were 94 % vs. 97 %, and specificities were 79 % vs. 83 %, with no significant heterogeneity. These findings suggest that ML models, particularly SVM, offer promising diagnostic performance in distinguishing true tumor recurrence from treatment-related changes.
由于高级别胶质瘤（HGG）具有侵袭性且预后较差，因此对其复发进行早期预测至关重要。将真正的复发与治疗相关变化（如放射性坏死）区分开来是一项重大的诊断挑战。机器学习（ML）提供了一种新方法，利用先进算法高精度分析复杂的成像数据。通过对PubMed、Embase、Scopus、Web of Science和谷歌学术进行全面检索，确定了符合条件的研究。从纳入的研究中提取敏感性、特异性、准确性、精确性、F1分数和曲线下面积（AUC）等指标。在筛选了1077条记录后，有7项研究符合系统评价的纳入标准，其中5项纳入荟萃分析。ML算法，特别是支持向量机（SVM），表现出了良好的性能。对5项研究的荟萃分析显示，合并敏感性为0.95（95%CI：0.84 - 0.99），特异性为0.80（95%CI：0.69 - 0.88）。此外，阳性诊断似然比（DLR）为4.75（95%CI：2.91 - 7.76），阴性DLR为0.06（95%CI：0.02 - 0.21），诊断比值比为80.97（95%CI：17.5 - 374.61）。诊断分数计算为4.39（95%CI：2.86 - 5.93），AUC为0.86（95%CI：0.83 - 0.89）。亚组分析显示，基于SVM的模型比非SVM模型具有更高的敏感性（0.98对0.87）和特异性（0.82对0.77）（p = 0.13）。比较胶质母细胞瘤和3级肿瘤，敏感性分别为94%对97%，特异性分别为79%对83%，无显著异质性。这些发现表明，ML模型，特别是SVM，在区分真正的肿瘤复发与治疗相关变化方面具有良好的诊断性能。
