A machine learning-based prediction model of H3K27M mutations in brainstem gliomas using conventional MRI and clinical features.

Abstract

H3K27M is the most frequent mutation in brainstem gliomas (BSGs), and it has great significance in the differential diagnosis, prognostic prediction and treatment strategy selection of BSGs. There has been a lack of reliable noninvasive methods capable of accurately predicting H3K27M mutations in BSGs. A total of 151 patients with newly diagnosed BSGs were included in this retrospective study. The H3K27M mutation status was obtained by whole-exome, whole-genome or Sanger's sequencing. A total of 1697 features, including 6 clinical parameters and 1691 imaging features, were extracted from pre- and post-contrast T1-weighted and T2-weighted images. Using a random forest algorithm, 36 selected MR image features were integrated with 3 selected clinical features to generate a model that was predictive of H3K27M mutations. Additionally, a simplified prediction model comprising the Karnofsky Performance Status (KPS) at diagnosis, symptom duration at diagnosis and edge sharpness on T2 was established for practical clinical utility using the least squares estimation method. H3K27M mutation was an independent prognostic factor that conferred a worse prognosis (p = 0.01, hazard ratio = 3.0, 95% confidence interval [CI], 1.57-5.74). The machine learning-based model achieved an accuracy of 84.44% (area under the curve [AUC] = 0.8298) in the test cohort. The simplified model achieved an AUC of 0.7839 in the test cohort. Using conventional MRI and clinical features, we established a machine learning-based model with high accuracy and a simplified model with improved clinical utility to predict H3K27M mutations in BSGs.