Machine Learning Analysis of Local Recurrence of Meningioma Treated with Stereotactic Radiotherapy

Purpose/Objectives: Modeling local recurrence for patients with meningiomas treated with radiotherapy remains a clinical challenge. Prognostic factors such as the extent of resection, grade, Ki-67, and age have been derived from limited institutional series with few integrative models available that incorporate imaging features. Machine learning approaches represent a potentially useful method for non-linear outcome prediction based on an ability to integrate clinical data derived from patient history and quantitative metrics (radiomics) derived from pixel-level imaging. We hypothesized that a machine learning model derived from clinical and quantitative imaging predictors provide an effective prediction of local recurrence.

Methods: We analyzed patients with WHO grade I–III meningioma treated with stereotactic radiosurgery (SRS) or fractionated stereotactic radiotherapy (FSRT) between 2012 and 2018, with a median follow-up of 3.9 years. The primary outcome of interest was local recurrence following radiotherapy. Clinical variables, including age, tumor grade, Ki-67, previous resection, tumor location, and radiation planning metrics, were abstracted from the clinical charts. A total of 851 3D radiomic features were extracted from T1 post-contrast MRI images of each meningioma and used as predictors of recurrence. Support vector machine, Naïve Bayes, tree-based, ensemble-based, and traditional logistic regression classification schemes were tested. Fourfold cross validation was employed, with misclassification cost weighted to minimize the rate of false negatives. The discriminatory ability of each model was assessed using receiver operating characteristic curves, with differing performance in models estimated using the DeLong method.

Results: A total of 163 patients with meningioma were assembled into the patient cohort. 71% of the patients had skull-based tumors, with 34% of the patients treated with single-fraction SRS. Tumor grade and Ki-67 of the treated tumor were available for 37 and 34% of the patients, respectively, with 25% of those biopsied having WHO grade II or III tumors. 47% of patients received a prior resection. There were only 18 events of local recurrence at the time of last follow-up. Machine learning models using quantitative imaging and clinical features demonstrated discriminatory ability in the prediction of local recurrence. Model performance improved with the incorporation of clinical variables compared to with radiomic features alone (e.g., [Fig. 1], p < 0.001).

Conclusions: In this feasibility study, we demonstrated that quantitative imaging features combined with clinical variables show ability to model local recurrence of meningioma following definitive treatment. Model performance was augmented with the incorporation of clinical variables with radiomic features. Given the low rate and late risk of local recurrence for meningiomas treated with radiotherapy, more extensive series with longer follow-up are likely necessary to investigate further optimization and comparative performance of machine learning and statistical models.

Fig. 1 Boosting-based tree sampling ensemble method ROC curve with and without clinical variables incorporated.

Publication History

Article published online:
12 February 2021