Predicting the 1p/19q Codeletion Status of Presumed Low-Grade Glioma with an Externally Validated Machine Learning
Sebastian R van der Voort1, Fatih Incekara2,3, Maarten M J Wijnenga4
1Biomedical Imaging Group Rotterdam, Departments of Radiology and Nuclear Medicine Medical Informatics, Erasmus MC-University Medical Centre Rotterdam, Rotterdam, the Netherlands.
Summary
A new machine learning algorithm can predict 1p/19q codeletion status in low-grade glioma (LGG) using MRI scans before surgery. This noninvasive method shows promise for guiding treatment decisions in LGG patients.
Area of Science:
- Neuro-oncology
- Medical imaging
- Artificial intelligence
Background:
- 1p/19q codeletion in low-grade glioma (LGG) is a key prognostic and predictive biomarker.
- Patients with 1p/19q codeleted LGG exhibit improved survival and treatment response.
- Accurate pre-operative determination of 1p/19q status is clinically significant for treatment planning.
Purpose of the Study:
- To develop and validate a machine learning algorithm for noninvasive prediction of 1p/19q codeletion status in LGG.
- To assess the algorithm's performance against expert clinical evaluation using preoperative MRI.
- To determine the generalizability and robustness of the predictive model.
Main Methods:
- A support vector machine algorithm was trained on preoperative MRI features (T1-weighted post-contrast, T2-weighted) and patient demographics (age, sex) from 284 LGG patients.
- The algorithm's predictive performance was evaluated on an independent external validation dataset of 129 LGG patients from The Cancer Imaging Archive (TCIA).
- Performance was compared against predictions made by four clinical experts (neurosurgeons and neuroradiologists).
Main Results:
- The machine learning algorithm achieved an Area Under the Curve (AUC) of 0.72 on the external validation dataset.
- The algorithm's predictive performance surpassed the average performance of neurosurgeons (AUC 0.52) but was lower than that of neuroradiologists (AUC 0.81).
- Significant variability was observed in the performance of individual clinical experts (AUC range 0.45-0.83).
Conclusions:
- The developed machine learning algorithm offers a robust and generalizable method for noninvasively predicting 1p/19q status in presumed LGG.
- The algorithm's performance, on average, exceeded that of oncological neurosurgeons, suggesting its potential utility in clinical decision-making.
- This AI-driven approach could aid in optimizing treatment strategies for LGG patients prior to surgical intervention.
