Purpose & Background:

Methods Continued:

Results:

Conclusions:

Limitations & Future:

The purpose of this study was to develop and evaluate a machine-learning model that predicts whether a glioma is lower-grade glioma (LGG) or glioblastoma (GBM) using patient clinical characteristics and tumor mutation data.

Machine Learning Based Classification of Glioma Grade Using Clinical and Somatic Mutation Features

Momin Mahmood, Aiden Adkins, Ryan Gosine

What is a glioma?

A glioma is a tumor that develops from cells associated with the brain’s supportive glial tissue.

• Lower grade glioma (LGG) generally grows more slowly but may progress into a more aggressive tumor. (Grades 1 or 2)
• Glioblastoma (GBM) is a fast growing, highly aggressive form of glioma associated with poorer patient outcomes. (Grade 4)

​

Our dataset

839 patients | 487 LGG | 352 GBM

Each patient record included:

• 20 binary mutation features, indicating whether genes were mutated
• 3 clinical characteristics: age, gender, and race
• The patient’s known glioma classification: LGG or GBM

​

Exploratory Data Analysis

​

1-Hot Encoding & Feature Engineering

Phi-K Correlation Analysis

​

​

Test-Train Split & Model Development

SHAP Plot

& XAI

Multiple Correspondence Analysis Biplot:

Unlike Pearson Spearman correlation, phi K evaluates mixed data types capturing non-linear relationships between binary mutation indicators, continuous clinical variables (Age), and the ordinal target (Grade).

IDH1 (phi_K = 0.89): Shows an extremely strong correlation with tumor grade (deep red), making it the single most predictive feature.

Age at diagnosis (phi_K = 0.68): Has a strong positive association with Grade

IDH1 & Age at diagnosis (phi_K = 0.74): Indicates a strong co-dependence between patient age and IDH1 mutation status.

​

• Show where 95% of each group falls. Overlap means LGG and GBM patients can't be fully separated by just these two dimensions.
• Molecular confirmation of an IDH_1 mutation is required for the diagnosis of the majority LGGs (consistent with biplot result)
• Each data point is a patient’s unique genetic profile manipulated into the twin axae

​

​

​

​

​

- Starts with a baseline prediction and iteratively builds decision trees. New tree focuses specifically on correcting the misclassifications (residuals) of the previous trees on patient data involving gradients.

- Hyperparameter Tuning: Fine-tunes tree learning rate (v) and number of learning epochs to prevent overfitting on rare gene mutations, while lowering the learning rate so single strong features don't dominate early on.

- Combines the weighted outputs of all trees to LGG (Grade 0) or GBM (Grade 1).

Methods:

Our Tuned XGBoost was successful in detecting 66 of 70 Glioblastoma cases, having achieved a 94.3% recall and 9.30 ROC-AUC (Receiver Operating Characteristic - Area Under Curve).

​

Prioritizing Glioblastoma recall led to a reduction in the number of missed cases, falling to an even 4. Doing this had its drawbacks, leading to 18 cases of Low-Grade Glioma being detected as Glioblastomas.

Gini importance algorithm used to designate 5 most contributory genes Axae account for inertia (variability in data set).

XGBoost Tuned Model:

Results Continued

Tuning

Boosting With Optimal Hyper-parameters

Model Comparison

Ten tuned classifiers were evaluated on the held out test set of 168 patients. Five models tied for the highest GBM recall of 94.3%. Tuned XGBoost was selected because it combined this high sensitivity with the strongest overall performance balance: 86.9% accuracy, 78.6% precision, 85.7% F1-score, and 0.930 ROC-AUC.

Employed 1 hot encoding on race to convert it from a categorical to a nominal variable for optimal computing.

In all, we assed models including single decision trees, linear and distance-based classifiers (Logistic Regression, SVM), ensemble approaches (Bagging, Random Forest, AdaBoost, Gradient Boosting, XGBoost, Stacking), and neural networks (MLP). These models were constructed using standard sklearn/imblearn pipelines with python in google colab. Recall was prioritized as the main metric in order to reduce false negatives.

This study was limited by its relatively small, single-source dataset, limited clinical variables, possible demographic underrepresentation, and lack of external validation, which may restrict how well the model generalizes to other patient populations. In the future, we will collect more diverse patient data, validate the model on an independent dataset, evaluate the demographic subgroups, and decrease the number of false positives.

Reference: UCI Machine Learning Repository. (2022). Uci.Edu. https://archive.ics.uci.edu/dataset/759/gliomgradia+ng+clinical+and+mutation+features+dataset

LGG

GBM

SHAP values measured how

each feature influenced

XGBoost’s predictions.

Each point represents one

patient: positive values push

the prediction toward GBM,

negative values push toward

LGG, and color represents the feature value or mutation status.

​

IDH1 was the strongest predictor, followed by age at diagnosis, TP53, PTEN, and IDH2. IDH1 mutation generally favored LGG, while the absence of IDH1 mutation and older age favored GBM. These results explain the model’s predictions but do not establish causation.

​

The XGBoost feature-importance chart ranks the variables according to their contribution to the model’s tree splits; unlike SHAP, it summarizes overall influence but does not show whether a feature pushes a prediction toward LGG or GBM.

Machine Learning Based Classification of Glioma Grade Using Clinical and Somatic Mutation Features

Momin Mahmood, Aiden Adkins, Ryan Gosine

Machine Learning Based Classification of Glioma Grade Using Clinical and Somatic Mutation Features

Momin Mahmood, Aiden Adkins, Ryan Gosine

Machine Learning Based Classification of Glioma Grade Using Clinical and Somatic Mutation Features

Momin Mahmood, Aiden Adkins, Ryan Gosine