A multi-view prognostic model for diffuse large B-cell lymphoma based on kernel canonical correlation analysis and...

Area of Science:

• Artificial Intelligence in Oncology
• Medical Imaging Analysis
• Machine Learning for Prognostics

Background:

• Positron emission tomography/computed tomography (PET/CT) is the standard for diffuse large B-cell lymphoma (DLBCL) staging.
• Existing prognostic models often overlook quantitative imaging features from PET/CT scans.
• Integrating clinical and imaging data can enhance prognostic accuracy for DLBCL patients.

Purpose of the Study:

• To develop a multi-view learning (MVL) model utilizing both clinical and imaging data for DLBCL prognosis prediction.
• To improve decision-making for clinicians by providing more accurate patient prognoses.
• To compare the performance of the proposed MVL model against single-view learning models and other MVL approaches.

Main Methods:

• Feature engineering involved extraction, recursive feature elimination, and principal component analysis on clinical and imaging data.
• A Support Vector Machine (SVM) model (SVM-2K) was developed using kernel canonical correlation analysis (KCCA) on mapped clinical and imaging features.
• Model performance was evaluated using accuracy, sensitivity, F1 score, AUC, and G-mean on a test dataset.

Main Results:

• The SVM-2K model achieved superior performance with AUC of 92.1%, accuracy of 96.9%, sensitivity of 90.9%, F1 score of 92.8%, and G-mean of 91.4%.
• All MVL models demonstrated better performance than the best single-view learning model.
• Feature engineering significantly improved the performance of the SVM model on DLBCL test data.

Conclusions:

• Multi-view learning models significantly outperform single-view learning models in predicting DLBCL patient prognosis.
• The proposed SVM-2K model demonstrates excellent performance and accuracy in prognostic prediction.
• This MVL approach offers a valuable tool for assisting clinical decision-making in DLBCL management.