Room: Karl Dean Ballroom C
Purpose: Human papillomavirus (HPV) and miRNAs have been known to be a driving oncogenic factor in oropharyngeal cancer as well as a significant biomarker for patient survival. However, the reliable and efficient usage of these biomarkers for accurate cancer prediction is very challenging due to the biomarker uncertainty and redundancy. We developed an effective model to select the most predictive biomarkers for accurately predicting the clinical outcome by use of advanced machine learning method.
Methods: For each patient case, genomic features, including Human Papilloma Virus (HPV) status and the microRNA expression, were extracted. Then, a belief-function theory-based feature selection method was proposed to automatically identify the most predictive feature subset from the extracted genomic features through the minimization of a specific loss function. Based on BFT, the loss function was defined by considering three requirements of predictive features: 1) high prediction accuracy, 2) low uncertainty, and 3) high sparsity to reduce over-fitting risk on unseen patient samples. Finally, a BFT-based classifier, evidential K-NN method (EK-NN), was trained to predict the treatment outcome given as input the selected predictive feature subset.
Results: Experimental results on 101 oropharyngeal cancer patients and comparisons with other reported methods illustrate the superior performance of the proposed model. Our model achieved high prediction accuracy and AUC of 0.86 and 0.88, respectively. We also evaluated the prediction performance of the new EK-NN classifier with respect to K under different sparsity degrees. Our proposed ToE-based method improves the robustness of traditional EK-NN classifier in terms of parameter K, thus ensuring the stability of outcome prediction.
Conclusion: Our proposed prediction model handles well-known challenges associated with the use of genomic features for cancer therapy outcome prediction. This model will provide a complete strategy to improve the accuracy and stability of cancer prediction.