A Hybrid Model for Recurrent Head and Neck Squamous Cell Carcinoma Prognosis in Ghana

Abstract

Despite the rapid advancement in the development of hybrid ensemble Machine Learning (ML) techniques in malignancy management, recurrence and mortality from Head and Neck Squamous Cell Carcinoma (HNSCC) subtypes have not significantly improved in recent decades due to poor prognosis. Moreso, the recurrent HNSCC prognoses increase in patients with HNSCC due to the metastatic stage of the tumor at diagnosis, but studies providing promising prognostic models as a supporting tool for recurrence classification and prediction in HNSCC are lacking. As a supporting tool for identifying the most accurate prognosis and a robust prognostic classification model for classifying HNSCC recurrence patterns, this study presents a hybrid stacked ensemble classifier model when the same ML classifiers for; feature selectors, base classifiers, and meta classifiers are used, that could accurately predict recurrence outcomes and identify the most newly accurate prognostic features in HNSCC recurrence. Retrospective data of 125 HNSCC patients treated with curative intent between 2016 and 2020 at KBTH and who had a follow-up within this calendar period are collected. Data is pre-processed using mode imputation and one-hot encoding. The proposed Hybrid Ensemble Super Classification Algorithm (HESCA) model uses the ML classifier models including Gradient Boosting Machine (GBM), Distributed Random Forest (DRF), Deep Neural Network (DNN), Generalised Linear Model (GLM), and Naïve Bayes (NB) for stacked ensemble learning. These classifier models are employed in constructing feature subsets, base classifiers, and with each as a meta-classifier in a stacking ensemble. The performances of the HESCA model on various feature subsets are compared. Next, the performance of the HESCA model on 8-input features is compared with the HESCA model on full-input features. Then, based on gradient-boosted features, the performance of the HESCA model is compared with the established stacked ensembles. Thus, the two baseline stacked ensemble models, and one state-of-the-art stacked ensemble model. The results show that when the GBM classifier is used as a meta-classifier in a stacking ensemble consisting of five base classifiers on gradient-boosted features (GBM-input features) including concurrent chemoradiotherapy treatment, age at diagnosis, p63, cervical lymph/neck nodes, tumor size, smoking habit, pathological tumor staging at T4, and stage IV of tumor at diagnosis achieves higher accuracy (90.63%) with the least log loss (0.2959) compared to that achieved by base models and the established stacked ensemble models on the same gradient boosted features of recurrent HNSCC prognostic data. This gives a hybrid stacked ensemble model termed the HESCA model, which consists of five base models under study and a GBM meta-model. It is also observed that this HESCA model on GBM-input features achieves better classification evaluation measures than that achieved on any other input feature subsets as well as the full-input feature subset considered in this study. The study shows that using the GBM classifier as a meta-classifier model in a stacking ensemble having five base classifiers with its gradient-boosted features results in better performance than base models and any other established stacked ensemble model used in this study; and using the HESCA model with gradient boosted features is clinically appropriate as a supporting tool for identifying, classifying and predicting patients' recurrent HNSCC prognostic data.