Vis sammendrag
Cancer is one of the leading factors in global mortality,, with head and neck squamous cell carcinoma (HNSCC) being the seventh most common cancer worldwide [2, 3]. Survival analysis using machine learning models can help predict overall survival (OS) and disease-free-survival (DFS). In survival analysis, radiomics features can be used together with clinical features for more accurate prediction of both endpoints. Radiomics features are numerous quantitative features extracted from mathematical analysis of medical images, including Computed Tomography (CT), Magnetic Resonance Imaging (MRI), and Positron Emission Tomography (PET), which are difficult to capture with the human eye. The goal of this thesis was to predict OS and DFS through survival analysis using combinations of datasets, feature scaling or transformation methods, feature selection methods, and survival analysis models. Each combination was evaluated using Concordance Index (C-Index) and Integrated Brier Score (IBS). Three types of data were used, data with only clinical features, data with only radiomics features, and data with both clinical and radiomics features. Models were trained using a train dataset from Oslo University Hospital (OUS) comprising 139 patients, and tested on a test set from MAASTRO clinic, comprising 99 patients. Different combinations of three datasets, four feature scaling or transformation methods, seven feature selection methods, and eight survival analysis models were tested. Furthermore, the most commonly selected features for OS and DFS were analyzed. For predicting OS, the highest C-Index of 0.708 on the test set was achieved using D1, Yeo-Johnson transformation, PLSR, and Random Survival Forest. Using IBS as a metric, the lowest IBS of 0.198 was obtained with D3, Yeo-Johnson transformation, CFI, and Random Survival Forest. For predicting DFS, the highest C-Index of 0.677 on the test set was obtained with D3, Min-Max scaling, all features, and component-wise gradient boosting. The lowest IBS of 0.213 was achieved using D1 with different combinations of Min-Max scaling, RENT, and Extra Survival Trees, as well as D1 with Yeo-Johnson scaling, RF, and Component-wise Gradient Boosting. The most commonly selected features for both endpoints were hpv related, uicc8 III − IV , and shape sphericity. No single combination outperformed the others for both endpoints, as multiple combinations produced similarly high C-Index or low IBS. To generalize these findings to other datasets, a larger number of samples is needed. Future work can include statistical testing between the different combinations and also consider the stability of each combination.
