Abstract:
Colorectal cancer (CRC) is one of the leading causes of cancer-related deaths worldwide; thus, reliable early assessment, proper clinical risk stratification, and prognosis of cancer outcomes are essential parts of modern oncology. This research studies the efficiency of advanced machine learning and deep learning methods for predicting 5-year survival using a dataset comprising 167,497 records of patients based on 28 clinical, demographic, socio-economic, and health system variables. The major tasks of this study were to create and compare robust predictive models, to identify the important decisive factors defining the patient's survival, and to develop the best methodological techniques of prognostic assessment in clinical practice. For the purpose of evaluation, an 80/20 stratified random data-splitting strategy (NTrain =133997, NTest =33500) was implemented in addition to feature preprocessing and dummy coding for categorical variables. Various classification algorithms were applied during the research, such as Classification and Regression Trees (CART), Random Forests, Linear Support Vector Machines (LinearSVC), Non-Linear RBF Kernel Support Vector Machines, Linear Models with Stochastic Gradient Descent (SGD), and a 4-layer Deep Neural Network (DNN) with Batch Normalization and Dropout. The evaluation of models was supplemented with a variety of diagnostic metrics that comprise accuracy, precision, recall, F1- score, Area Under the Receiver Operating Characteristic Curve (ROC-AUC), and confusion matrix analysis. The first diagnostic model based on the synthetic baseline target label (Survival Prediction) reported a significant majority class collapse across the algorithms, which limited the provided accuracy to 59.96% and made the ROC-AUC equal to 0.5000 due to statistical independence of input predictors and the target variable. After changing the prediction target to the real clinical outcome (Survival_5_years with mortality rules), optimized ensembles were able to achieve a predictive accuracy of 75.82%, 99.78% precision, 99.81% specificity, and an ROC-AUC of 0.7978. The analysis of feature importance has shown that patient costs in healthcare (30.1%), size of the primary tumor in millimeters (11.2%), age of the patient (9.6%), incidence rates in the region (9.5%), and mortality rates in the region (8.1%) were the main determinants of 5-year survival probability. In conclusion, although both high-capacity deep neural networks and sophisticated support vector machines exhibit remarkable architectural adaptability, their predictive value is contingent on the availability of ground-truth labels and physiological biomarkers, which can transform predictive pipelines into clinically useful predictors of survival risk based on the event outcome.
