RemissIA: a remission identification approach in cancer patients using machine learning ensembles

Cancer continues to be a leading cause of death worldwide, ranking first or second in over 100 countries. To enhance current healthcare systems, artificial intelligence, particularly machine learning, has emerged as a promising tool to aid in medical decision-making. A crucial factor in assessing a patient’s health status is the possibility of remission. In this study, we employed five machine learning algorithms - logistic regression, random forest, XGBoost, multilayer perceptron and convolutional neural network - along with an ensemble model using majority voting among these models. We utilized the Surveillance, Epidemiology, and End Results (SEER) dataset from the National Cancer Institute (NCI). Our approach involved analyzing the entire SEER database and also separating it into nine categories corresponding to different cancer types. We performed a train-validation-test split and evaluated the models using the F1 metric, comparing the results using the Wilcoxon statistic with a 95% confidence interval analysis. In the findings, both MLP and CNN achieved F1 metrics above 75% across all nine tumor groupings as well as the complete SEER database encompassing all tumor types. The Wilcoxon statistic indicated no significant difference between MLP and CNN in seven out of the ten analyzed groups. However, the proposed ensemble model did not outperform the individual models due to the lower performance of the logistic regression and random forest models. This research successfully achieved its objectives of training algorithms without restricting the analysis to a specific tumor or cancer type, thus making a valuable contribution to the fields of oncology and artificial intelligence.