Cancer disease tracking with machine learning methods

Breast, lung, prostate and stomach cancers have been the most common types of cancer worldwide. Detection and diagnosis of these cancers at an early stage poses a challenge in the literature. When dealing with cancer patients, physicians choose among various treatment methods with risk factors. The treatment program is critical in clinical decision making, as the risks of treatment may outweigh the benefits. This program is prepared by looking at the previous comorbidities of the patient, the medications he took and the treatment procedures he had undergone. Manually deciding which drug and treatment to use can be time consuming and difficult. In this thesis, a computational solution is presented for breast, lung, prostate and stomach cancer patients' in-hospital post-diagnosis mortality prediction using machine learning approaches that keep the prediction rate as high as possible. The solution is based on the analysis of diagnostic, drug and procedural parameters that are easily available from electronic health systems. In order to predict the mortality outcomes of cancer patients, a classification-based approach has been introduced, model training has been carried out, and the performances of these classifiers have been evaluated. Logistic Regression, Decision Tree, Random Forest, Support Vector Machine and Multi Layer Perceptron classifiers were evaluated on the Medical Information Mart in Intensive Care IV (MIMIC-IV) dataset and various experiments were carried out with them. Diagnosis, drug and treatment features were extracted for the specified cancer patients, and feature selection related to Logistic Regression was made. In order to use easily accessible electronic health data and to make the procedures to be done quickly and effectively in terms of memory and time usage, the classifier structure was designed to provide successful results with less data and it was aimed to reduce the burden of doctors. The mortality prediction abilities of the machine learning models were evaluated with the F1 Macro Mean and AUC-ROC score metrics. The best F1 scores were 0.74 for breast, 0.73 for lung, 0.82 for prostate and 0.79 for stomach cancer. The best AUROC scores were 0.94 for breast, 0.91 for lung, 0.96 for prostate and 0.88 for stomach cancer. As a result, using the most relevant features, it was seen that the results were similar to the main result for each cancer type, and it was revealed that this approach can be easily used in healthcare facilities where data and resources are limited.