การพยากรณ์ช่วงเวลาการรอดชีวิตในผู้ป่วยมะเร็งปากมดลูกโดยใช้เทคนิคการเรียนรู้ของเครื่อง

Abstract

Cervical cancer is considered a dangerous and common disease among women all over the world, and it still carries a relatively high mortality rate. Prediction patient’s survival is an important aspect of research due to its significance in enhancing treatment efficacy. Currently, survival prediction heavily relies on statistical methods, which has limitations when dealing with non-linear correlated data. Therefore, this research aims to apply and ensemble machine learning techniques through the selective stacking method to construct a classification model to predict the survival interval and a regression model to make a numerical survival prediction. In this research, multiple machine learning were trained and the top three most efficient models were selected for another round of machine learning. This approach enhance prediction accuracy when compared to the results from models generated solely using conventional machine learning techniques. This research utilized patient data from the medical records of the Faculty of Medicine, Chiang Mai University, which includes factors such as patient age, tumor size, pathology, tumor stage, chemotherapy method, brachytherapy method, radiation dose administered, side effect status, start date of diagnosis, end date of diagnosis, date of local recurrence, date of distant recurrence, date of death, and survival status. The classification model categorized survival periods into fur ranges: “< 6 months”, “6 months – 3 years”, “3 years – 5 years”, and “> 5 years” while the regression model use the interval between the start date of diagnosis and death of patients as the target output. The feature importance was then identified from regression model using the local interpretable model-agnostic explanations (LIME) technique. The results showed that models created using the selective stacking method provided more accurate predictions compared to models generated using traditional machine learning techniques in both classification and regression. The classification model achieved an accuracy of 87.81 percent, and the regression model had a root mean squared error of 18.92 and a correlation coefficient of 0.669. The most influential factor on survival was the side effect status around any organs. By observing individual patient prognostic results, factors that tend to contribute to longer or shorter survival times could be identified. This information can be valuable for treatment planning, resource allocation, and providing appropriate treatment guidelines for patients.