Predicting Side Effects of Chemotherapy: Leveraging Machine Learning for Cancer Symptom Management

This study investigates the potential of machine learning (ML) to predict symptom escalation in chemotherapy patients using daily self-reported data. We analyzed self-reported health information from 339 patients, tracking 12 symptoms quantified by severity and distress to compile a comprehensive symptom score ranging from 0 to 230. To address the inherent challenge of dataset imbalance, we employed a stratified sampling technique. This involved dividing the dataset into several subgroups based on the severity of symptoms, and then randomly selecting an equal number of instances from each subgroup. This created a balanced dataset, thereby enhancing the reliability and robustness of our predictive models. We rigorously tested nine diverse ML models, including decision trees, discriminant analysis, and support vector machines (SVM), to predict changes in the total symptom score based on data from the preceding 3 to 5 days. These models were trained on the balanced dataset to mitigate the effects of the original data imbalance, and performance was also compared on the unbalanced dataset to evaluate the impact of data preparation on model efficacy. Our analysis highlighted the SVM classifiers' high accuracy rate of 82% on the unbalanced dataset, though they often misclassified severe symptom changes as less severe. In contrast, the Ensemble algorithm augmented with the RUSBoost classifier demonstrated exceptional precision across both datasets. It achieved accuracy rates of 61.16%, 58.41%, and 60.05% for datasets spanning 3, 4, and 5 days, respectively. This model effectively identified significant symptom changes, emphasizing the critical role of balanced datasets in ML training within health care contexts, where precision significantly affects patient care and outcomes. This research demonstrates the powerful capabilities of ML in cancer care, advocating for advanced algorithmic approaches and the optimized use of balanced datasets to enhance the effectiveness and impact of patient care strategies in cancer symptom management.