Machine learning applications in the diagnosis of heart disease in pediatric chest pain

Although chest pain is an infrequent symptom in childhood, it can lead to considerable concern among families due to the potential for underlying cardiac causes. Therefore, the evaluation of pediatric patients in terms of heart diseases and the timely identification of possible pathologies hold great clinical significance. However, a significant portion of chest pain stems from musculoskeletal, gastrointestinal, respiratory, or psychological causes. The rarity of cardiac-related cases leads to unnecessary tests during the diagnostic process and a waste of healthcare resources. In this context, a machine learning-based classification model was developed in the study to determine the etiology of pediatric chest pain; the diagnostic performance of the model was evaluated using various algorithms, hyperparameter optimization techniques, and explainability analyses. Within the scope of the thesis, a dataset containing demographic information, clinical symptoms, and ECG findings of pediatric patients aged between 3 and 18 years was obtained from a university hospital. Although the study is retrospective, the number of data points related to sick children is quite limited. Therefore, special attention was given to the data preprocessing phase; missing data were handled, categorical variables were encoded, and class imbalance was corrected using the SMOTE method. Subsequently, a comparative performance analysis was conducted by applying various supervised learning algorithms such as Random Forest, Extra Trees, XGBoost, and LightGBM. The top three performing models were identified, and hyperparameter optimization and cross-validation were carried out using GridSearchCV, RandomizedSearchCV, and Optuna. Additionally, explainability analysis was performed using SHAP and LIME methods to visualize the decision making processes of the models and interpret them in clinical terms. According to the study results, LightGBM, Extra Trees, and XGBoost algorithms successfully predicted the presence of heart disease in pediatric patients with accuracy scores above 94%, ROC AUC scores around 99%, and balanced classification metrics. SHAP analyses revealed that variables such as troponin levels, ECG rhythm findings, and characteristics of chest pain played a critical role in the model predictions. Moreover, with LIME analysis, the model decisions were interpreted on an individual patient basis. Thus, the model was shown to be not only high performing but also clinically valid and reliable. Throughout the study, the integrated use of machine learning algorithms not only as standalone classifiers but also across multiple stages including data preprocessing, feature selection, hyperparameter optimization, and explainability significantly enhanced the overall success of the system. This decision support system, developed with consideration of the sensitivities specific to pediatric patients, accelerates early diagnosis processes, optimizes the use of clinical resources, and supports the diagnostic decisions of healthcare professionals. This thesis demonstrates that machine learning-based models can be effectively used to differentiate the causes of chest pain in children and presents an explainable, reliable, and integrable artificial intelligence-based system that can contribute to clinical decision-making. The findings show that the integration of artificial intelligence into pediatric healthcare applications is feasible and provide a solid foundation for future research.