Koroner arter hastalığının makine öğrenmesi teknikleriyle teşhisi = Diagnosis of coronary artery disease using machine learning techniques

Abstract

Bu yüksek lisans tezi, koroner arter hastalığı (KAH) teşhisinde makine öğrenimi yöntemlerinin etkinliğini ve hibrit bir yaklaşımın potansiyel faydalarını kapsamlı bir şekilde araştırmaktadır. KAH, dünya genelinde önde gelen ölüm nedenlerinden biri olup, erken ve doğru teşhis hayati önem taşımaktadır. Bu çalışmada, farklı özellik sıralama algoritmaları ve makine öğrenimi modelleri kullanılarak KAH teşhisinin iyileştirilmesi hedeflenmiştir. Bu amaçla, çeşitli veri setleri üzerinde kapsamlı deneyler gerçekleştirilmiş ve sonuçlar detaylı bir şekilde analiz edilmiştir. Çalışmada, IEEE Dataport, UCI Machine Learning Repository ve Kaggle'dan elde edilen dört farklı veri seti kullanılmıştır. Bu veri setleri, KAH teşhisinde kullanılan çeşitli klinik ve demografik özellikleri içermektedir. Veri setlerinin çeşitliliği, çalışmanın bulgularının genellenebilirliğini artırmaktadır. Eta, F Skor, Fisher ve tez kapsamında geliştirilen Hibrit Özellik Seçimi algoritmaları uygulanarak, her bir veri seti üzerinde önem derecesine göre farklı özellikler içeren alt veri setleri oluşturulmuştur. Özellik seçimi, modellerin performansını artırmanın ve gereksiz özelliklerin yol açabileceği gürültüyü azaltmanın yanı sıra hesaplama karmaşıklığını da önemli ölçüde azaltarak verimliliği artırmanın kritik bir yoludur. Oluşturulan alt veri setleri kullanılarak ANN, kNN, SVM, EDT ve Hibrit makine öğrenimi modelleriyle hastalık sınıflandırması yapılmıştır. Bu modeller, KAH teşhisinde yaygın olarak kullanılan güçlü makine öğrenimi algoritmalarıdır. Hibrit Makine Öğrenmesi yaklaşımı, farklı modellerin güçlü yönlerini birleştirerek tek bir modelin sınırlılıklarını aşmayı amaçlar. Bu çalışmada, dört farklı modelin tahminlerinin ortalaması alınarak hibrit bir model elde edilmiştir. Hibrit Özellik Seçimi algoritması ise, üç farklı özellik sıralama algoritmasından elde edilen özellik skorlarının birleştirilerek yeni bir sıralama oluşturulması prensibine dayanmaktadır. Bu hibrit yaklaşımlar, KAH teşhisinde daha doğru ve kararlı sonuçlar elde etmek için geliştirilmiştir. Modellerin performansı, Doğruluk, Duyarlılık, Özgüllük, F-skor, Kappa ve AUC metrikleri kullanılarak değerlendirilmiştir. Bu metrikler, modellerin farklı yönlerden performansını ölçerek kapsamlı bir değerlendirme sağlar. Farklı sayıda özellik içeren veri setlerinden oluşturulan modellerin test sonuçları karşılaştırılarak, en iyi performansı gösteren modeller ve özellik kombinasyonları belirlenmiştir. Bu analiz, KAH teşhisinde hangi özelliklerin ve modellerin daha etkili olduğunu ortaya koymaktadır. Bu çalışma, KAH teşhisinde makine öğrenimi yöntemlerinin ve hibrit bir yaklaşımın kullanımının potansiyel faydalarını kapsamlı bir şekilde göstermektedir. Elde edilen sonuçlar, özellik seçimi ve model kombinasyonunun teşhis performansını önemli ölçüde artırabileceğini ortaya koymaktadır. Geliştirilen Hibrit Özellik Seçimi algoritması ve Hibrit Makine Öğrenmesi modeli, KAH teşhisinde yeni ve etkili bir yaklaşım sunmaktadır. Bu çalışma, KAH'ın erken teşhis ve tedavisine yönelik gelecekteki araştırmalara ve klinik uygulamalara katkı sağlayabilecek değerli bilgiler sunmaktadır. Ayrıca, bu alanda daha ileri çalışmalar için bir temel oluşturmaktadır.

This master's thesis provides an in-depth exploration of the effectiveness of machine learning techniques in diagnosing Coronary Artery Disease (CAD), one of the most critical health challenges globally due to its significant mortality and morbidity rates. The accurate and timely diagnosis of CAD plays a pivotal role in reducing its devastating impacts, and this study aims to contribute to that effort by leveraging machine learning methodologies alongside advanced feature selection techniques. The research is designed to improve diagnostic precision, reduce diagnostic latency, and enable a non-invasive yet reliable approach to detecting CAD at an early stage. The study utilizes four distinct datasets sourced from IEEE Dataport, the UCI Machine Learning Repository, and Kaggle. These datasets incorporate a variety of clinical and demographic features commonly associated with CAD diagnosis, such as patient medical histories, vital signs, and other health indicators. The inclusion of diverse datasets enhances the generalizability of the findings and allows for comprehensive training and validation of machine learning models across multiple data sources. This diversity also ensures that the study's conclusions are robust and applicable to real-world scenarios, as the datasets collectively represent a wide spectrum of patient profiles and clinical contexts. Feature extraction and selection are central to this study, aiming to optimize model performance by focusing on the most relevant attributes while minimizing noise and computational complexity. Four feature selection techniques were employed: Eta, F-Score, Fisher, and a novel Hybrid Feature Selection algorithm developed specifically for this research. The Hybrid Feature Selection algorithm innovatively combines feature importance scores derived from multiple selection methods to create a unified ranking of features, thereby balancing the strengths of individual techniques. This approach ensures that critical features are identified and utilized, significantly improving the diagnostic capabilities of the machine learning models. The machine learning models used for classification in this research include Artificial Neural Networks (ANN), k-Nearest Neighbors (kNN), Support Vector Machines (SVM), Ensemble Decision Trees (EDT), and a Hybrid Machine Learning model. The hybrid model is a notable contribution, combining predictions from multiple base models to create a more accurate and robust diagnostic tool. By integrating the strengths of different algorithms, the hybrid model overcomes the limitations of individual classifiers and provides a unified framework for CAD diagnosis. This innovative approach ensures that the diagnostic system is not only precise but also adaptable to various datasets and feature configurations. The performance of the models was assessed using several key metrics, including Accuracy, Sensitivity, Specificity, F-Score, Kappa, and Area Under the Curve (AUC). These metrics provide a holistic evaluation of each model's effectiveness in identifying CAD cases while minimizing false positives and false negatives. A comparative analysis was conducted to understand how different feature selection techniques and machine learning models perform across varying datasets and feature subsets. This analysis revealed that the choice of feature selection method and classifier significantly impacts diagnostic performance and that these choices must be tailored to the characteristics of the dataset for optimal results. One of the primary findings of this study is the variability in the effectiveness of feature selection techniques across datasets. For instance, in some datasets, the Eta and Hybrid Feature Selection algorithms performed exceptionally well at higher feature percentages, capturing more nuanced patterns in the data. In contrast, F-Score and Fisher methods were more effective with lower feature percentages, demonstrating their utility in reducing data dimensionality without compromising diagnostic accuracy. This underscores the importance of dataset-specific optimization in the feature selection process. Similarly, the performance of machine learning models varied based on the datasets and feature subsets used. SVM and ANN consistently demonstrated high performance when applied to datasets with a larger number of features, leveraging their capacity to handle complex data relationships. In contrast, kNN and the hybrid model performed better in scenarios with reduced feature sets, showcasing their efficiency in computationally constrained environments. These findings highlight the importance of selecting the appropriate machine learning model based on the specific characteristics of the data, as well as the value of combining models to achieve superior diagnostic outcomes. An important distinction of this study is its ability to diagnose CAD using only clinical and demographic data, without relying on observable disease symptoms. This non-invasive diagnostic approach offers significant advantages, particularly for early detection. By using readily available health records and clinical data, this method enables widespread screening and timely intervention, potentially preventing the disease from reaching more severe stages. This capability is especially critical in healthcare settings with limited resources, as it minimizes the need for costly or invasive diagnostic procedures. The research findings demonstrate the significant potential of integrating feature selection techniques and machine learning models for CAD diagnosis. The Hybrid Feature Selection algorithm developed in this study effectively balances the strengths of individual methods, capturing critical features that might otherwise be overlooked. Similarly, the Hybrid Machine Learning model combines the predictive power of multiple algorithms to achieve higher accuracy and reliability. These hybrid approaches not only enhance the diagnostic process but also establish a foundation for future advancements in the field. The study's results are consistent with existing literature, reinforcing the validity and applicability of the proposed methodologies. The findings contribute valuable insights into the interplay between data characteristics, feature selection techniques, and machine learning models, providing a roadmap for optimizing diagnostic systems. Moreover, the research aligns with the growing emphasis on leveraging artificial intelligence and machine learning in healthcare, demonstrating how these technologies can transform diagnostic practices. In conclusion, this master's thesis advances the field of CAD diagnosis by presenting a comprehensive framework that combines innovative feature selection techniques with cutting-edge machine learning models. The proposed methodologies achieve high diagnostic accuracy, computational efficiency, and adaptability to diverse datasets. By demonstrating the feasibility of a non-invasive, data-driven approach to CAD diagnosis, this research offers a practical solution for early detection and intervention, ultimately improving patient outcomes. The findings lay a solid foundation for future research and clinical applications, paving the way for continued innovation in the use of machine learning for medical diagnostics.