Hierarchical Ensemble Model for Thyroid Cancer Detection using Multimodal Data

Early and accurate disease diagnosis is essential in the medical industry. This study emphasizes how crucial machine learning is for early diagnosis and how challenging it is to diagnose thyroid disease due to its wide range of symptoms. Both a blood test dataset and an ultrasound image dataset were analyzed. For the blood test data, models such as Logistic Regression, Support Vector Machine, K-Nearest Neighbor, Random Forest and Gradient Boosting were employed; Random Forest scored the highest, with 97.46 % accuracy. When applied to ultrasound images, deep learning models like CNN, VGG19, and Vision Transformer (ViT) generated an accuracy of 97.93%. The risk probability ratings of the two top-performing models, ViT and Random Forest, were integrated using soft voting, an ensemble technique that averaged their risk probability scores.in an effort to increase forecast precision, this approach improved the overall accuracy to 99.72%. The models were assessed and key features were identified using SHAP (SHapley Additive exPlanations) to increase transparency. The gradient improves more complex detecting capabilities. By demonstrating how well tabular and image data can be integrated with advanced algorithms, this work improves the identification of thyroid cancer and increases confidence in machine learning predictions.