Optimizing binary classification algorithms through hyperparameter tuning: A case study on thyroid cancer prediction
Computers and Industrial Engineering, cilt.219, 2026 (SCI-Expanded, Scopus)
- Yayın Türü: Makale / Tam Makale
- Cilt numarası: 219
- Basım Tarihi: 2026
- Doi Numarası: 10.1016/j.cie.2026.112225
- Dergi Adı: Computers and Industrial Engineering
- Derginin Tarandığı İndeksler: Science Citation Index Expanded (SCI-EXPANDED), Scopus, ABI/INFORM, Aerospace Database, Applied Science & Technology Source, Compendex, INSPEC, DIALNET, Business Source Ultimate (EBSCO), Engineering Source (EBSCO), Technology Collection (ProQuest)
- Anahtar Kelimeler: Binary classification, Hyperparameter tuning, Logistic regression, Random forest, Randomized search, Thyroid cancer
- İstanbul Gelişim Üniversitesi Adresli: Evet
Özet
Diagnosing thyroid cancer requires reliable and interpretable machine learning models that effectively reduce false negatives in clinical decision-making. This study carefully examines how hyperparameter optimization impacts binary classification algorithms using a real-world dataset from 234 patients, including demographic, ultrasonographic, cytological, and laboratory data. Five popular machine learning models—Logistic Regression, Support Vector Machine (SVM), K-Nearest Neighbors (KNN), Random Forest, and Gradient Boosting—were tested through nested cross-validation with Random Search and Bayesian Optimization. The results show that hyperparameter tuning can lead to clinically relevant improvements, although these vary by model. The largest improvement in ROC-AUC was seen in Gradient Boosting (from 0.765 to 0.786, Δ=+0.021). The Bayesian-optimized Random Forest outperformed the other models with ROC-AUC = 0.823, PR-AUC = 0.781, Recall = 0.722, F1 = 0.710, balanced accuracy = 0.764, and MCC = 0.525. Compared to the default Random Forest, recall increased from 0.589 to 0.722, adding about 13 % sensitivity—crucial for lowering false negatives in thyroid cancer diagnosis. SHAP-based explainability and ablation studies assessed feature importance and model interpretability, showing ultrasonography and cytology data as the most influential, while laboratory features alone had limited discriminative power. Overall, this framework provides a more transparent, reproducible, and interpretable machine learning approach for thyroid cancer prediction, illustrating the importance of hyperparameter tuning in medical classification.