Ouattassi Naouar, Zoizou Abdelhay, Maaroufi Mustapha, Labiyad El Mehdi, Benatiya Andaloussi Taha, Benmansour Najib, Ridal Mohammed, Zarghili Arsalane, El Amine El Alami Mohamed Nouredine
Biomedical and Translational Research Laboratory, Faculty of Medicine, Pharmacy and Dentistry, University Sidi Mohamed Ben Abdellah, Fez, MAR.
Otolaryngology - Head and Neck Surgery Department, Hassan II University Hospital, Fez, MAR.
Cureus. 2025 May 5;17(5):e83503. doi: 10.7759/cureus.83503. eCollection 2025 May.
To evaluate the diagnostic accuracy of middle ear partial filling on temporal bone computed tomography (CT) scan for middle ear cholesteatoma identification using supervised machine learning models.
We conducted an observational case-control study that retrospectively analyzed temporal bone CT scans from 212 patients from a single tertiary healthcare institution using supervised machine learning models, including k-Nearest Neighbors (kNN), Neural Networks, Logistic Regression, Support Vector Machine (SVM), and Random Forest. The study assessed the diagnostic value of partial middle ear filling for cholesteatoma. Limitations such as dataset imbalance and data complexity were addressed. Results. In internal validation, kNN and Neural Networks achieved the highest performance (area under the receiver operating characteristic curve [AUC]: 1.000, classification accuracy [CA]: 99.6-99.7%, F1: 0.996-0.997), followed by Logistic Regression (AUC: 0.998, CA: 98.3%, F1: 0.983) and SVM (AUC: 0.997, CA: 97.5%, F1: 0.975). Random Forest performed the weakest (AUC: 0.980, CA: 92.0%, F1: 0.919). External validation (125 cases) revealed Neural Networks' superior generalizability (four errors), outperforming Logistic Regression (five), SVM (seven), Random Forest (28), and kNN (45). kNN demonstrated notably lower generalizability, suggesting limited robustness for unseen data. Discussion. The study highlights the effectiveness of machine learning in diagnosing cholesteatoma. Addressing data imbalance and variability in CT scans was crucial for model performance. Further research is needed to refine these models and explore their integration into clinical practice.
使用监督式机器学习模型评估颞骨计算机断层扫描（CT）上中耳部分填充对中耳胆脂瘤识别的诊断准确性。
我们进行了一项观察性病例对照研究，使用监督式机器学习模型，包括k近邻算法（kNN）、神经网络、逻辑回归、支持向量机（SVM）和随机森林，对来自一家三级医疗机构的212例患者的颞骨CT扫描进行回顾性分析。该研究评估了中耳部分填充对胆脂瘤的诊断价值。解决了数据集不平衡和数据复杂性等局限性问题。结果。在内部验证中，kNN和神经网络表现最佳（受试者工作特征曲线下面积[AUC]：1.000，分类准确率[CA]：99.6 - 99.7%，F1值：0.996 - 0.997），其次是逻辑回归（AUC：0.998，CA：98.3%，F1值：0.983）和支持向量机（AUC：0.997，CA：97.5%，F1值：0.975）。随机森林表现最差（AUC：0.980，CA：92.0%，F1值：0.919）。外部验证（125例）显示神经网络具有卓越的泛化能力（4个错误），优于逻辑回归（5个）、支持向量机（7个）、随机森林（28个）和kNN（45个）。kNN的泛化能力明显较低，表明对未见数据的稳健性有限。讨论。该研究突出了机器学习在诊断胆脂瘤方面的有效性。解决CT扫描中的数据不平衡和变异性对模型性能至关重要。需要进一步研究来优化这些模型，并探索将其整合到临床实践中。
