The integration of IoT devices in healthcare introduces significant reliability challenges, necessitating robust anomaly detection mechanisms to ensure continuous and accurate operation. This study proposes a machine learning-driven framework for detecting faulty device anomalies, leveraging a dataset of 200,000 records. Four machine learning models are evaluated across three methodological paradigms: supervised learning (XGBoost, K-Nearest Neighbors (KNN)), semi-supervised learning (Generative Adversarial Networks (GAN)), and unsupervised learning (Isolation Forest). Performance assessment is conducted using multiple metrics, including accuracy, F1-score, precision, recall, Receiver Operating Characteristic–Area Under the Curve (ROC-AUC), and computational efficiency. Experimental results indicate that XGBoost achieves the highest accuracy (99%) with minimal computational overhead (0.04 s), making it the most efficient model for real-time fault detection. Isolation Forest demonstrates a strong balance between precision and recall. These findings provide critical insights into optimizing fault detection strategies, ensuring the reliability of IoT-enabled medical devices. This research contributes to the Software Engineering for the Internet of Things (IoT) by providing a machine learning-driven anomaly detection framework tailored for IoT-based healthcare environments. Furthermore, it supports software development by integrating diverse learning paradigms into software systems, facilitating real-time fault detection in medical IoT devices. By enabling early identification of operational anomalies, this framework enhances system resilience, minimizes device downtime, and supports the safe and continuous operation of healthcare systems.