The swift adoption of the Internet of Medical Things (IoMT) into healthcare infrastructure has significantly enhanced patient observation and real-time data analysis capabilities, but it has also exposed these networks to severe cybersecurity threats. Traditional intrusion detection systems (IDS) frequently struggle to address the distinct challenges presented by IoMT environments, especially concerning scalability, class imbalance, and real-time anomaly detection. In response to these challenges, we propose IsoTabNet-IDS, a Transfer Learning-enhanced IDS that leverages TabNet’s attention mechanism and Isolation Forest for robust anomaly detection. Our model integrates RFE and LDA for feature selection, followed by anomaly scoring through Isolation Forest, which is used to refine TabNet’s classification of attack and normal traffic. IsoTabNet-IDS is evaluated on three publicly available highly imbalanced datasets such as WUSTL-EHMS-2020, WUSTL-HRDL-2021 and WUSTL-IIoT-2024, achieving precision scores of 99.9996%, 99.99987% and 99.99993%, respectively. The system consistently surpasses the latest methods of evaluating models like accuracy, F1-score and precision, while maintaining a low latency suitable for real-time deployment. These findings underscore IsoTabNet-IDS’s effectiveness in enhancing the security of IoMT networks, offering a scalable and adaptable approach to safeguard sensitive healthcare data against emerging cyber threats.