Securing AI- driven telemedicine platforms: vulnerability assessment and resilient architecture design

Telemedicine systems based on AI are used more and more to assist in diagnostics, triage, and remote patient monitoring, although their integration presents additive risks to safety both in the area of cybersecurity, model reliability and clinical decision integrity. Although the traditional concept of healthcare security focuses on the confidentiality of data and system availability, the AI-enabled telemedicine also has to ensure diagnostic accuracy and safety of clinical outcomes that is still unaddressed in current models. Recent research shows that artificially low accuracy in medical imaging models can be caused by adversarial inputs by up to 4065% (Zhou et al., 2024), that medical chatbots in the triage can become more prone to clinically unsafe responses when prompt-injection attacked by up to 30%+ (Li et al., 2023), and that unsecured IoMT channels have false-negative spoofing detection rates above 40% (Moradi et al., 2024). These weaknesses present risks that are likely to directly disrupt the clinical decision-making process, and not necessarily expose patient data. This article determines the existence of vulnerabilities in AI inference pipeline, IoMT telemetry, and telehealth communication operations that are critical and provides a secure-by-design architectural framework that incorporates zero-trust networking, model integrity verification, encrypted biosignal provenance, adversarial input filtering, and clinician-in-the-loop safety restrictions. The framework also complies with the healthcare regulatory standards, such as HIPAA, GDPR, ISO 14971, IEC 62304, and FDA AI/ML medical device recommendations, to facilitate compliant and ready-to-implement deployment to the clinic.