Thor Tronrud delves into the challenges and opportunities of integrating AI into medical devices.

Artificial intelligence (AI) has rapidly become a transformative force across industries, but its application in medical devices presents unique challenges. In his article “A Practical Future of Developing AI for Medical Devices” for MD+DI, Thor Tronrud, a Research and Data Analysis-Focused Software Engineer at StarFish Medical, sheds light on bridging knowledge gaps, navigating regulatory constraints, and developing robust AI solutions for medical devices.

Thor Tronrud emphasizes the importance of demystifying AI and addressing the complexities of applying language models to medical devices, particularly within stringent regulatory frameworks such as those outlined by the FDA.

Key takeaways include:

• Understanding Language Models in MedTech: Language models, often used in AI applications, require careful consideration when integrated into medical devices due to their inherent non-deterministic nature, which can introduce risks to patient safety.
• Challenges with Third-Party Models: While pre-trained large language models (LLMs) are cost-effective and widely accessible, they come with significant concerns, including privacy risks, dependence on external services, and the potential for unverified outputs.
• Smaller, Tailored Models: Tronrud suggests that smaller models (150M-500M parameters) are highly effective for tasks like rapid classification and can often be trained on synthetic datasets, generated by larger models, to reduce costs and risks.

Tronrud highlights that AI components in medical devices are subject to rigorous regulatory scrutiny, particularly when their outputs directly impact patient outcomes. A risk-based approach, as mandated by regulatory bodies like the FDA, is essential for navigating these challenges and ensuring compliance. By focusing on smaller, task-specific AI models, developers can create more controlled and predictable solutions that align with regulatory requirements.

To further enhance AI adoption in MedTech, Tronrud underscores the importance of thoughtful design and fine-tuning. For instance, developers can focus on clinician-assisting AI systems that enhance workflows while leaving critical decision-making in the hands of human experts. This balanced approach allows AI to improve efficiency and accuracy without compromising safety or compliance.

Despite the allure of generative AI, Tronrud encourages manufacturers to look towards more mature AI applications, such as computer vision, for insights. These established models demonstrate how AI can be used to complement human expertise and optimize medical workflows, setting a practical standard for the integration of language models in medical devices.