User-Centered Design of AI-Assisted Clinical ECG Interpretation Systems

Sammanfattning

This thesis investigates how Artificial Intelligence (AI) models can be integrated into medical software to support healthcare professionals in diagnosing the type and location of myocardial infarctions (MI) from electrocardiogram (ECG) data. Using a user-centered, iterative design process, the work was carried out in collaboration with physicians and nurses in an emergency department who interact with ECG recordings as part of their daily tasks. Through contextual inquiry and the development of user interface prototypes, followed by implementation in an application, the goal was to design a usable graphical interface for the machine-learning model AI-OMI. Key design decisions were made in accordance with established theories in human-computer interaction (HCI) and usability engineering. The final design uses multiple visualization methods to present the AI model's outputs alongside traditional ECG graphs. The results indicate that aligning the AI model with established clinical workflows minimizes mental load and facilitates the adoption of new technology in critical care environments. However, conflicts can arise when an AI model's prediction contradicts a physician's conventional medical training, leading to distrust of AI. The approach of explainable AI is considered and discussed as a method to avoid conflicts.