Subscribe to RSS
Comprehensive Visualization of AI Decisions for Early Complication Detection of Cardiac Surgery Patients
Background: Postoperative care of cardiac surgery patients is complex and often fraught with an increased morbidity and mortality due to potential complications. Even with intensive and continuous monitoring of patients it is often difficult to interpret the early signs of life-threatening complications from the vast amount of monitoring data before actual symptoms occur. AI-based monitoring systems can assist medical personal by predicting incipient complications in an early stage, which gains valuable time for treatment. Although the trend in medical AI systems is going toward the use of explainable AI models, the results and type of explanations vary widely and are hard to understand. We introduce a novel dashboard system that interprets the AI decisions and visualizes them in a uniform and comprehensible way, tailored toward medical personal.
Method: Following a user-centered approach, we captured the system requirements by holding a survey with 25 heart surgeons, as well as conducting interviews with medical personal from different wards. The answers allowed us to identify the necessary data that is needed in critical decision making, when facing possible complications and capture preferred display options of the users. A first prototype dashboard, based on vue.js was then developed and fed with data of existing AI models. The different types of results and explanations of the AI models were analyzed and matching visualizations styles were chosen.
Results: The resulting system uses the data and explanations of different, existing AI prediction models and matches them with suitable visualization styles. The user can then choose the level of detail he prefers, ranging from a simple scoring system to a highly detailed view with precise information about alarming values.
Conclusion: We developed an interactive dashboard system that is capable of visualizing complex AI decisions in a comprehensive way for medical personal. The system can interpret the results of different AI models, matches them with suitable visualization options and displays them to the user uniformly. The user can individualize the dashboard and choose a level of detail that suits his needs. First user tests are currently performed to ensure the comprehensibility of our dashboard; further studies to evaluate and improve the usability and user experience in a clinical setting are planned.
No conflict of interest has been declared by the author(s).
Article published online:
03 February 2022
© 2022. Thieme. All rights reserved.
Georg Thieme Verlag KG
Rüdigerstraße 14, 70469 Stuttgart, Germany