Integrating Interactive Web-Based Modules to Enhance Explainability of a Medical AI Dashboard for Predicting Readmission to Cardiovascular ICUs

Background: Artificial intelligence (AI)-driven decision assistance systems have great potential in the medical field, however they are still considered a black box by most medical experts. Hence, the goal of this project was to provide real-time AI-driven decision assistance for monitoring patients in cardiovascular intensive care units (ICUs) by means of explainable interactive graphical user interface (GUI) modules. To this end, we developed and analyzed different web-based modules to enhance the interpretability of the decisions made by the AI model, given the example of readmission prediction.

Method: As part of a medical AI dashboard, two alternative web-based modules visualize the data from a diverse cohort of patient stays, providing further information through the use of clickable UI elements and tooltips in a dynamic and interactive manner. The modules visualize the predictions of a pre-trained AI model, suggesting whether the patient can be discharged from the ICU without risk for readmission at the current time or in the upcoming 24-hour window, with 8-hour intervals. In addition, the modules quantify the confidence level associated with the AI model's decision as well as the most important features and their relevance used to come up with the corresponding decision.

Results: The first module uses a color-coded icon indicating an AI-model's decision, e. g., whether it would be safe to discharge the patient at current time (green), or not (red). It is surrounded by a clickable pie-chart, visualizing the most relevant input features leading to the decision in the current time, as well as in the next 8, 16, and 24 hours, using colors to indicate if the variables are in the normal range. The second module comprises a 5 by 4 grid, with the first row quantifying the probability assigned to an event of interest (e.g., sepsis) in the above-mentioned time stamps. The rest of the rows represent the four most relevant features and the columns represent time stamps respectively. The size of each element in the grid encodes its relevance to AI model's decision and the colors indicate how far the variables are from the normal range.

Conclusion: We have provided two alternative web-based visualizations to improve the interpretability of an AI-driven patient monitoring dashboard intended to be used in cardiovascular ICUs. In the future, an empirical user study will be conducted to evaluate the usability of the two modules. Nevertheless, both modules can be customized and used in parallel for different use cases such as readmission prediction and early detection of sepsis.

Publikationsverlauf

Artikel online veröffentlicht:
28. Januar 2023