The performance-flexible AI-based planning approach using a deep neural network achieved 98.29% sensitivity for predicting ICU admission after elective surgery when prioritizing scarce ICU capacity.
Does a performance-flexible AI-based planning approach improve the prediction of ICU admission and capacity management in patients undergoing elective surgery?
A performance-flexible AI model allows dynamic prioritization of sensitivity or specificity for predicting postoperative ICU admission, optimizing ICU capacity management and reducing overbooking by 30%.
Operating room and intensive care unit (ICU) capacities belong to the scarcest resources in hospitals and strongly depend on each other. When planning elective surgeries, it is therefore important to consider both resources in an integrated way and to guarantee a certain flexibility in planning to avoid under- and overutilization, e.g., in the form of cancellations. In this work, we introduce a performance-flexible artificial intelligence (AI)-based planning approach for predicting whether an elective patient will be transferred to the ICU after elective surgery. This approach includes a performance-flexible loss function in a machine learning (ML) model and a subsequent simulation about ICU occupancy. The algorithm is evaluated by a large data set of the University Hospital of Augsburg, Germany, consisting of more than 26,600 elective surgeries between 2017 and 2021, and extensive simulation studies. This approach is generalizable as it uses data typically available during surgery planning in the outpatient clinic. Our findings demonstrate that, unlike state-of-the-art ML algorithms, our performance-flexible AI-based planning approach can prioritize a specific label in binary classification (i.e., ICU or non-ICU) subject to capacity considerations while maintaining high accuracy. This ensures a stable ratio of realized demand to planned ICU capacity that is close to 1 across different scenarios. Our performance-flexible AI-based planning algorithm outperforms state-of-the-art ML algorithms and supports hospital decision-makers with a flexible planning tool.
Grieger et al. (Wed,) conducted a other in Elective surgery (n=26,677). Performance-flexible AI-based planning approach (NPFBCE loss function) vs. Standard machine learning algorithms (e.g., WBCE, BCET) was evaluated on Sensitivity for predicting ICU admission (with positive class weight α = 0.8). The performance-flexible AI-based planning approach using a deep neural network achieved 98.29% sensitivity for predicting ICU admission after elective surgery when prioritizing scarce ICU capacity.