AI-based Risk Stratification for ECMO Therapy in the ICU after Elective Pediatric Cardiac Surgery

Background: About 3% of patients undergoing pediatric cardiac surgery with complex cases may require ECMO therapy during their stay in the ICU to prevent cardiac arrest. This applies to patients with different congenital heart diseases (CHDs), surgical approaches, and ages when receiving surgery. EMCO therapy correlates with higher mortality, increasing with the duration of therapy. Accordingly, the avoidance of ECMO therapy is of central importance to patient outcomes. This study aims to identify an AI-based prediction of ECMO-necessitating hemodynamic states before becoming clinically apparent.

Methods: Peri- and postoperative data from the electronic health record (EHR) of around 900 patients after cardiac surgery has been collected between 2020 and 2023. Ethical approval has been obtained to extend the number of patients to a total of approximately 2,500 patients from 2023 to 2027. The data were normalized for comparison and inclusion of different CHDs and respective hemodynamics. Welch’s t-tests have been used to test 30-minute windows of the vital parameters at the start of ECMO therapy against 30-minute windows of earlier more stable hemodynamic states to screen for pathologic patterns.

Results: The collected data reflect the physiology of beginning circulatory failure and shock. When visualized, the vital parameters of ECMO patients can be clearly distinguished several hours before the beginning of ECMO therapy. This is also statistically significant when analyzing just the ECMO patients over time. The mean systolic blood pressure 3 hours before the start of ECMO therapy significantly differs from the mean systolic blood pressure at the beginning of ECMO therapy (p = 0.04). The maximum heart frequency differs 4.5 hours before (p = 0,03) and the respiratory rate is significantly different 12 hours before the beginning of ECMO therapy (p = 0.02). The data are normally distributed.

Conclusion: The pattern of basic vital parameters of patients with ECMO therapy shows statistically significant differences over time before the start of EMCO therapy. While this might be generally expected and reflect the clinical presentation, the onset of statistical features seems to precede clinical presentation. As such, they might be useful indicators for AI-based classification. In addition, the vital parameters have so far been analyzed individually. It seems reasonable to expect synergistic effects in a comprehensive analysis and classification. The dataset, however, suffers from imbalance. This will be addressed by the separation of the dataset along CHDs and respective hemodynamics.

Publication History

Article published online:
11 February 2025