Paediatric Intensive Care at UMC Utrecht Improvement and testing of a Clinical Deterioration Model for the Continuous Data-Driven Monitoring in Critical Congenital Heart Disease

Abstract

[""Introduction: Continuous monitoring of vital signs is crucial in paediatric intensive care units (PICUs) to detect clinical deterioration. This study focuses on critically ill paediatric patients with congenital heart disease (CHD) who underwent cardiac interventions in the first year of life. To help medical practitioners with the big amount of monitoring data, a data-driven model for automated detection of clinical deterioration and instability was developed. This model used the Maahlanobis distance of 5 vital parameters combined and a one-class support vector machine model. The goal of this study is to improve the existing machine learning algorithm through a) broadening statistical foundation for data-based cutoff values, b) exploring an alternative method for splitting stable from unstable timepoints (lactate) and c) creating new baselines for the comparison of the Mahalanobis distance.

Methods: For the broadening statistical foundation for the data-based cutoff values (80th percentile) a Local Outlier Factor methodology was used. The resulting outlier scores and k-distances were placed in density graphs to locate the start of the outliers. With the new percentile, a new SVM model was developed. The alternative method for splitting stable from unstable timepoints was lactate measurements. Lactate levels in blood are commonly used to determine if a patient is unstable. Thereafter measurements < 2 mmol/L were labeled as stable and fed into a new SVM model. The new baselines were created using the weighted moving average of the medians of the Mahalanobis measurements. The modified models were evaluated by a physician using visualizations that compared the models results with the vital parameters of new patients.

Results: The modified models showed good evaluation metrics, except for specificity. They accurately identified stable periods but not unstable periods. The use of lactate measurements and smoothing the Mahalanobis baseline improved the model's performance, especially in terms of specificity. However, it is suspected that not all unstable periods are outliers, which would explain the low specificity.

Conclusion: The modified models provide accurate detection of stable periods. The models could be used to schedule interventions, minimize disturbances during rest periods and reduce alarm fatigue. Further improvements can be made by incorporating additional data sources, and refining the classification of unstable periods. Future research should focus on real-time validation and testing in a clinical setting to optimize the model's performance and usability. ""]