Developing Machine-Learning Algorithms for Real-time Prediction of Hypotensive Events in ICU Settings
Human organs require a relatively high and constant arterial blood pressure (ABP) to ensure adequate perfusion. A prolonged period of low blood pressure (BP) is a dysfunction of the cardiovascular system called arterial hypotension and frequently occurs in intensive care units or operating rooms where the patient is under general anesthesia. Hypotension episodes are associated with a higher mortality rate and organ failure such as acute kidney injury or myocardial injury over the next few days. Predicting hypotensive events in advance provides clinical staff with enough time to reduce mortality and morbidity rate with proper therapeutic measures such as administering vasopressors.
In this dissertation, we propose a machine-learning algorithm that predicts hypotensive events in ICU settings with high accuracy, specificity and positive predictive value (PPV). At its heart, the algorithm utilizes a novel labeling approach that is capable of real-time monitoring of patients’ physiological records. Several supervised classification algorithms are trained, and the hyperparameters are tuned to optimize the F1-score. A high-level decision-making scenario is also introduced and applied to initial prediction labels to further improve the algorithm’s PPV by filtering out the isolated false alarms. This reduces the chance of alarm fatigue among clinical staff and increases their confidence in the alarms sounded by the algorithm.
Further, we expand the application of our algorithm to the cases where invasive measurement of the BP signals is not a possibility. Our study reveals that the information from the BP signals is essential for the proposed algorithm to predict hypotensive events properly. Hence, we provided the algorithm with the simulated intermittent noninvasively measured mean arterial pressure (NIMAP) by downsampling the invasive MAP. It is shown that hypotensive events can be predicted effectively by using simulated noninvasive BP-related and non-BP physiological features. This opens the door for using the information from readily available devices such as automatic cuff inflation devices along with pulse oximeters, and ECG.
Lastly, we investigated the effect of extending our feature set, by adding patients’ contextual data including demographics, pre-existing comorbidities, and history of medications, on the predictive performance of the algorithm. The algorithm was trained with various subsets of contextual and physiological features and the results are analyzed. We finally performed a feature importance analysis to find the most and least predictive features.