Real-Time Machine Learning for ICU Hypoxia Prediction: A Pilot Study

Abstract: Continuous and precise monitoring in the intensive care unit (ICU) are essential, particularly for patients with respiratory disorders necessitating mechanical ventilation, as they represent a critical cohort. Hypoxia can be defined as a medical condition characterized by an inadequate supply of oxygen to bodily tissues and organs, leading to an inability to meet their metabolic needs. This condition can manifest in diverse scenarios, such as high-altitude environments for example mountain climbing or in high-altitude flights. Common symptoms of hypoxia include breathing difficulties, confusion, elevated heart rate, and impaired cognitive and physical functions. If left unaddressed, hypoxia can lead to tissue damage, organ failure, and, in severe cases, even death. Traditionally, hypoxia detection relies on post facto measures, with methods implying peripheral capillary oxygen saturation (SpO2) providing valuable but delayed insights. Models providing real-time levels of hypoxia, or even early detection and intervention, would thus be relevant to prevent such a state. They could in fact provide timely detection to trigger automatic ventilation or to alert healthcare personnel promptly via adaptive automation. The goal of this study was to produce a real-time hypoxia detection model using machine-learning techniques in the context of ICU. Methods. We used the open-source eICU database, consisting of critically ill patients treated in ICU across the United States. It contains vital signs recorded at 5-min intervals and many hypoxia events. We utilized parameters according to a feature selection such as SpO2, heart rate (HR), and respiratory rate (RR) to make hypoxia predictions from few prior lags. We selected three hypoxia levels to achieve a supervised learning classification for hypoxia: No Hypoxia (SpO2 > 93%), Low level of Hypoxia (88% < SpO2 < 93%), and Strong level of Hypoxia (SpO2 < 88%). Furthermore, we exported the trained model to the Open Neural Network Exchange (ONNX) format to facilitate real-time predictions deployment in clinical settings, thus offering a valuable tool for early hypoxia detection and proactive patient care in ICU.Results. Mann-Whitney U tests and paired t-tests raised significant differences across levels of hypoxia for the SpO2, HR and RR measures. We tested several machine-learning models and our results showed that a Random Forest model could provide accurate predictions of impending hypoxia events (5 minutes prior to the event). We performed a random search fine-tuning method and a group 5-fold cross-validation and achieved a predictive accuracy of 0.937, a precision rate of 0.85, and a balanced accuracy of 0.813. The ONNX implementation of the model had an inference of 1 millisecond, which allows real-time hypoxia prediction. Discussion. Hypoxia is frequently encountered in the ICU as a result of a wide range of pathologic characteristics. Therefore, detecting hypoxia events before they occur is of paramount importance to ensure timely countermeasures through adaptive automation and prevent fatal outcomes for patients. The use of an ONNX model further enables real-time predictions, enhancing the clinical utility of our approach. The versatility of this approach extends beyond the medical domain, for example in aviation where hypoxia poses a serious threat. Such models have potential for being integrated in real time for pinpointing instances of pilots hypoxia, thus contributing to their health and to overall flight safety.

Keywords: AI, Hypoxia, Machine Learning, Real time prediction, Medical, Human Factors

DOI: 10.54941/ahfe1004734

Cite this paper:

​