Evaluating predictions of critical oxygen desaturation events

Predicting Hypoxia Using Machine Learning: Systematic Review

Background

Hypoxia is an important risk factor and indicator for the declining health of inpatients. Predicting future hypoxic events using machine learning is a prospective area of study to facilitate time-critical interventions to counter patient health deterioration.

Objective

This systematic review aims to summarize and compare previous efforts to predict hypoxic events in the hospital setting using machine learning with respect to their methodology, predictive performance, and assessed population.

Methods

A systematic literature search was performed using Web of Science, Ovid with Embase and MEDLINE, and Google Scholar. Studies that investigated hypoxia or hypoxemia of hospitalized patients using machine learning models were considered. Risk of bias was assessed using the Prediction Model Risk of Bias Assessment Tool.

Results

After screening, a total of 12 papers were eligible for analysis, from which 32 models were extracted. The included studies showed a variety of population, methodology, and outcome definition. Comparability was further limited due to unclear or high risk of bias for most studies (10/12, 83%). The overall predictive performance ranged from moderate to high. Based on classification metrics, deep learning models performed similar to or outperformed conventional machine learning models within the same studies. Models using only prior peripheral oxygen saturation as a clinical variable showed better performance than models based on multiple variables, with most of these studies (2/3, 67%) using a long short-term memory algorithm.

Conclusions

Machine learning models provide the potential to accurately predict the occurrence of hypoxic events based on retrospective data. The heterogeneity of the studies and limited generalizability of their results highlight the need for further validation studies to assess their predictive performance.

Innovative Predictive Approach towards a Personalized Oxygen Dosing System

Despite the large impact chronic obstructive pulmonary disease (COPD) that has on the population, the implementation of new technologies for diagnosis and treatment remains limited. Current practices in ambulatory oxygen therapy used in COPD rely on fixed doses overlooking the diverse activities which patients engage in. To address this challenge, we propose a software architecture aimed at delivering patient-personalized edge-based artificial intelligence (AI)-assisted models that are built upon data collected from patients' previous experiences along with an evaluation function. The main objectives reside in proactively administering precise oxygen dosages in real time to the patient (the edge), leveraging individual patient data, previous experiences, and actual activity levels, thereby representing a substantial advancement over conventional oxygen dosing. Through a pilot test using vital sign data from a cohort of five patients, the limitations of a one-size-fits-all approach are demonstrated, thus highlighting the need for personalized treatment strategies. This study underscores the importance of adopting advanced technological approaches for ambulatory oxygen therapy.

SWIFT: A deep learning approach to prediction of hypoxemic events in critically-Ill patients using SpO2 waveform prediction

Hypoxemia is a significant driver of mortality and poor clinical outcomes in conditions such as brain injury and cardiac arrest in critically ill patients, including COVID-19 patients. Given the host of negative clinical outcomes attributed to hypoxemia, identifying patients likely to experience hypoxemia would offer valuable opportunities for early and thus more effective intervention. We present SWIFT (SpO2 Waveform ICU Forecasting Technique), a deep learning model that predicts blood oxygen saturation (SpO2) waveforms 5 and 30 minutes in the future using only prior SpO2 values as inputs. When tested on novel data, SWIFT predicts more than 80% and 60% of hypoxemic events in critically ill and COVID-19 patients, respectively. SWIFT also predicts SpO2 waveforms with average MSE below .0007. SWIFT predicts both occurrence and magnitude of potential hypoxemic events 30 minutes in the future, allowing it to be used to inform clinical interventions, patient triaging, and optimal resource allocation. SWIFT may be used in clinical decision support systems to inform the management of critically ill patients during the COVID-19 pandemic and beyond.

SWIFT: A Deep Learning Approach to Prediction of Hypoxemic Events in Critically-Ill Patients Using SpO2 Waveform Prediction.. [PMID: 33688661 PMCID: PMC7941630 DOI: 10.1101/2021.02.25.21252234] [Citation(s) in RCA: 0] [Impact Index Per Article: 0]

Hypoxemia is a significant driver of mortality and poor clinical outcomes in conditions such as brain injury and cardiac arrest in critically ill patients, including COVID-19 patients. Given the host of negative clinical outcomes attributed to hypoxemia, identifying patients likely to experience hypoxemia would offer valuable opportunities for early and thus more effective intervention. We present SWIFT (SpO₂Waveform ICU Forecasting Technique), a deep learning model that predicts blood oxygen saturation (SpO₂) waveforms 5 and 30 minutes in the future using only prior SpO₂ values as inputs. When tested on novel data, SWIFT predicts more than 80% and 60% of hypoxemic events in critically ill and COVID-19 patients, respectively. SWIFT also predicts SpO₂ waveforms with average MSE below .0007. SWIFT provides information on both occurrence and magnitude of potential hypoxemic events 30 minutes in advance, allowing it to be used to inform clinical interventions, patient triaging, and optimal resource allocation. SWIFT may be used in clinical decision support systems to inform the management of critically ill patients during the COVID-19 pandemic and beyond.

Gaussian Mixture Models for Detecting Sleep Apnea Events Using Single Oronasal Airflow Record

Sleep apnea is a common sleep-related disorder that significantly affects the population. It is characterized by repeated breathing interruption during sleep. Such events can induce hypoxia, which is a risk factor for multiple cardiovascular and cerebrovascular diseases. Polysomnography, the gold standard, is expensive, inaccessible, uncomfortable and an expert technician is needed to score sleep-related events. To address these limitations, many previous studies have proposed and implemented automatic scoring processes based on fewer sensors and machine learning classification algorithms. However, alternative device technologies developed for both home and hospital still have limited diagnostic accuracy for detecting apnea events even though many of the previous investigational algorithms are based on multiple physiological channel inputs. In this paper, we propose a new probabilistic algorithm based on (only) oronasal respiration signal for automated detection of apnea events during sleep. The proposed model leverages AASM recommendations for characterizing apnea events with respect to dynamic changes in the local respiratory airflow baseline. Unlike classical threshold-based classification models, we use a Gaussian mixture probability model for detecting sleep apnea based on the posterior probabilities of the respective events. Our results show significant improvement in the ability to detect sleep apnea events compared to a rule-based classifier that uses the same classification features and also compared to two previously published studies for automated apnea detection using the same respiratory flow signal. We use 96 sleep patients with different apnea severity levels as reflected by their Apnea-Hypopnea Index (AHI) levels. The performance was not only analyzed over obstructive sleep apnea (OSA) but also over other types of sleep apnea events including central and mixed sleep apnea (CSA, MSA). Also the performance was comprehensively analyzed and evaluated over patients with varying disease severity conditions, where it achieved an overall performance of TPR=88.5%, TNR=82.5%, and AUC=86.7%. The proposed approach contributes a new probabilistic framework for detecting sleep apnea events using a single airflow record with an improved capability to generalize over different apnea severity conditions

Predictive Monitoring of Critical Cardiorespiratory Alarms in Neonates Under Intensive Care

We aimed at reducing alarm fatigue in neonatal intensive care units by developing a model using machine learning for the early prediction of critical cardiorespiratory alarms. During this study in over 34,000 patient monitoring hours in 55 infants 278,000 advisory (yellow) and 70,000 critical (red) alarms occurred. Vital signs including the heart rate, breathing rate, and oxygen saturation were obtained at a sampling frequency of 1 Hz while heart rate variability was calculated by processing the ECG - both were used for feature development and for predicting alarms. Yellow alarms that were followed by at least one red alarm within a short post-alarm window constituted the case-cohort while the remaining yellow alarms constituted the control cohort. For analysis, the case and control cohorts, stratified by proportion, were split into training (80%) and test sets (20%). Classifiers based on decision trees were used to predict, at the moment the yellow alarm occurred, whether a red alarm(s) would shortly follow. The best performing classifier used data from the 2-min window before the occurrence of the yellow alarm and could predict 26% of the red alarms in advance (18.4s, median), at the expense of 7% additional red alarms. These results indicate that based on predictive monitoring of critical alarms, nurses can be provided a longer window of opportunity for preemptive clinical action. Further, such as algorithm can be safely implemented as alarms that are not algorithmically predicted can still be generated upon the usual breach of the threshold, as in current clinical practice.

Characteristics of Desaturation and Respiratory Rate in Postoperative Patients Breathing Room Air Versus Supplemental Oxygen: Are They Different?

BACKGROUND

Routine monitoring of postoperative patients with pulse oximetry-based surveillance monitoring has been shown to reduce adverse events. However, there is some concern that pulse oximetry is limited in its ability to detect deterioration quickly enough to allow for intervention in patients receiving supplemental oxygen. To address such concerns, this study expands on the current limited knowledge of differences in desaturation and respiratory rate characteristics between patients breathing room air and those receiving supplemental oxygen.

METHODS

Pulse oximetry-derived data and patient characteristics were used to examine overnight desaturation patterns of 67 postoperative patients who were receiving either supplemental oxygen or breathing room air. The 2 modalities with respect to the speed of desaturation, in addition to magnitude and duration of desaturation events, are compared. Night-time pulse rate, oxygen saturation, respiratory rate, and the transition times from normal oxygen saturation levels to desaturated states are also compared. The behavior of respiratory rate in proximity to desaturation events is described. Statistical methods included multivariable regression and inverse probability of treatment weighted to adjust for any imbalance in patient characteristics between the oxygen and room air patients and linear mixed effect models to account for clustering by patient.

RESULTS

The study included 33 patients on room air and 34 receiving supplemental oxygen. The speed of desaturation was not different for room air versus oxygen for 2 types of desaturation (adjusted % difference, 95% confidence interval [CI]: type I; 22.4%, -51.5% to 209%; P = .67, type II; -17.3%, -53.8% to 47.6%; P = .52). Patients receiving supplemental oxygen had a higher mean oxygen saturation (adjusted difference, 95% CI, 2.4 [0.7-4.0]; P = .006). No differences were found for the average overnight respiratory or pulse rate, or proportion of time in desaturation states between the 2 groups.The time to transition from a normal oxygen saturation (92%) to 88% or below was not longer for supplemental oxygen patients (P = .42, adjusted difference 26.1%: 95% CI, -28.1% to 121%). Respiratory rates did not differ between the overall mean and desaturation or recovery phases or between the oxygen and room air group.

CONCLUSIONS

In this study, desaturation characteristics did not differ between patients receiving supplemental oxygen and breathing room air with regard to speed, depth, or duration of desaturation. Transition time for desaturations to reach low oxygen saturation alarms was not different, while respiratory rate remained in the normal range during these events. These findings suggest that pulse oximetry-based surveillance monitoring for deterioration detection can be used equally effectively for patients on supplemental oxygen and for those on room air.

Multi-Step Ahead Predictions for Critical Levels in Physiological Time Series

Standard modeling and evaluation methods have been classically used in analyzing engineering dynamical systems where the fundamental problem is to minimize the (mean) error between the real and predicted systems. Although these methods have been applied to multi-step ahead predictions of physiological signals, it is often more important to predict clinically relevant events than just to match these signals. Adverse clinical events, which occur after a physiological signal breaches a clinically defined critical threshold, are a popular class of such events. This paper presents a framework for multi-step ahead predictions of critical levels of abnormality in physiological signals. First, a performance metric is presented for evaluating multi-step ahead predictions. Then, this metric is used to identify personalized models optimized with respect to predictions of critical levels of abnormality. To address the paucity of adverse events, weighted support vector machines and cost-sensitive learning are used to optimize the proposed framework with respect to statistical metrics that can take into account the relative rarity of such events.

Journal of Clinical Monitoring and Computing 2015 end of year summary: respiration

This paper reviews 17 papers or commentaries published in Journal of Clinical Monitoring and Computing in 2015, within the field of respiration. Papers were published covering monitoring and training of breathing, monitoring of gas exchange, hypoxemia and acid-base, and CO2 monitoring.

Effect of concurrent oxygen therapy on accuracy of forecasting imminent postoperative desaturation

Episodic postoperative desaturation occurs predominantly from respiratory depression or airway obstruction. Monitor display of desaturation is typically delayed by over 30 s after these dynamic inciting events, due to perfusion delays, signal capture and averaging. Prediction of imminent critical desaturation could aid development of dynamic high-fidelity response systems that reduce or prevent the inciting event from occurring. Oxygen therapy is known to influence the depth and duration of desaturation epochs, thereby potentially influencing the accuracy of forecasting of desaturation. In this study, postoperative pulse oximetry data were retrospectively modeled using autoregressive methods to create prediction models for [Formula: see text] and imminent critical desaturation in the postoperative period. The accuracy of these models in predicting near future [Formula: see text] values was tested using root mean square error. The model accuracy for prediction of critical desaturation ([Formula: see text] [Formula: see text]) was evaluated using meta-analytical methods (sensitivity, specificity, likelihood ratios, diagnostic odds ratios and area under summary receiver operating characteristic curves). Between-study heterogeneity was used as a measure of reliability of the model across different patients and evaluated using the tau-squared statistic. Model performance was evaluated in [Formula: see text] patients who received postoperative oxygen supplementation and [Formula: see text] patients who did not receive oxygen. Our results show that model accuracy was high with root mean square errors between 0.2 and 2.8%. Prediction accuracy as defined by area under the curve for critical desaturation events was observed to be greater in patients receiving oxygen in the 60-s horizon ([Formula: see text] vs. [Formula: see text]). This was likely related to the higher frequency of events in this group (median [IQR] [Formula: see text] [Formula: see text]) than patients who were not treated with oxygen ([Formula: see text] [Formula: see text]; [Formula: see text]). Model reliability was reflected by the homogeneity of the prediction models which were homogenous across both prediction horizons and oxygen treatment groups. In conclusion, we report the use of autoregressive models to predict [Formula: see text] and forecast imminent critical desaturation events in the postoperative period with high degree of accuracy. These models reliably predict critical desaturation in patients receiving supplemental oxygen therapy. While high-fidelity prophylactic interventions that could modify these inciting events are in development, our current study offers proof of concept that the afferent limb of such a system can be modeled with a high degree of accuracy.