Semi-supervised detection of intracranial pressure alarms using waveform dynamics

Developing a Scalable Annotation Method for Large Datasets That Enhances Alarms With Actionability Data to Increase Informativeness: Mixed Methods Approach

BACKGROUND

Alarm fatigue, a multifactorial desensitization of staff to alarms, can harm both patients and health care staff in intensive care units (ICUs), especially due to false and nonactionable alarms. Increasing amounts of routinely collected alarm and ICU patient data are paving the way for training machine learning (ML) models that may help reduce the number of nonactionable alarms, potentially increasing alarm informativeness and reducing alarm fatigue. At present, however, there is no publicly available dataset or process that routinely collects information on alarm actionability (ie, whether an alarm triggers a medical intervention or not), which is a key feature for developing meaningful ML models for alarm management. Furthermore, case-based manual annotation is too slow and resource intensive for large amounts of data.

OBJECTIVE

We propose a scalable method to annotate patient monitoring alarms associated with patient-related variables regarding their actionability. While the method is aimed to be used primarily in our institution, other clinicians, scientists, and industry stakeholders could reuse it to build their own datasets.

METHODS

The interdisciplinary research team followed a mixed methods approach to develop the annotation method, using data-driven, qualitative, and empirical strategies. The iterative process consisted of six steps: (1) defining alarm terms; (2) reaching a consensus on an annotation concept and documentation structure; (3) defining physiological alarm conditions, related medical interventions, and time windows to assess; (4) developing mapping tables; (5) creating the annotation rule set; and (6) evaluating the generated content. All decisions were made based on feasibility criteria, clinical relevance, occurrence frequency, data availability and quantity, structure, and storage mode. The annotation guideline development process was preceded by the analysis of the institution's data and systems, the evaluation of device manuals, and a systematic literature review.

RESULTS

In a multidisciplinary consensus-based approach, we defined preprocessing steps and a rule-based annotation method to classify alarms as either actionable or nonactionable based on data from the patient data management system. We have presented our experience in developing the annotation method and provided the generated resources. The method focuses on respiratory and medication management interventions and includes 8 general rules in a tabular format that are accompanied by graphical examples. Mapping tables enable handling unstructured information and are referenced in the annotation rule set.

CONCLUSIONS

Our annotation method will enable a large number of alarms to be labeled semiautomatically, retrospectively, and quickly, and will provide information on their actionability based on further patient data. This will make it possible to generate annotated datasets for ML models in alarm management and alarm fatigue research. We believe that our annotation method and the resources provided are universal enough and could be used by others to prepare data for future ML projects, even beyond the topic of alarms.

Impact of Label Noise on the Learning Based Models for a Binary Classification of Physiological Signal

Label noise is omnipresent in the annotations process and has an impact on supervised learning algorithms. This work focuses on the impact of label noise on the performance of learning models by examining the effect of random and class-dependent label noise on a binary classification task: quality assessment for photoplethysmography (PPG). PPG signal is used to detect physiological changes and its quality can have a significant impact on the subsequent tasks, which makes PPG quality assessment a particularly good target for examining the impact of label noise in the field of biomedicine. Random and class-dependent label noise was introduced separately into the training set to emulate the errors associated with fatigue and bias in labeling data samples. We also tested different representations of the PPG, including features defined by domain experts, 1D raw signal and 2D image. Three different classifiers are tested on the noisy training data, including support vector machine (SVM), XGBoost, 1D Resnet and 2D Resnet, which handle three representations, respectively. The results showed that the two deep learning models were more robust than the two traditional machine learning models for both the random and class-dependent label noise. From the representation perspective, the 2D image shows better robustness compared to the 1D raw signal. The logits from three classifiers are also analyzed, the predicted probabilities intend to be more dispersed when more label noise is introduced. From this work, we investigated various factors related to label noise, including representations, label noise type, and data imbalance, which can be a good guidebook for designing more robust methods for label noise in future work.

Computational approaches to alleviate alarm fatigue in intensive care medicine: A systematic literature review

Patient monitoring technology has been used to guide therapy and alert staff when a vital sign leaves a predefined range in the intensive care unit (ICU) for decades. However, large amounts of technically false or clinically irrelevant alarms provoke alarm fatigue in staff leading to desensitisation towards critical alarms. With this systematic review, we are following the Preferred Reporting Items for Systematic Reviews (PRISMA) checklist in order to summarise scientific efforts that aimed to develop IT systems to reduce alarm fatigue in ICUs. 69 peer-reviewed publications were included. The majority of publications targeted the avoidance of technically false alarms, while the remainder focused on prediction of patient deterioration or alarm presentation. The investigated alarm types were mostly associated with heart rate or arrhythmia, followed by arterial blood pressure, oxygen saturation, and respiratory rate. Most publications focused on the development of software solutions, some on wearables, smartphones, or headmounted displays for delivering alarms to staff. The most commonly used statistical models were tree-based. In conclusion, we found strong evidence that alarm fatigue can be alleviated by IT-based solutions. However, future efforts should focus more on the avoidance of clinically non-actionable alarms which could be accelerated by improving the data availability.

Systematic Review Registration:https://www.crd.york.ac.uk/prospero/display_record.php?ID=CRD42021233461, identifier: CRD42021233461.

Foundations of Time Series Analysis

For almost a century, classical statistical methods including exponential smoothing and autoregression integrated moving averages (ARIMA) have been predominant in the analysis of time series (TS) and in the pursuit of forecasting future events from historical data. TS are chronological sequences of observations, and TS data are therefore prevalent in many aspects of clinical medicine and academic neuroscience. With the rise of highly complex and nonlinear datasets, machine learning (ML) methods have become increasingly popular for prediction or pattern detection and within neurosciences, including neurosurgery. ML methods regularly outperform classical methods and have been successfully applied to, inter alia, predict physiological responses in intracranial pressure monitoring or to identify seizures in EEGs. Implementing nonparametric methods for TS analysis in clinical practice can benefit clinical decision making and sharpen our diagnostic armory.

Application Scenarios for Artificial Intelligence in Nursing Care: Rapid Review

Background

Artificial intelligence (AI) holds the promise of supporting nurses’ clinical decision-making in complex care situations or conducting tasks that are remote from direct patient interaction, such as documentation processes. There has been an increase in the research and development of AI applications for nursing care, but there is a persistent lack of an extensive overview covering the evidence base for promising application scenarios.

Objective

This study synthesizes literature on application scenarios for AI in nursing care settings as well as highlights adjacent aspects in the ethical, legal, and social discourse surrounding the application of AI in nursing care.

Methods

Following a rapid review design, PubMed, CINAHL, Association for Computing Machinery Digital Library, Institute of Electrical and Electronics Engineers Xplore, Digital Bibliography & Library Project, and Association for Information Systems Library, as well as the libraries of leading AI conferences, were searched in June 2020. Publications of original quantitative and qualitative research, systematic reviews, discussion papers, and essays on the ethical, legal, and social implications published in English were included. Eligible studies were analyzed on the basis of predetermined selection criteria.

Results

The titles and abstracts of 7016 publications and 704 full texts were screened, and 292 publications were included. Hospitals were the most prominent study setting, followed by independent living at home; fewer application scenarios were identified for nursing homes or home care. Most studies used machine learning algorithms, whereas expert or hybrid systems were entailed in less than every 10th publication. The application context of focusing on image and signal processing with tracking, monitoring, or the classification of activity and health followed by care coordination and communication, as well as fall detection, was the main purpose of AI applications. Few studies have reported the effects of AI applications on clinical or organizational outcomes, lacking particularly in data gathered outside laboratory conditions. In addition to technological requirements, the reporting and inclusion of certain requirements capture more overarching topics, such as data privacy, safety, and technology acceptance. Ethical, legal, and social implications reflect the discourse on technology use in health care but have mostly not been discussed in meaningful and potentially encompassing detail.

Conclusions

The results highlight the potential for the application of AI systems in different nursing care settings. Considering the lack of findings on the effectiveness and application of AI systems in real-world scenarios, future research should reflect on a more nursing care–specific perspective toward objectives, outcomes, and benefits. We identify that, crucially, an advancement in technological-societal discourse that surrounds the ethical and legal implications of AI applications in nursing care is a necessary next step. Further, we outline the need for greater participation among all of the stakeholders involved.

Big data, machine learning, and artificial intelligence: a field guide for neurosurgeons

Big data has transformed into a trend phrase in healthcare and neurosurgery, becoming a pervasive and inescapable phrase in everyday life. The upsurge in big data applications is a direct consequence of the drastic boom in information technology as well as the growing number of internet-connected devices called the Internet of Things in healthcare. Compared with business, marketing, and other sectors, healthcare applications are lagging due to a lack of technical knowledge among healthcare workers, technological limitations in acquiring and analyzing the data, and improper governance of healthcare big data. Despite these limitations, the medical literature is flooded with big data-related articles, and most of these are filled with abstruse terminologies such as machine learning, artificial intelligence, artificial neural network, and algorithm. Many of the recent articles are restricted to neurosurgical registries, creating a false impression that big data is synonymous with registries. Others advocate that the utilization of big data will be the panacea to all healthcare problems and research in the future. Without a proper understanding of these principles, it becomes easy to get lost without the ability to differentiate hype from reality. To that end, the authors give a brief narrative of big data analysis in neurosurgery and review its applications, limitations, and the challenges it presents for neurosurgeons and healthcare professionals naive to this field. Awareness of these basic concepts will allow neurosurgeons to understand the literature regarding big data, enabling them to make better decisions and deliver personalized care.

DeepClean: Self-Supervised Artefact Rejection for Intensive Care Waveform Data Using Deep Generative Learning

Waveform physiological data are important in the treatment of critically ill patients in the intensive care unit. Such recordings are susceptible to artefacts, which must be removed before the data can be reused for alerting or reprocessed for other clinical or research purposes. Accurate removal of artefacts reduces bias and uncertainty in clinical assessment, as well as the false positive rate of ICU alarms, and is therefore a key component in providing optimal clinical care. In this work, we present DeepClean, a prototype self-supervised artefact detection system using a convolutional variational autoencoder deep neural network that avoids costly and painstaking manual annotation, requiring only easily obtained 'good' data for training. For a test case with invasive arterial blood pressure, we demonstrate that our algorithm can detect the presence of an artefact within a 10s sample of data with sensitivity and specificity around 90%. Furthermore, DeepClean was able to identify regions of artefacts within such samples with high accuracy, and we show that it significantly outperforms a baseline principal component analysis approach in both signal reconstruction and artefact detection. DeepClean learns a generative model and therefore may also be used for imputation of missing data.

Patient-adaptable intracranial pressure morphology analysis using a probabilistic model-based approach

OBJECTIVE

We present a framework for analyzing the morphology of intracranial pressure (ICP). The analysis of ICP signals is challenging due to the non-linear and non-Gaussian characteristics of the signal dynamics, inevitable corruption by noise and artifacts, and variations in ICP pulse morphology among individuals with different neurological conditions. Existing frameworks make unrealistic assumptions regarding ICP dynamics and are not tuned for individual patients.

APPROACH

We propose a dynamic Bayesian network for automated detection of three major ICP pulsatile components. The proposed model captures the non-linear and non-Gaussian dynamics of ICP morphology and further adapts to a patient as the individual's ICP measurements are received. To make the approach more robust, we leverage evidence reversal and present an inference algorithm to obtain the posterior distribution over the locations of pulsatile components.

MAIN RESULTS

We evaluate our approach on a dataset with over 700 h of recordings from 66 neurological patients, where the pulsatile components were annotated by prior studies. The algorithm obtains accuracies of 96.56%, 92.39%, and 94.04% for the detection of each pulsatile component in the test set, showing significant improvement over existing approaches.

SIGNIFICANCE

Continuous ICP monitoring is essential in guiding the treatment of neurological conditions such as traumatic brain injuries. An automated approach for ICP morphology analysis is a step towards enhancing patient care with minimal supervision. Compared to previous methods, our framework offers several advantages. It learns the parameters that model each patient's ICP in an unsupervised manner, resulting in an accurate morphology analysis. The Bayesian model-based framework provides uncertainty estimates and reveals interesting facts about the ICP dynamics. The framework can readily be applied to replace existing morphological analysis methods and support the use of ICP pulse morphological features to aid the monitoring of pathophysiological changes of relevance to the care of patients with acute brain injuries.

Intracranial Pressure Monitoring Signals After Traumatic Brain Injury: A Narrative Overview and Conceptual Data Science Framework

Continuous intracranial pressure (ICP) monitoring is a cornerstone of neurocritical care after severe brain injuries such as traumatic brain injury and acts as a biomarker of secondary brain injury. With the rapid development of artificial intelligent (AI) approaches to data analysis, the acquisition, storage, real-time analysis, and interpretation of physiological signal data can bring insights to the field of neurocritical care bioinformatics. We review the existing literature on the quantification and analysis of the ICP waveform and present an integrated framework to incorporate signal processing tools, advanced statistical methods, and machine learning techniques in order to comprehensively understand the ICP signal and its clinical importance. Our goals were to identify the strengths and pitfalls of existing methods for data cleaning, information extraction, and application. In particular, we describe the use of ICP signal analytics to detect intracranial hypertension and to predict both short-term intracranial hypertension and long-term clinical outcome. We provide a well-organized roadmap for future researchers based on existing literature and a computational approach to clinically-relevant biomedical signal data.

Artifact removal from neurophysiological signals: impact on intracranial and arterial pressure monitoring in traumatic brain injury

OBJECTIVE

Monitoring intracranial and arterial blood pressure (ICP and ABP, respectively) provides crucial information regarding the neurological status of patients with traumatic brain injury (TBI). However, these signals are often heavily affected by artifacts, which may significantly reduce the reliability of the clinical determinations derived from the signals. The goal of this work was to eliminate signal artifacts from continuous ICP and ABP monitoring via deep learning techniques and to assess the changes in the prognostic capacities of clinical parameters after artifact elimination.

METHODS

The first 24 hours of monitoring ICP and ABP in a total of 309 patients with TBI was retrospectively analyzed. An artifact elimination model for ICP and ABP was constructed via a stacked convolutional autoencoder (SCAE) and convolutional neural network (CNN) with 10-fold cross-validation tests. The prevalence and prognostic capacity of ICP- and ABP-related clinical events were compared before and after artifact elimination.

RESULTS

The proposed SCAE-CNN model exhibited reliable accuracy in eliminating ABP and ICP artifacts (net prediction rates of 97% and 94%, respectively). The prevalence of ICP- and ABP-related clinical events (i.e., systemic hypotension, intracranial hypertension, cerebral hypoperfusion, and poor cerebrovascular reactivity) all decreased significantly after artifact removal.

CONCLUSIONS

The SCAE-CNN model can be reliably used to eliminate artifacts, which significantly improves the reliability and efficacy of ICP- and ABP-derived clinical parameters for prognostic determinations after TBI.

Development, Implementation, and Evaluation of a Personalized Machine Learning Algorithm for Clinical Decision Support: Case Study With Shingles Vaccination

BACKGROUND

Although clinical decision support (CDS) alerts are effective reminders of best practices, their effectiveness is blunted by clinicians who fail to respond to an overabundance of inappropriate alerts. An electronic health record (EHR)-integrated machine learning (ML) algorithm is a potentially powerful tool to increase the signal-to-noise ratio of CDS alerts and positively impact the clinician's interaction with these alerts in general.

OBJECTIVE

This study aimed to describe the development and implementation of an ML-based signal-to-noise optimization system (SmartCDS) to increase the signal of alerts by decreasing the volume of low-value herpes zoster (shingles) vaccination alerts.

METHODS

We built and deployed SmartCDS, which builds personalized user activity profiles to suppress shingles vaccination alerts unlikely to yield a clinician's interaction. We extracted all records of shingles alerts from January 2017 to March 2019 from our EHR system, including 327,737 encounters, 780 providers, and 144,438 patients.

RESULTS

During the 6 weeks of pilot deployment, the SmartCDS system suppressed an average of 43.67% (15,425/35,315) potential shingles alerts (appointments) and maintained stable counts of weekly shingles vaccination orders (326.3 with system active vs 331.3 in the control group; P=.38) and weekly user-alert interactions (1118.3 with system active vs 1166.3 in the control group; P=.20).

CONCLUSIONS

All key statistics remained stable while the system was turned on. Although the results are promising, the characteristics of the system can be subject to future data shifts, which require automated logging and monitoring. We demonstrated that an automated, ML-based method and data architecture to suppress alerts are feasible without detriment to overall order rates. This work is the first alert suppression ML-based model deployed in practice and serves as foundational work in encounter-level customization of alert display to maximize effectiveness.

Is the Sequence of SuperAlarm Triggers More Predictive Than Sequence of the Currently Utilized Patient Monitor Alarms?

OBJECTIVE

Our previous studies have shown that "code blue" events can be predicted by SuperAlarm patterns that are multivariate combinations of monitor alarms and laboratory test results cooccurring frequently preceding the events but rarely among control patients. Deploying these patterns to the monitor data streams can generate SuperAlarm sequences. The objective of this study is to test the hypothesis that SuperAlarm sequences may contain more predictive sequential patterns than monitor alarms sequences.

METHODS

Monitor alarms and laboratory test results are extracted from a total of 254 adult coded and 2213 control patients. The training dataset is composed of subsequences that are sampled from complete sequences and then further represented as fixed-dimensional vectors by the term frequency inverse document frequency method. The information gain technique and weighted support vector machine are adopted to select the most relevant features and train a classifier to differentiate sequences between coded patients and control patients. Performances are assessed based on an independent dataset using three metrics: sensitivity of lead time (Sen L @T), alarm frequency reduction rate (AFRR), and work-up to detection ratio (WDR).

RESULTS

The performance of 12-h-long sequences of SuperAlarm can yield a Sen L@2 of 93.33%, an AFRR of 87.28%, and a WDR of 3.01. At an AFRR = 87.28%, Sen L@2 for raw alarm sequences and discretized alarm sequences are 73.33% and 70.19%, respectively. At a WDR = 3.01, Sen L@2 are 49.88% and 43.33%.

CONCLUSION AND SIGNIFICANCE

The results demonstrate that SuperAlarm sequences indeed outperform monitor alarm sequences and suggest that one can focus on sequential patterns from SuperAlarm sequences to develop more precise patient monitoring solutions.

Detection of Intracranial Hypertension using Deep Learning

Intracranial Hypertension, a disorder characterized by elevated pressure in the brain, is typically monitored in neurointensive care and diagnosed only after elevation has occurred. This reaction-based method of treatment leaves patients at higher risk of additional complications in case of misdetection. The detection of intracranial hypertension has been the subject of many recent studies in an attempt to accurately characterize the causes of hypertension, specifically examining waveform morphology. We investigate the use of Deep Learning, a hierarchical form of machine learning, to model the relationship between hypertension and waveform morphology, giving us the ability to accurately detect presence hypertension. Data from 60 patients, showing intracranial pressure levels over a half hour time span, was used to evaluate the model. We divided each patient's recording into average normalized beats over 30 sec segments, assigning each beat a label of high (i.e. greater than 15 mmHg) or low intracranial pressure. The model was tested to predict the presence of elevated intracranial pressure. The algorithm was found to be 92.05± 2.25% accurate in detecting intracranial hypertension on our dataset.

Integrating monitor alarms with laboratory test results to enhance patient deterioration prediction

Patient monitors in modern hospitals have become ubiquitous but they generate an excessive number of false alarms causing alarm fatigue. Our previous work showed that combinations of frequently co-occurring monitor alarms, called SuperAlarm patterns, were capable of predicting in-hospital code blue events at a lower alarm frequency. In the present study, we extend the conceptual domain of a SuperAlarm to incorporate laboratory test results along with monitor alarms so as to build an integrated data set to mine SuperAlarm patterns. We propose two approaches to integrate monitor alarms with laboratory test results and use a maximal frequent itemsets mining algorithm to find SuperAlarm patterns. Under an acceptable false positive rate FPRmax, optimal parameters including the minimum support threshold and the length of time window for the algorithm to find the combinations of monitor alarms and laboratory test results are determined based on a 10-fold cross-validation set. SuperAlarm candidates are generated under these optimal parameters. The final SuperAlarm patterns are obtained by further removing the candidates with false positive rate>FPRmax. The performance of SuperAlarm patterns are assessed using an independent test data set. First, we calculate the sensitivity with respect to prediction window and the sensitivity with respect to lead time. Second, we calculate the false SuperAlarm ratio (ratio of the hourly number of SuperAlarm triggers for control patients to that of the monitor alarms, or that of regular monitor alarms plus laboratory test results if the SuperAlarm patterns contain laboratory test results) and the work-up to detection ratio, WDR (ratio of the number of patients triggering any SuperAlarm patterns to that of code blue patients triggering any SuperAlarm patterns). The experiment results demonstrate that when varying FPRmax between 0.02 and 0.15, the SuperAlarm patterns composed of monitor alarms along with the last two laboratory test results are triggered at least once for [56.7-93.3%] of code blue patients within an 1-h prediction window before code blue events and for [43.3-90.0%] of code blue patients at least 1-h ahead of code blue events. However, the hourly number of these SuperAlarm patterns occurring in control patients is only [2.0-14.8%] of that of regular monitor alarms with WDR varying between 2.1 and 6.5 in a 12-h window. For a given FPRmax threshold, the SuperAlarm set generated from the integrated data set has higher sensitivity and lower WDR than the SuperAlarm set generated from the regular monitor alarm data set. In addition, the McNemar's test also shows that the performance of the SuperAlarm set from the integrated data set is significantly different from that of the SuperAlarm set from the regular monitor alarm data set. We therefore conclude that the SuperAlarm patterns generated from the integrated data set are better at predicting code blue events.

Complex signals bioinformatics: evaluation of heart rate characteristics monitoring as a novel risk marker for neonatal sepsis

PURPOSES

Heart rate characteristics monitoring for early detection of late-onset neonatal sepsis was first described in 2003. This technique, which uses mathematical methods to report the fold-increase in the risk of imminent neonatal sepsis, adds independent information to laboratory tests and clinical findings, and, in a large randomized trial, reduced NICU mortality of very low birth weight infants. Through re-analysis and new secondary analyses of published studies, we have systematically evaluated the utility of this new risk marker for screening the growing population of premature infants.

METHODS

We followed the guidelines proposed by Hlatky et al. (Circulation, 119:2408-2416, 2009), reviewed past works, and re-analyzed data from 1,489 patients receiving conventional monitoring alone, 348 of whom had 488 episodes of proven sepsis, in the large randomized trial.

RESULTS

Heart rate characteristics monitoring passed all phases of risk marker development from proof of concept to improvement of clinical outcomes. The predictiveness curve affirmed good calibration, and addition of the heart rate characteristics index to predictive models using standard risk factors favorably impacted the receiver operating characteristic curve area (increase of 0.030), continuous net reclassification index (0.389) and the integrated discrimination index (0.008), and compares well to other modern risk factors.

CONCLUSION

Heart rate characteristics monitoring is a validated risk marker for sepsis in the NICU.