Predicting vital sign deterioration with artificial intelligence or machine learning

Time to detection of serious adverse events by continuous vital sign monitoring versus clinical practice

BACKGROUND

Continuous vital sign monitoring detects far more severe vital sign deviations (SVDs) than intermittent clinical rounds, and deviations are to some extent related to subsequent serious adverse events (SAEs). Early detection of SAEs is pivotal to allow for effective interventions but the time relationship between detection of SAEs by continuous vital sign monitoring versus clinical practice is not well-described at the general ward.

AIM

To quantify the time difference between detection of SAEs by continuous vital sign monitoring and clinical suspicion of deterioration (CSD) in major abdominal surgery patients.

METHODS

Five hundred and five patients had their vital signs continuously monitored in combination with usual clinical practice consisting of National Early Warning Score assessments at least every 8'th hour, assessments during rounds, and other kinds of staff-patient interactions. The primary outcome was the time difference between the first chart note of CSD versus the first SVD, detected by continuous vital sign monitoring, in patients with a subsequent confirmed SAE during or up to 48 h after end of continuous vital sign monitoring.

RESULTS

Out of the 505 continuously monitored patients, 142 patients had a combination of both postoperative SAE, CSD and SVD, and thus were included in the primary analysis. The median time from the first SVD to SAE was 42.8 h (interquartile range 19.8-72.1 h) compared to 13 minutes (interquartile range - 4.8 to 3.5 h) for CSD with a median difference of 48.1 h (95% confidence interval 43.0-54.8 h), p-value < .001.

CONCLUSION

Continuous vital sign monitoring detects signs of oncoming SAEs in the form of SVD hours before CSD, potentially allowing for earlier and more effective treatments to reduce the extent of SAEs.

Pediatric Monitoring Technologies and Congenital Heart Disease: A Systematic Review

Outpatient monitoring of infants with congenital heart disease has been shown to significantly reduce rates of mortality in the single ventricle population. Despite the accelerating development of miniaturized biosensors and electronics, and a growing market demand for at-home monitoring devices, the application of these technologies to infants and children is significantly delayed compared with the development of devices for adults. This article aims to review the current landscape of available monitoring technologies and devices for pediatric patients to describe the gap between technologies and clinical needs with the goal of progressing development of clinically and scientifically validated pediatric monitoring devices.

CON: The hypotension prediction index is not a validated predictor of hypotension

The Hypotension Prediction Index (HPI) algorithm is a commercial prediction algorithm developed to predict hypotension, a mean arterial pressure (MAP) below 65 mmHg. Although HPI has been investigated in several studies, recent concerns of have been raised regarding HPI's predictive abilities, which may have been overstated. A selection bias may have forced the HPI algorithm to learn almost exclusively from MAP. This CON position paper describes the selection bias further and summarises the scientific status of HPI's predictive abilities, including the meaning of a recent erratum retracting the primary conclusion of a published HPI validation study. We argue that the HPI algorithm needs re-validation or complete re-development to achieve a clinically relevant 'added value' in comparison with the predictive performance of a simple and costless MAP alarm threshold in the range of 70 to 75 mmHg.

Artificial intelligence in vascular surgical decision making

Despite advances in prevention, detection, and treatment, cardiovascular disease is a leading cause of mortality and represents a major health problem worldwide. Artificial intelligence and machine learning have brought new insights to the management of vascular diseases by allowing analysis of huge and complex datasets and by offering new techniques to develop advanced imaging analysis. Artificial intelligence-based applications have the potential to improve prognostic evaluation and evidence-based decision making and contribute to vascular therapeutic decision making. In this scoping review, we provide an overview on how artificial intelligence could help in vascular surgical clinical decision making, highlighting potential benefits, current limitations, and future challenges.

Prediction of intraoperative hypotension from the linear extrapolation of mean arterial pressure

BACKGROUND

Hypotension prediction index (HPI) software is a proprietary machine learning-based algorithm used to predict intraoperative hypotension (IOH). HPI has shown superiority in predicting IOH when compared to the predictive value of changes in mean arterial pressure (ΔMAP) alone. However, the predictive value of ΔMAP alone, with no reference to the absolute level of MAP, is counterintuitive and poor at predicting IOH. A simple linear extrapolation of mean arterial pressure (LepMAP) is closer to the clinical approach.

OBJECTIVES

Our primary objective was to investigate whether LepMAP better predicts IOH than ΔMAP alone.

DESIGN

Retrospective diagnostic accuracy study.

SETTING

Two tertiary University Hospitals between May 2019 and December 2019.

PATIENTS

A total of 83 adult patients undergoing high risk non-cardiac surgery.

DATA SOURCES

Arterial pressure data were automatically extracted from the anaesthesia data collection software (one value per minute). IOH was defined as MAP < 65 mmHg.

ANALYSIS

Correlations for repeated measurements and the area under the curve (AUC) from receiver operating characteristics (ROC) were determined for the ability of LepMAP and ΔMAP to predict IOH at 1, 2 and 5 min before its occurrence (A-analysis, using the whole dataset). Data were also analysed after exclusion of MAP values between 65 and 75 mmHg (B-analysis).

RESULTS

A total of 24 318 segments of ten minutes duration were analysed. In the A-analysis, ROC AUCs to predict IOH at 1, 2 and 5 min before its occurrence by LepMAP were 0.87 (95% confidence interval, CI, 0.86 to 0.88), 0.81 (95% CI, 0.79 to 0.83) and 0.69 (95% CI, 0.66 to 0.71) and for ΔMAP alone 0.59 (95% CI, 0.57 to 0.62), 0.61 (95% CI, 0.59 to 0.64), 0.57 (95% CI, 0.54 to 0.69), respectively. In the B analysis for LepMAP these were 0.97 (95% CI, 0.9 to 0.98), 0.93 (95% CI, 0.92 to 0.95) and 0.86 (95% CI, 0.84 to 0.88), respectively, and for ΔMAP alone 0.59 (95% CI, 0.53 to 0.58), 0.56 (95% CI, 0.54 to 0.59), 0.54 (95% CI, 0.51 to 0.57), respectively. LepMAP ROC AUCs were significantly higher than ΔMAP ROC AUCs in all cases.

CONCLUSIONS

LepMAP provides reliable real-time and continuous prediction of IOH 1 and 2 min before its occurrence. LepMAP offers better discrimination than ΔMAP at 1, 2 and 5 min before its occurrence. Future studies evaluating machine learning algorithms to predict IOH should be compared with LepMAP rather than ΔMAP.

Machine learning-based prediction of critical illness in children visiting the emergency department

OBJECTIVES

Triage is an essential emergency department (ED) process designed to provide timely management depending on acuity and severity; however, the process may be inconsistent with clinical and hospitalization outcomes. Therefore, studies have attempted to augment this process with machine learning models, showing advantages in predicting critical conditions and hospitalization outcomes. The aim of this study was to utilize nationwide registry data to develop a machine learning-based classification model to predict the clinical course of pediatric ED visits.

METHODS

This cross-sectional observational study used data from the National Emergency Department Information System on emergency visits of children under 15 years of age from January 1, 2016, to December 31, 2017. The primary and secondary outcomes were to identify critically ill children and predict hospitalization from triage data, respectively. We developed and tested a random forest model with the under sampled dataset and validated the model using the entire dataset. We compared the model's performance with that of the conventional triage system.

RESULTS

A total of 2,621,710 children were eligible for the analysis and included 12,951 (0.5%) critical outcomes and 303,808 (11.6%) hospitalizations. After validation, the area under the receiver operating characteristic curve was 0.991 (95% confidence interval [CI] 0.991-0.992) for critical outcomes and 0.943 (95% CI 0.943-0.944) for hospitalization, which were higher than those of the conventional triage system.

CONCLUSIONS

The machine learning-based model using structured triage data from a nationwide database can effectively predict critical illness and hospitalizations among children visiting the ED.

Accessing Artificial Intelligence for Clinical Decision-Making

Advancements in computing and data from the near universal acceptance and implementation of electronic health records has been formative for the growth of personalized, automated, and immediate patient care models that were not previously possible. Artificial intelligence (AI) and its subfields of machine learning, reinforcement learning, and deep learning are well-suited to deal with such data. The authors in this paper review current applications of AI in clinical medicine and discuss the most likely future contributions that AI will provide to the healthcare industry. For instance, in response to the need to risk stratify patients, appropriately cultivated and curated data can assist decision-makers in stratifying preoperative patients into risk categories, as well as categorizing the severity of ailments and health for non-operative patients admitted to hospitals. Previous overt, traditional vital signs and laboratory values that are used to signal alarms for an acutely decompensating patient may be replaced by continuously monitoring and updating AI tools that can pick up early imperceptible patterns predicting subtle health deterioration. Furthermore, AI may help overcome challenges with multiple outcome optimization limitations or sequential decision-making protocols that limit individualized patient care. Despite these tremendously helpful advancements, the data sets that AI models train on and develop have the potential for misapplication and thereby create concerns for application bias. Subsequently, the mechanisms governing this disruptive innovation must be understood by clinical decision-makers to prevent unnecessary harm. This need will force physicians to change their educational infrastructure to facilitate understanding AI platforms, modeling, and limitations to best acclimate practice in the age of AI. By performing a thorough narrative review, this paper examines these specific AI applications, limitations, and requisites while reviewing a few examples of major data sets that are being cultivated and curated in the US.

Outcome in patients with open abdomen treatment for peritonitis: a multidomain approach outperforms single domain predictions

Numerous patient-related clinical parameters and treatment-specific variables have been identified as causing or contributing to the severity of peritonitis. We postulated that a combination of clinical and surgical markers and scoring systems would outperform each of these predictors in isolation. To investigate this hypothesis, we developed a multivariable model to examine whether survival outcome can reliably be predicted in peritonitis patients treated with open abdomen. This single-center retrospective analysis used univariable and multivariable logistic regression modeling in combination with repeated random sub-sampling validation to examine the predictive capabilities of domain-specific predictors (i.e., demography, physiology, surgery). We analyzed data of 1,351 consecutive adult patients (55.7% male) who underwent open abdominal surgery in the study period (January 1998 to December 2018). Core variables included demographics, clinical scores, surgical indices and indicators of organ dysfunction, peritonitis index, incision type, fascia closure, wound healing, and fascial dehiscence. Postoperative complications were also added when available. A multidomain peritonitis prediction model (MPPM) was constructed to bridge the mortality predictions from individual domains (demographic, physiological and surgical). The MPPM is based on data of n = 597 patients, features high predictive capabilities (area under the receiver operating curve: 0.87 (0.85 to 0.90, 95% CI)) and is well calibrated. The surgical predictor “skin closure” was found to be the most important predictor of survival in our cohort, closely followed by the two physiological predictors SAPS-II and MPI. Marginal effects plots highlight the effect of individual outcomes on the prediction of survival outcome in patients undergoing staged laparotomies for treatment of peritonitis. Although most single indices exhibited moderate performance, we observed that the predictive performance was markedly increased when an integrative prediction model was applied. Our proposed MPPM integrative prediction model may outperform the predictive power of current models.

Real-time machine learning-based intensive care unit alarm classification without prior knowledge of the underlying rhythm

AIMS

This work attempts to develop a standalone heart rhythm alerting system for the intensive care unit (ICU), where life-threatening arrhythmias have to be identified/alerted more precisely and more instantaneously (i.e. with lower latency) than existing bedside monitors.

METHODS AND RESULTS

We use the dataset from the PhysioNet 2015 Challenge, which contains records that led to true and false arrhythmic alarms in the ICU. These records have been re-annotated as one of eight classes, namely (i) asystole, (ii) extreme bradycardia, (iii) extreme tachycardia, (iv) ventricular fibrillation (VF), (v) ventricular tachycardia (VT), (vi) normal sinus rhythm, (vii) sinus tachycardia, and (viii) noise/artefacts. Arrhythmia-specific features and features that measure the signal quality were extracted from all the records. To improve VF detection, an improved, over an existing, single-lead R-wave detection was developed that takes into account the R-waves detected in all electrocardiographic (ECG) leads. To avoid false R-wave detection due to pacing spikes, ECG signals were filtered with a low pass filter prior to R-wave detection, while the raw signals were used for feature extraction. Random forest was used as the classifier, and 10-time five-fold cross-validation, resulted in a macro-average sensitivity of 81.54%.

CONCLUSIONS

In conclusion, comparing with the bedside monitors used in the PhysioNet 2015 competition, we find that our method achieves higher positive predictive values for asystole, extreme bradycardia, VT, and VF; furthermore, our method is able to alert the presence of arrhythmia instantaneously, i.e. up to 4 s earlier.

Leveraging the trust of nurses to advance a digital agenda in Europe: a critical review of health policy literature

This article is a critical and integrative review of health policy literature examining artificial intelligence (AI) and its implications for healthcare systems and the frontline nursing workforce. A key focus is on co-creation as essential for the deployment and adoption of AI. Our review hinges on the European Commission's White Paper on Artificial Intelligence from 2020, which provides a useful roadmap. The value of health data spaces and electronic health records (EHRs) is considered; and the role of advanced nurse practitioners in harnessing the potential of AI tools in their practice is articulated. Finally, this paper examines "trust" as a precondition for the successful deployment and adoption of AI in Europe. AI applications in healthcare can enhance safety and quality, and mitigate against common risks and challenges, once the necessary level of trust is achieved among all stakeholders. Such an approach can enable effective preventative care across healthcare settings, particularly community and primary care. However, the acceptance of AI tools in healthcare is dependent on the robustness, validity and reliability of data collected and donated from EHRs. Nurse stakeholders have a key role to play in this regard, since trust can only be fostered through engaging frontline end-users in the co-design of EHRs and new AI tools. Nurses hold an intimate understanding of the direct benefits of such technology, such as releasing valuable nursing time for essential patient care, and empowering patients and their family members as recipients of nursing care. This article brings together insights from a unique group of stakeholders to explore the interaction between AI, the co-creation of data spaces and EHRs, and the role of the frontline nursing workforce. We identify the pre-conditions needed for successful deployment of AI and offer insights regarding the importance of co-creating the future European Health Data Space.

A Pilot Study of End-Tidal Carbon Dioxide in Prediction of Inhospital Cardiac Arrests

A validated means to predict inhospital cardiac arrest is lacking. The purpose of this study was to evaluate the changes in end-tidal carbon dioxide, as it correlates with the progression to inhospital cardiac arrest in ICU patients.

Artificial Intelligence in the Intensive Care Unit

This article is one of ten reviews selected from the Annual Update in Intensive Care and Emergency Medicine 2020. Other selected articles can be found online at https://www.biomedcentral.com/collections/annualupdate2020. Further information about the Annual Update in Intensive Care and Emergency Medicine is available from http://www.springer.com/series/8901.

Intraoperative hypotension and its prediction

Intraoperative hypotension (IOH) very commonly accompanies general anaesthesia in patients undergoing major surgical procedures. The development of IOH is unwanted, since it is associated with adverse outcomes such as acute kidney injury and myocardial injury, stroke and mortality. Although the definition of IOH is variable, harm starts to occur below a mean arterial pressure (MAP) threshold of 65 mmHg. The odds of adverse outcome increase for increasing duration and/or magnitude of IOH below this threshold, and even short periods of IOH seem to be associated with adverse outcomes. Therefore, reducing the hypotensive burden by predicting and preventing IOH through proactive appropriate treatment may potentially improve patient outcome. In this review article, we summarise the current state of the prediction of IOH by the use of so-called machine-learning algorithms. Machine-learning algorithms that use high-fidelity data from the arterial pressure waveform, may be used to reveal 'traits' that are unseen by the human eye and are associated with the later development of IOH. These algorithms can use large datasets for 'training', and can subsequently be used by clinicians for haemodynamic monitoring and guiding therapy. A first clinically available application, the hypotension prediction index (HPI), is aimed to predict an impending hypotensive event, and additionally, to guide appropriate treatment by calculated secondary variables to asses preload (dynamic preload variables), contractility (dP/dtmax), and afterload (dynamic arterial elastance, Eadyn). In this narrative review, we summarise the current state of the prediction of hypotension using such novel, automated algorithms and we will highlight HPI and the secondary variables provided to identify the probable origin of the (impending) hypotensive event.