An algorithm strategy for precise patient monitoring in a connected healthcare enterprise

Intelligent Clinical Decision Support

Early recognition of pathologic cardiorespiratory stress and forecasting cardiorespiratory decompensation in the critically ill is difficult even in highly monitored patients in the Intensive Care Unit (ICU). Instability can be intuitively defined as the overt manifestation of the failure of the host to adequately respond to cardiorespiratory stress. The enormous volume of patient data available in ICU environments, both of high-frequency numeric and waveform data accessible from bedside monitors, plus Electronic Health Record (EHR) data, presents a platform ripe for Artificial Intelligence (AI) approaches for the detection and forecasting of instability, and data-driven intelligent clinical decision support (CDS). Building unbiased, reliable, and usable AI-based systems across health care sites is rapidly becoming a high priority, specifically as these systems relate to diagnostics, forecasting, and bedside clinical decision support. The ICU environment is particularly well-positioned to demonstrate the value of AI in saving lives. The goal is to create AI models embedded in a real-time CDS for forecasting and mitigation of critical instability in ICU patients of sufficient readiness to be deployed at the bedside. Such a system must leverage multi-source patient data, machine learning, systems engineering, and human action expertise, the latter being key to successful CDS implementation in the clinical workflow and evaluation of bias. We present one approach to create an operationally relevant AI-based forecasting CDS system.

Disentangling the clinical data chaos: User-centered interface system design for trauma centers

This paper presents a year-long study of our project, aiming at (1) understanding the work practices of clinical staff in trauma intensive care units (TICUs) at a trauma center, with respect to their usage of clinical data interface systems, and (2) developing and evaluating an intuitive and user-centered clinical data interface system for their TICU environments. Based on a long-term field study in an urban trauma center that involved observation-, interview-, and survey-based studies to understand our target users and their working environment, we designed and implemented MediSenseView as a working prototype. MediSenseView is a clinical-data interface system, which was developed through the identification of three core challenges of existing interface system use in a trauma care unit-device separation, usage inefficiency, and system immobility-from the perspectives of three staff groups in our target environment (i.e., doctors, clinical nurses and research nurses), and through an iterative design study. The results from our pilot deployment of MediSenseView and a user study performed with 28 trauma center staff members highlight their work efficiency and satisfaction with MediSenseView compared to existing clinical data interface systems in the hospital.

Reducing medical device alarms by an order of magnitude: A human factors approach

The intensive care unit (ICU) is one of the most technically advanced environments in healthcare, using a multitude of medical devices for drug administration, mechanical ventilation and patient monitoring. However, these technologies currently come with disadvantages, namely noise pollution, information overload and alarm fatigue—all caused by too many alarms. Individual medical devices currently generate alarms independently, without any coordination or prioritisation with other devices, leading to a cacophony where important alarms can be lost amongst trivial ones, occasionally with serious or even fatal consequences for patients. We have called this approach to the design of medical devices the single-device paradigm, and believe it is obsolete in modern hospitals where patients are typically connected to several devices simultaneously. Alarm rates of one alarm every four minutes for only the physiological monitors (as recorded in the ICUs of two hospitals contributing to this paper) degrades the quality of the patient’s healing environment and threatens patient safety by constantly distracting healthcare professionals. We outline a new approach to medical device design involving the application of human factors principles which have been successful in eliminating alarm fatigue in commercial aviation. Our approach comprises the networked-device paradigm, comprehensive alarms and humaniform information displays. Instead of each medical device alarming separately at the patient’s bedside, our proposed approach will integrate, prioritise and optimise alarms across all devices attached to each patient, display information more intuitively and hence increase alarm quality while reducing the number of alarms by an order of magnitude below current levels.

Role of Artificial Intelligence in Patient Safety Outcomes: Systematic Literature Review

BACKGROUND

Artificial intelligence (AI) provides opportunities to identify the health risks of patients and thus influence patient safety outcomes.

OBJECTIVE

The purpose of this systematic literature review was to identify and analyze quantitative studies utilizing or integrating AI to address and report clinical-level patient safety outcomes.

METHODS

We restricted our search to the PubMed, PubMed Central, and Web of Science databases to retrieve research articles published in English between January 2009 and August 2019. We focused on quantitative studies that reported positive, negative, or intermediate changes in patient safety outcomes using AI apps, specifically those based on machine-learning algorithms and natural language processing. Quantitative studies reporting only AI performance but not its influence on patient safety outcomes were excluded from further review.

RESULTS

We identified 53 eligible studies, which were summarized concerning their patient safety subcategories, the most frequently used AI, and reported performance metrics. Recognized safety subcategories were clinical alarms (n=9; mainly based on decision tree models), clinical reports (n=21; based on support vector machine models), and drug safety (n=23; mainly based on decision tree models). Analysis of these 53 studies also identified two essential findings: (1) the lack of a standardized benchmark and (2) heterogeneity in AI reporting.

CONCLUSIONS

This systematic review indicates that AI-enabled decision support systems, when implemented correctly, can aid in enhancing patient safety by improving error detection, patient stratification, and drug management. Future work is still needed for robust validation of these systems in prospective and real-world clinical environments to understand how well AI can predict safety outcomes in health care settings.

Generalizability of SuperAlarm via Cross-Institutional Performance Evaluation

Bedside patient monitors are ubiquitous tools in modern critical care units to provide timely patient status. However, current systems suffer from high volume of false alarms leading to alarm fatigue, one of top technical hazards in clinical settings. Many studies are racing to develop improved algorithms towards precision patient monitoring, while little has been done to investigate the aspect of algorithm generalizability across different health institutions. Our group has been developing an evolving framework termed SuperAlarm that extracts multivariate patterns in data streams (monitor alarms, electronic health records and physiologic waveforms) of modern health enterprise to predict patient deterioration and has demonstrated great potential in mitigating alarm fatigue. In this study, we further investigate the generalizability of SuperAlarm by designing a comprehensive approach to achieve performance comparison in predicting in-hospital code blue (CB) events across two health institutions. SuperAlarm model trained with alarm data in one institution is tested on both internal and external test sets. Results show comparable performance with sensitivity up to 80% within one-hour window of events and over 90% in reduction of false alarms in both institutions. Cross-institutional performance agreement can be further improved by predicting a more stringent CB subtype (cardiopulmonary arrest), with internal sensitivity lying within 95% confident interval of external one up to 8-hour before event onset. The cross-institutional performance comparison offers first-hand knowledge on both advantages and challenges in generalizing a prediction algorithm across different institutions, which hold key information to guide the design of model training and deployment strategy.

False alarms reduction in non-convulsive status epilepticus detection via continuous EEG analysis

OBJECTIVE

Frequent false alarms from computer-assisted monitoring systems may harm the safety of patients with non-convulsive status epilepticus (NCSE). In this study, we aimed at reducing false alarms in the NCSE detection based on preventing from three common errors: over-interpretation of abnormal background activity, dense short ictal discharges and continuous interictal discharges as ictal discharges.

APPROACH

We analyzed 10 participants' hospital-archived 127-hour electroencephalography (EEG) recordings with 310 ictal discharges. To reduce the false alarms caused by abnormal background activity, we used morphological features extracted by visibility graph methods in addition to time-frequency features. To reduce the false alarms caused by over-interpreting short ictal discharges and interictal discharges, we created two synthetic classes-'Suspected Non-ictal' and 'Suspected Ictal'-based on the misclassified categories and constructed a synthetic 4-class dataset combining the standard two classes-'Non-ictal' and 'Ictal'-to train a 4-class classifier. Precision-recall curves were used to compare our proposed 4-class classification model and the standard 2-class classification model with or without the morphological features in the leave-one-out cross validation stage. The sensitivity and precision were primarily used as performance metrics for the detection of a seizure event.

MAIN RESULTS

The 4-class classification model improved the performance of the standard 2-class model, in particular increasing the precision by 15% at an 80% sensitivity level when only time-frequency features were used. Using the morphological features, the 4-class classification model achieved the best performances: a sensitivity of 93% ± 12% and a precision of 55% ± 30% in the group level. 100% accuracy was reached in a participant's 4.3-hour recording with 5 ictal discharges.

SIGNIFICANCE

False alarms in the NCSE detection were remarkably reduced using the morphological features and the proposed 4-class classification model.

Electrocardiographic right ventricular strain precedes hypoxic pulseless electrical activity cardiac arrests: Looking beyond pulmonary embolism

AIM

The role of the right ventricle (RV) in pulseless electrical activity (PEA) is poorly defined outside of pulmonary embolism. We aimed to (1) describe the continuous electrocardiographic (ECG) manifestations of RV strain (RVS) preceding PEA/Asystole in-hospital cardiac arrest (IHCA), and (2) determine the prevalence and clinical causes of RVS in PEA/Asystole IHCA.

METHODS

In this retrospective cross-sectional study, we evaluated 140 patients with continuous ECG data preceding PEA/Asystole IHCA. We iteratively defined the RVS continuous ECG pattern using the development cohort (93 patients). Clinical cause determination was blinded from ECG analysis in the validation cohort (47 patients).

RESULTS

The overall cohort had mean age 62.1 ± 17.1 years, 70% return of spontaneous circulation and 30% survival to discharge. RVS continuous ECG pattern was defined as progressive RV depolarization delay in lead V1 with at least one supporting finding of RV ischaemia or right axis deviation. Using this criterion, 66/140 (47%) cases showed preceding RVS. In patients with RVS, no pulmonary embolism was found in 9/13 (69%) autopsies and 8/10 (80%) CT chest angiograms. The positive and negative predictive value of RVS pattern for diagnosing a respiratory cause of PEA/Asystole in the validation cohort was 81% [95% CI 64-98%] and 58% [95% CI 36-81%], respectively.

CONCLUSION

RVS continuous ECG pattern preceded 47% of PEA/Asystole IHCA and is predictive of a respiratory cause of cardiac arrest, not just pulmonary embolism. These suggest that rapid elevations in pulmonary pressures and resultant RV failure may cause PEA in respiratory failure.

Usefulness of Trends in Continuous Electrocardiographic Telemetry Monitoring to Predict In-Hospital Cardiac Arrest

Survival from in-hospital cardiac arrest (IHCA) due to pulseless electrical activity/asystole remains poor. We aimed to evaluate whether electrocardiographic changes provide predictive information for risk of IHCA from pulseless electrical activity/asystole. We conducted a retrospective case-control study, utilizing continuous electrocardiographic data from case and control patients. We selected 3 consecutive 3-hour blocks (block 3, 2, and 1 in that order); block 1 immediately preceded cardiac arrest in cases, whereas block 1 was chosen at random in controls. In each block, we measured dominant positive and negative trends in electrocardiographic parameters, evaluated for arrhythmias, and compared these between consecutive blocks. We created random forest and logistic regression models, and tested them on differentiating case versus control patients (case block 1 vs control block 1), and temporal relation to cardiac arrest (case block 2 vs case block 1). Ninety-one cases (age 63.0 ± 17.6, 58% male) and 1,783 control patients (age 63.5 ± 14.8, 67% male) were evaluated. We found significant differences in electrocardiographic trends between case and control block 1, particularly in QRS duration, QTc, RR, and ST. New episodes of atrial fibrillation and bradyarrhythmias were more common before IHCA. The optimal model was the random forest, achieving an area under the curve of 0.829, 63.2% sensitivity, 94.6% specificity at differentiating case versus control block 1 on a validation set, and area under the curve 0.954, 91.2% sensitivity, 83.5% specificity at differentiating case block 1 versus case block 2. In conclusion, trends in electrocardiographic parameters during the 3-hour window immediately preceding IHCA differ significantly from other time periods, and provide robust predictive information.