Multiparameter Intelligent Monitoring in Intensive Care II: a public-access intensive care unit database

Cohort selection for clinical trials: n2c2 2018 shared task track 1

OBJECTIVE

Track 1 of the 2018 National NLP Clinical Challenges shared tasks focused on identifying which patients in a corpus of longitudinal medical records meet and do not meet identified selection criteria.

MATERIALS AND METHODS

To address this challenge, we annotated American English clinical narratives for 288 patients according to whether they met these criteria. We chose criteria from existing clinical trials that represented a variety of natural language processing tasks, including concept extraction, temporal reasoning, and inference.

RESULTS

A total of 47 teams participated in this shared task, with 224 participants in total. The participants represented 18 countries, and the teams submitted 109 total system outputs. The best-performing system achieved a micro F1 score of 0.91 using a rule-based approach. The top 10 teams used rule-based and hybrid systems to approach the problems.

DISCUSSION

Clinical narratives are open to interpretation, particularly in cases where the selection criterion may be underspecified. This leaves room for annotators to use domain knowledge and intuition in selecting patients, which may lead to error in system outputs. However, teams who consulted medical professionals while building their systems were more likely to have high recall for patients, which is preferable for patient selection systems.

CONCLUSIONS

There is not yet a 1-size-fits-all solution for natural language processing systems approaching this task. Future research in this area can look to examining criteria requiring even more complex inferences, temporal reasoning, and domain knowledge.

Hospital Admission Location Prediction via Deep Interpretable Networks for the Year-Round Improvement of Emergency Patient Care

OBJECTIVE

This paper presents a deep learning method of predicting where in a hospital emergency patients will be admitted after being triaged in the Emergency Department (ED). Such a prediction will allow for the preparation of bed space in the hospital for timely care and admission of the patient as well as allocation of resource to the relevant departments, including during periods of increased demand arising from seasonal peaks in infections.

METHODS

The problem is posed as a multi-class classification into seven separate ward types. A novel deep learning training strategy was created that combines learning via curriculum and a multi-armed bandit to exploit this curriculum post-initial training.

RESULTS

We successfully predict the initial hospital admission location with area-under-receiver-operating-curve (AUROC) ranging between 0.60 to 0.78 for the individual wards and an overall maximum accuracy of 52% where chance corresponds to 14% for this seven-class setting. Our proposed network was able to interpret which features drove the predictions using a 'network saliency' term added to the network loss function.

CONCLUSION

We have proven that prediction of location of admission in hospital for emergency patients is possible using information from triage in ED. We have also shown that there are certain tell-tale tests which indicate what space of the hospital a patient will use.

SIGNIFICANCE

It is hoped that this predictor will be of value to healthcare institutions by allowing for the planning of resource and bed space ahead of the need for it. This in turn should speed up the provision of care for the patient and allow flow of patients out of the ED thereby improving patient flow and the quality of care for the remaining patients within the ED.

CMIID: A comprehensive medical information identifier for clinical search harmonization in Data Safe Havens

Over the last decades clinical research has been driven by informatics changes nourished by distinct research endeavors. Inherent to this evolution, several issues have been the focus of a variety of studies: multi-location patient data access, interoperability between terminological and classification systems and clinical practice and records harmonization. Having these problems in mind, the Data Safe Haven paradigm emerged to promote a newborn architecture, better reasoning and safe and easy access to distinct Clinical Data Repositories. This study aim is to present a novel solution for clinical search harmonization within a safe environment, making use of a hybrid coding taxonomy that enables researchers to collect information from multiple repositories based on a clinical domain query definition. Results show that is possible to query multiple repositories using a single query definition based on clinical domains and the capabilities of the Unified Medical Language System, although it leads to deterioration of the framework response times. Participants of a Focus Group and a System Usability Scale questionnaire rated the framework with a median value of 72.5, indicating the hybrid coding taxonomy could be enriched with additional metadata to further improve the refinement of the results and enable the possibility of using this system as data quality tagging mechanism.

Robust T-End Detection via T-End Signal Quality Index and Optimal Shrinkage

An automatic accurate T-wave end (T-end) annotation for the electrocardiogram (ECG) has several important clinical applications. While there have been several algorithms proposed, their performance is usually deteriorated when the signal is noisy. Therefore, we need new techniques to support the noise robustness in T-end detection. We propose a new algorithm based on the signal quality index (SQI) for T-end, coined as tSQI, and the optimal shrinkage (OS). For segments with low tSQI, the OS is applied to enhance the signal-to-noise ratio (SNR). We validated the proposed method using eleven short-term ECG recordings from QT database available at Physionet, as well as four 14-day ECG recordings which were visually annotated at a central ECG core laboratory. We evaluated the correlation between the real-world signal quality for T-end and tSQI, and the robustness of proposed algorithm to various additive noises of different types and SNR’s. The performance of proposed algorithm on arrhythmic signals was also illustrated on MITDB arrhythmic database. The labeled signal quality is well captured by tSQI, and the proposed OS denoising help stabilize existing T-end detection algorithms under noisy situations by making the mean of detection errors decrease. Even when applied to ECGs with arrhythmia, the proposed algorithm still performed well if proper metric is applied. We proposed a new T-end annotation algorithm. The efficiency and accuracy of our algorithm makes it a good fit for clinical applications and large ECG databases. This study is limited by the small size of annotated datasets.

Towards better heartbeat segmentation with deep learning classification

The confidence of medical equipment is intimately related to false alarms. The higher the number of false events occurs, the less truthful is the equipment. In this sense, reducing (or suppressing) false positive alarms is hugely desirable. In this work, we propose a feasible and real-time approach that works as a validation method for a heartbeat segmentation third-party algorithm. The approach is based on convolutional neural networks (CNNs), which may be embedded in dedicated hardware. Our proposal aims to detect the pattern of a single heartbeat and classifies them into two classes: a heartbeat and not a heartbeat. For this, a seven-layer convolution network is employed for both data representation and classification. We evaluate our approach in two well-settled databases in the literature on the raw heartbeat signal. The first database is a conventional on-the-person database called MIT-BIH, and the second is one less uncontrolled off-the-person type database known as CYBHi. To evaluate the feasibility and the performance of the proposed approach, we use as a baseline the Pam-Tompkins algorithm, which is a well-known method in the literature and still used in the industry. We compare the baseline against the proposed approach: a CNN model validating the heartbeats detected by a third-party algorithm. In this work, the third-party algorithm is the same as the baseline for comparison purposes. The results support the feasibility of our approach showing that our method can enhance the positive prediction of the Pan-Tompkins algorithm from \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$97.84\%$$\end{document}97.84%/\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$90.28\%$$\end{document}90.28% to \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$100.00\%$$\end{document}100.00%/\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$96.77\%$$\end{document}96.77% by slightly decreasing the sensitivity from \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$95.79\%$$\end{document}95.79%/\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$96.95\%$$\end{document}96.95% to \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$92.98\%/$$\end{document}92.98%/\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$95.71\%$$\end{document}95.71% on the MIT-BIH/CYBHi databases.

Vital Signs Prediction and Early Warning Score Calculation Based on Continuous Monitoring of Hospitalised Patients Using Wearable Technology

In this prospective, interventional, international study, we investigate continuous monitoring of hospitalised patients' vital signs using wearable technology as a basis for real-time early warning scores (EWS) estimation and vital signs time-series prediction. The collected continuous monitored vital signs are heart rate, blood pressure, respiration rate, and oxygen saturation of a heterogeneous patient population hospitalised in cardiology, postsurgical, and dialysis wards. Two aspects are elaborated in this study. The first is the high-rate (every minute) estimation of the statistical values (e.g., minimum and mean) of the vital signs components of the EWS for one-minute segments in contrast with the conventional routine of 2 to 3 times per day. The second aspect explores the use of a hybrid machine learning algorithm of kNN-LS-SVM for predicting future values of monitored vital signs. It is demonstrated that a real-time implementation of EWS in clinical practice is possible. Furthermore, we showed a promising prediction performance of vital signs compared to the most recent state of the art of a boosted approach of LSTM. The reported mean absolute percentage errors of predicting one-hour averaged heart rate are 4.1, 4.5, and 5% for the upcoming one, two, and three hours respectively for cardiology patients. The obtained results in this study show the potential of using wearable technology to continuously monitor the vital signs of hospitalised patients as the real-time estimation of EWS in addition to a reliable prediction of the future values of these vital signs is presented. Ultimately, both approaches of high-rate EWS computation and vital signs time-series prediction is promising to provide efficient cost-utility, ease of mobility and portability, streaming analytics, and early warning for vital signs deterioration.

Robust PPG motion artifact detection using a 1-D convolution neural network

BACKGROUND AND OBJECTIVES

Continuous monitoring of physiological parameters such as photoplethysmography (PPG) has attracted increased interest due to advances in wearable sensors. However, PPG recordings are susceptible to various artifacts, and thus reducing the reliability of PPG-driven parameters, such as oxygen saturation, heart rate, blood pressure and respiration. This paper proposes a one-dimensional convolution neural network (1-D-CNN) to classify five-second PPG segments into clean or artifact-affected segments, avoiding data-dependent pulse segmentation techniques and heavy manual feature engineering.

METHODS

Continuous raw PPG waveforms were blindly allocated into segments with an equal length (5s) without leveraging any pulse location information and were normalized with Z-score normalization methods. A 1-D-CNN was designed to automatically learn the intrinsic features of the PPG waveform, and perform the required classification. Several training hyperparameters (initial learning rate and gradient threshold) were varied to investigate the effect of these parameters on the performance of the network. Subsequently, this proposed network was trained and validated with 30 subjects, and then tested with eight subjects, with our local dataset. Moreover, two independent datasets downloaded from the PhysioNet MIMIC II database were used to evaluate the robustness of the proposed network.

RESULTS

A 13 layer 1-D-CNN model was designed. Within our local study dataset evaluation, the proposed network achieved a testing accuracy of 94.9%. The classification accuracy of two independent datasets also achieved satisfactory accuracy of 93.8% and 86.7% respectively. Our model achieved a comparable performance with most reported works, with the potential to show good generalization as the proposed network was evaluated with multiple cohorts (overall accuracy of 94.5%).

CONCLUSION

This paper demonstrated the feasibility and effectiveness of applying blind signal processing and deep learning techniques to PPG motion artifact detection, whereby manual feature thresholding was avoided and yet a high generalization ability was achieved.

Non-contact vital-sign monitoring of patients undergoing haemodialysis treatment

A clinical study was designed to record a wide range of physiological values from patients undergoing haemodialysis treatment in the Renal Unit of the Churchill Hospital in Oxford. Video was recorded for a total of 84 dialysis sessions from 40 patients during the course of 1 year, comprising an overall video recording time of approximately 304.1 h. Reference values were provided by two devices in regular clinical use. The mean absolute error between the heart rate estimates from the camera and the average from two reference pulse oximeters (positioned at the finger and earlobe) was 2.8 beats/min for over 65% of the time the patient was stable. The mean absolute error between the respiratory rate estimates from the camera and the reference values (computed from the Electrocardiogram and a thoracic expansion sensor-chest belt) was 2.1 breaths/min for over 69% of the time for which the reference signals were valid. To increase the robustness of the algorithms, novel methods were devised for cancelling out aliased frequency components caused by the artificial light sources in the hospital, using auto-regressive modelling and pole cancellation. Maps of the spatial distribution of heart rate and respiratory rate information were developed from the coefficients of the auto-regressive models. Most of the periods for which the camera could not produce a reliable heart rate estimate lasted under 3 min, thus opening the possibility to monitor heart rate continuously in a clinical environment.

Parameter Estimation of Hemodynamic Cardiovascular Model for Synthesis of Photoplethysmogram Signal

Synthesis of accurate, personalize photoplethysmogram (PPG) signal is important to interpret, analyze and predict cardiovascular disease progression. Generative models like Generative Adversarial Networks (GANs) can be used for signal synthesis, however, they are difficult to map to the underlying pathophysiological conditions. Hence, we propose a PPG synthesis strategy that has been designed using a cardiovascular system, modeled through the hemodynamic principle. The modeled architecture is composed of a two-chambered heart along with the systemic-pulmonic blood circulation and a baroreflex auto-regulation mechanism to control the arterial blood pressure. The comprehensive PPG signal is synthesized from the cardiac pressure-flow dynamics. In order to tune the modeled cardiac parameters with respect to a measured PPG data, a novel feature extraction strategy has been employed along with the particle swarm optimization heuristics. Our results demonstrate that the synthesized PPG is accurately followed the morphological changes of the ground truth (GT) signal with an RMSE of 0.003 occurring due to the Coronary Artery Disease (CAD) which is caused by an obstruction in the artery.

Ontology-driven weak supervision for clinical entity classification in electronic health records

In the electronic health record, using clinical notes to identify entities such as disorders and their temporality (e.g. the order of an event relative to a time index) can inform many important analyses. However, creating training data for clinical entity tasks is time consuming and sharing labeled data is challenging due to privacy concerns. The information needs of the COVID-19 pandemic highlight the need for agile methods of training machine learning models for clinical notes. We present Trove, a framework for weakly supervised entity classification using medical ontologies and expert-generated rules. Our approach, unlike hand-labeled notes, is easy to share and modify, while offering performance comparable to learning from manually labeled training data. In this work, we validate our framework on six benchmark tasks and demonstrate Trove's ability to analyze the records of patients visiting the emergency department at Stanford Health Care for COVID-19 presenting symptoms and risk factors.

Reinforcement Learning for Clinical Decision Support in Critical Care: Comprehensive Review

BACKGROUND

Decision support systems based on reinforcement learning (RL) have been implemented to facilitate the delivery of personalized care. This paper aimed to provide a comprehensive review of RL applications in the critical care setting.

OBJECTIVE

This review aimed to survey the literature on RL applications for clinical decision support in critical care and to provide insight into the challenges of applying various RL models.

METHODS

We performed an extensive search of the following databases: PubMed, Google Scholar, Institute of Electrical and Electronics Engineers (IEEE), ScienceDirect, Web of Science, Medical Literature Analysis and Retrieval System Online (MEDLINE), and Excerpta Medica Database (EMBASE). Studies published over the past 10 years (2010-2019) that have applied RL for critical care were included.

RESULTS

We included 21 papers and found that RL has been used to optimize the choice of medications, drug dosing, and timing of interventions and to target personalized laboratory values. We further compared and contrasted the design of the RL models and the evaluation metrics for each application.

CONCLUSIONS

RL has great potential for enhancing decision making in critical care. Challenges regarding RL system design, evaluation metrics, and model choice exist. More importantly, further work is required to validate RL in authentic clinical environments.

Patient clustering using dynamic partitioning on correlated and uncertain biomedical data

Background and objectivesHealth professionals look for specific patterns by correlating multiple physiological data in the process of deciding treatments to remedy clinical abnormalities. Biomedical data exhibit some common patterns in the event of identical clinical illnesses. The primary interest of this work is automatic discovery of such patterns in vital sign data (e.g. heart rate, blood pressure) using unsupervised learning and utilising them to identify patients with similar clinical conditions. MethodsA patient clustering method is developed that efficiently isolates patients into multiple groups by discovering dynamic patterns in multi-dimensional vital sign data. A dynamic partitioning algorithm and a patient clustering approach is proposed by introducing a measure namely aggregated instance-wise uncertainty (AIU) computed from multi-dimensional physiological time-series data. ResultsThe developed model is evaluated qualitatively using principal component analysis and silhouette value; and quantitatively in terms of its ability of clustering patients associated with different clinical situations. Experiments are conducted using real-world biomedical data of patients having various clinical conditions. Thee observed accuracy was 82.85% and 91.17% on two experimental datasets comprised of 35 and 34 patients data respectively.The comparisons show that the proposed approached outperformed than other methods in state-of-the-art approach. ConclusionsThe experimental outcomes demonstrate the effectiveness of the proposed approach in discovering distinct patterns with predictive significance.

Supervised machine learning for the prediction of infection on admission to hospital: a prospective observational cohort study

BACKGROUND

Infection diagnosis can be challenging, relying on clinical judgement and non-specific markers of infection. We evaluated a supervised machine learning (SML) algorithm for diagnosing bacterial infection using routinely available blood parameters on presentation to hospital.

METHODS

An SML algorithm was developed to classify cases into infection versus no infection using microbiology records and six available blood parameters (C-reactive protein, white cell count, bilirubin, creatinine, ALT and alkaline phosphatase) from 160203 individuals. A cohort of patients admitted to hospital over a 6 month period had their admission blood parameters prospectively inputted into the SML algorithm. They were prospectively followed up from admission to classify those who fulfilled clinical case criteria for a community-acquired bacterial infection within 72 h of admission using a pre-determined definition. Predictive ability was assessed using receiver operating characteristics (ROC) with cut-off values for optimal sensitivity and specificity explored.

RESULTS

One hundred and four individuals were included prospectively. The median (range) cohort age was 65 (21-98)  years. The majority were female (56/104; 54%). Thirty-six (35%) were diagnosed with infection in the first 72 h of admission. Overall, 44/104 (42%) individuals had microbiological investigations performed. Treatment was prescribed for 33/36 (92%) of infected individuals and 4/68 (6%) of those with no identifiable bacterial infection. Mean (SD) likelihood estimates for those with and without infection were significantly different. The infection group had a likelihood of 0.80 (0.09) and the non-infection group 0.50 (0.29) (P < 0.01; 95% CI: 0.20-0.40). ROC AUC was 0.84 (95% CI: 0.76-0.91).

CONCLUSIONS

An SML algorithm was able to diagnose infection in individuals presenting to hospital using routinely available blood parameters.

Prediction of an Acute Hypotensive Episode During an ICU Hospitalization With a Super Learner Machine-Learning Algorithm

BACKGROUND

Acute hypotensive episodes (AHE), defined as a drop in the mean arterial pressure (MAP) <65 mm Hg lasting at least 5 consecutive minutes, are among the most critical events in the intensive care unit (ICU). They are known to be associated with adverse outcome in critically ill patients. AHE prediction is of prime interest because it could allow for treatment adjustment to predict or shorten AHE.

METHODS

The Super Learner (SL) algorithm is an ensemble machine-learning algorithm that we specifically trained to predict an AHE 10 minutes in advance. Potential predictors included age, sex, type of care unit, severity scores, and time-evolving characteristics such as mechanical ventilation, vasopressors, or sedation medication as well as features extracted from physiological signals: heart rate, pulse oximetry, and arterial blood pressure. The algorithm was trained on the Medical Information Mart for Intensive Care dataset (MIMIC II) database. Internal validation was based on the area under the receiver operating characteristic curve (AUROC) and the Brier score (BS). External validation was performed using an external dataset from Lariboisière hospital, Paris, France.

RESULTS

Among 1151 patients included, 826 (72%) patients had at least 1 AHE during their ICU stay. Using 1 single random period per patient, the SL algorithm with Haar wavelets transform preprocessing was associated with an AUROC of 0.929 (95% confidence interval [CI], 0.899-0.958) and a BS of 0.08. Using all available periods for each patient, SL with Haar wavelets transform preprocessing was associated with an AUROC of 0.890 (95% CI, 0.886-0.895) and a BS of 0.11. In the external validation cohort, the AUROC reached 0.884 (95% CI, 0.775-0.993) with 1 random period per patient and 0.889 (0.768-1) with all available periods and BSs <0.1.

CONCLUSIONS

The SL algorithm exhibits good performance for the prediction of an AHE 10 minutes ahead of time. It allows an efficient, robust, and rapid evaluation of the risk of hypotension that opens the way to routine use.

A bibliometric analysis and visualization of medical data mining research

BACKGROUND

Data mining technology used in the field of medicine has been widely studied by scholars all over the world. But there is little research on medical data mining (MDM) from the perspectives of bibliometrics and visualization, and the research topics and development trends in this field are still unclear.

METHODS

This paper has applied bibliometric visualization software tools, VOSviewer 1.6.10 and CiteSpace V, to study the citation characteristics, international cooperation, author cooperation, and geographical distribution of the MDM.

RESULTS

A total of 1575 documents are obtained, and the most frequent document type is article (1376). SHAN NH is the most productive author, with the highest number of publications of 12, and the Gillies's article (750 times citation) is the most cited paper. The most productive country and institution in MDM is the USA (559) and US FDA (35), respectively. The Journal of Biomedical Informatics, Expert Systems with Applications and Journal of Medical Systems are the most productive journals, which reflected the nature of the research, and keywords "classification (790)" and "system (576)" have the strongest strength. The hot topics in MDM are drug discovery, medical imaging, vaccine safety, and so on. The 3 frontier topics are reporting system, precision medicine, and inflammation, and would be the foci of future research.

CONCLUSION

The present study provides a panoramic view of data mining methods applied in medicine by visualization and bibliometrics. Analysis of authors, journals, institutions, and countries could provide reference for researchers who are fresh to the field in different ways. Researchers may also consider the emerging trends when deciding the direction of their study.

Synthetic PPG generation from haemodynamic model with baroreflex autoregulation: a Digital twin of cardiovascular system

Synthetic data generation has recently emerged as a substitution technique for handling the problem of bulk data needed in training machine learning algorithms. Healthcare, primarily cardiovascular domain is a major area where synthetic physiological data like Photoplethysmogram (PPG), Electrocardiogram (ECG), Phonocardiogram (PCG), etc. are being used to improve accuracy of machine learning algorithm. Conventional synthetic data generation approach using mathematical formulations lack interpretability. Hence, aim of this paper is to generate synthetic PPG signal from a Digital twin platform replicating cardiovascular system. Such system can serve the dual purpose of replicating the physical system, so as to simulate specific `what if' scenarios as well as to generate large scale synthetic data with patho-physiological interpretability. Cardio-vascular Digital twin is modeled with a two chambered heart, haemodynamic equations and a baroreflex based pressure control mechanism to generate blood pressure and flow variations. Synthetic PPG signal is generated from the model for healthy and Atherosclerosis condition. Initial validation of the platform has been made on the basis of efficiency of the platform in clustering Coronary Artery Disease (CAD) and non CAD PPG data by extracting features from the synthetically generated PPG and comparing that with PPG obtained from Physionet data.

Boosting ICD multi-label classification of health records with contextual embeddings and label-granularity

BACKGROUND AND OBJECTIVE

This work deals with clinical text mining, a field of Natural Language Processing applied to biomedical informatics. The aim is to classify Electronic Health Records with respect to the International Classification of Diseases, which is the foundation for the identification of international health statistics, and the standard for reporting diseases and health conditions. Within the framework of data mining, the goal is the multi-label classification, as each health record has assigned multiple International Classification of Diseases codes. We investigate five Deep Learning architectures with a dataset obtained from the Basque Country Health System, and six different perspectives derived from shifts in the input and the output.

METHODS

We evaluate a Feed Forward Neural Network as the baseline and several Recurrent models based on the Bidirectional GRU architecture, putting our research focus on the text representation layer and testing three variants, from standard word embeddings to meta word embeddings techniques and contextual embeddings.

RESULTS

The results showed that the recurrent models overcome the non-recurrent model. The meta word embeddings techniques are capable of beating the standard word embeddings, but the contextual embeddings exhibit as the most robust for the downstream task overall. Additionally, the label-granularity alone has an impact on the classification performance.

CONCLUSIONS

The contributions of this work are a) a comparison among five classification approaches based on Deep Learning on a Spanish dataset to cope with the multi-label health text classification problem; b) the study of the impact of document length and label-set size and granularity in the multi-label context; and c) the study of measures to mitigate multi-label text classification problems related to label-set size and sparseness.

[Comparison of four scoring systems for predicting ICU mortality in patients with sepsis]

OBJECTIVE

To evaluate the value of Sequential Organ Failure Assessment (SOFA), Simplified Acute Physiology Score Ⅱ (SAPS-Ⅱ), Oxford Acute Severity of Illness Score (OASIS) and Logistic Organ Dysfunction System (LODS) scoring systems for predicting ICU mortality in patients with sepsis.

METHODS

We collected the data of a total of 2470 cases of sepsis recorded in the MIMIC-III database from 2001 to 2012 and retrieved the scores of SOFA, SAPS-Ⅱ, OASIS and LODS of the patients within the first day of ICU admission. We compared with the score between the survivors and the non-survivors and analyzed the differences in the area under the ROC curve (AUC) of the 4 scoring systems. Binomial logistic regression was performed to compare the predictive value of the 4 scoring systems for ICU mortality of the patients.

RESULTS

In the 2470 patients with sepsis, 1966 (79.6%) survived and 504 (20.4%) died in the ICU. Compared with the survivors, the non-survivors had a significantly older mean age, higher proportion of patients receiving mechanical ventilation, and higher initial lactate level, creatinine, urea nitrogen, SOFA score, SAPS-Ⅱ score, OASIS score and LODS score (P < 0.05) but with significantly lower body weight and platelet counts (P < 0.05). The AUCs of the SOFA score, SAPS-Ⅱ score, OASIS score, and LODS score were 0.729 (P < 0.001), 0.768 (P < 0.001), 0.757 (P < 0.001), and 0.739 (P < 0.001), respectively. The AUC of SAPS-Ⅱ score was significantly higher than those of SOFA score (Z=3.679, P < 0.001) and LODS score (Z=3.698, P < 0.001) but was comparable with that of OASIS score (Z=1.102, P=0.271); the AUC of OASIS score was significantly higher than that of LODS score (Z=2.172, P=0.030) and comparable with that of SOFA score (Z=1.709, P=0.088). For predicting ICU mortality in patients without septic shock, the AUC of SAPS-Ⅱ score was 0.769 (0.743-0.793), the highest among the 4 scoring systems; in patients with septic shock, the AUCs SAPS-Ⅱ score and OASIS score, 0.768 (0.745-0.791) and 0.762 (0.738-0.785), respectively, were significantly higher than those of the other two scoring systems. Binomial logistic regression showed the corrected SOFA, SAPS-Ⅱ, and OASIS scores, but not LODS scores, were significantly correlated with ICU mortality in patients with sepsis, and their ORs were 1.08 (95% CI: 1.03-1.14, P=0.001), 1.04 (95% CI: 1.02-1.05, P < 0.001), 1.04 (95% CI: 1.01-1.06, P=0.001), 0.96 (95% CI: 0.89-1.04, P=0.350), respectively.

CONCLUSIONS

The scores of SOFA, SAPS-Ⅱ, OASIS, and LODS can predict ICU mortality in patients with sepsis, but SAPS-Ⅱ and OASIS scores have better predictive value than SOFA and LODS scores.

Estimation and Tracking of Blood Pressure Using Routinely Acquired Photoplethysmographic Signals and Deep Neural Networks

Supplemental Digital Content is available in the text.

Continuous tracking of blood pressure in critically ill patients allows rapid identification of clinically important changes and helps guide treatment. Classically, such tracking requires invasive monitoring with its associated risks, discomfort, and low availability outside critical care units. We hypothesized that information contained in a prevalent noninvasively acquired signal (photoplethysmograph: a byproduct of pulse oximetry) combined with advanced machine learning will allow continuous estimation of the patient’s blood pressure.

Do Hyponatremia or Its Underlying Mechanisms Associate With Mortality Risk in Observational Data?

Supplemental Digital Content is available in the text.

Objectives:

Whether unaccounted determinants of hyponatremia, rather than water excess per se, primarily associate with mortality in observational studies has not been explicitly examined.

Design:

Retrospective cohort study of the association between hyponatremia and mortality, stratified by outpatient diuretic use in three strata.

Setting:

An inception cohort of 13,661 critically ill patients from a tertiary medical center.

Measurements and Main Results:

Admission serum sodium concentrations, obtained within 12 hours of admission to the ICU, were the primary exposure. Hyponatremia was associated with 1.82 (95% CI, 1.56–2.11; p < 0.001) higher odds of mortality, yet differed according to outpatient diuretic use (multiplicative interaction between thiazide and serum sodium < 133 mEq/L; p = 0.002). Although hyponatremia was associated with a three-fold higher (odds ratio, 3.11; 95% CI, 2.32–4.17; p < 0.001) odds of mortality among those prescribed loop diuretics, no increase of risk was observed among thiazide diuretic users (odds ratio, 0.87; 95% CI, 0.47–1.51; p = 0.63). When examined as a continuous variable, each one mEq/L higher serum sodium was associated with 8% (odds ratio, 0.92; 95% CI, 0.90–0.94; p < 0.001) lower odds of mortality in loop diuretic patients and 5% (odds ratio, 0.95; 95% CI, 0.93–0.96, p < 0.001) lower in diuretic naïve patients, but was not associated with mortality risk among thiazide users (odds ratio, 0.99; 95% CI, 0.95–1.02; p = 0.45).

Conclusions:

Hyponatremia is not uniformly associated with increased mortality, but differs according to diuretic exposure. Our results suggest that the underlying pathophysiologic factors that lead to water excess, rather water excess itself, account in part for the association between hyponatremia and poor outcomes. More accurate estimations about the association between hyponatremia and outcomes might influence clinical decision-making.

Linking Big Data and Prediction Strategies: Tools, Pitfalls, and Lessons Learned

OBJECTIVES

Modern critical care amasses unprecedented amounts of clinical data-so called "big data"-on a minute-by-minute basis. Innovative processing of these data has the potential to revolutionize clinical prognostics and decision support in the care of the critically ill but also forces clinicians to depend on new and complex tools of which they may have limited understanding and over which they have little control. This concise review aims to provide bedside clinicians with ways to think about common methods being used to extract information from clinical big datasets and to judge the quality and utility of that information.

DATA SOURCES

We searched the free-access search engines PubMed and Google Scholar using the MeSH terms "big data", "prediction", and "intensive care" with iterations of a range of additional potentially associated factors, along with published bibliographies, to find papers suggesting illustration of key points in the structuring and analysis of clinical "big data," with special focus on outcomes prediction and major clinical concerns in critical care.

STUDY SELECTION

Three reviewers independently screened preliminary citation lists.

DATA EXTRACTION

Summary data were tabulated for review.

DATA SYNTHESIS

To date, most relevant big data research has focused on development of and attempts to validate patient outcome scoring systems and has yet to fully make use of the potential for automation and novel uses of continuous data streams such as those available from clinical care monitoring devices.

CONCLUSIONS

Realizing the potential for big data to improve critical care patient outcomes will require unprecedented team building across disparate competencies. It will also require clinicians to develop statistical awareness and thinking as yet another critical judgment skill they bring to their patients' bedsides and to the array of evidence presented to them about their patients over the course of care.

A Characteristic Filtering Method for Pulse Wave Signal Quality Assessment

Pulse wave is an important physiological signal widely used in clinic. In practical applications, the pulse wave recordings are easily corrupted by different interferences. Sometimes, it is very difficult to eliminate the noise by commonly used filtering methods. In this study, we proposed a filtering method based on the characteristics of pulse wave recordings to remove the noisy outliers. Firstly, five characteristics, short-term energy (SE), ascending intensity difference (AID), descending intensity difference (DID), ascending time difference (ATD), and descending time difference (DTD), were chosen as metrics and calculated from cardiac pulse wave. Then the median lines of the five metrics were obtained using a median filter, respectively. An acceptable value range around the median line of each metric was set based on histogram distribution analysis and was used to examine pulse wave recordings cardiac-cycle-by-cycle. For each cardiac cycle, when one or more of its five characteristic values exceed(s) the acceptable range, the pulse wave recording segment was discarded from further analysis. With this proposed method, the noisy outliers could be efficiently identified from the pulse wave recordings. This suggests that the proposed preprocessing method would be useful in improving the assessment performance of pulse-wave-based clinical applications. Additionally, the method might also be extended used in other physiological signals pre-processing, such as ECG, blood pressure wave, etc.

Expert-augmented machine learning

Machine learning is increasingly used across fields to derive insights from data, which further our understanding of the world and help us anticipate the future. The performance of predictive modeling is dependent on the amount and quality of available data. In practice, we rely on human experts to perform certain tasks and on machine learning for others. However, the optimal learning strategy may involve combining the complementary strengths of humans and machines. We present expert-augmented machine learning, an automated way to automatically extract problem-specific human expert knowledge and integrate it with machine learning to build robust, dependable, and data-efficient predictive models.

Machine learning is proving invaluable across disciplines. However, its success is often limited by the quality and quantity of available data, while its adoption is limited by the level of trust afforded by given models. Human vs. machine performance is commonly compared empirically to decide whether a certain task should be performed by a computer or an expert. In reality, the optimal learning strategy may involve combining the complementary strengths of humans and machines. Here, we present expert-augmented machine learning (EAML), an automated method that guides the extraction of expert knowledge and its integration into machine-learned models. We used a large dataset of intensive-care patient data to derive 126 decision rules that predict hospital mortality. Using an online platform, we asked 15 clinicians to assess the relative risk of the subpopulation defined by each rule compared to the total sample. We compared the clinician-assessed risk to the empirical risk and found that, while clinicians agreed with the data in most cases, there were notable exceptions where they overestimated or underestimated the true risk. Studying the rules with greatest disagreement, we identified problems with the training data, including one miscoded variable and one hidden confounder. Filtering the rules based on the extent of disagreement between clinician-assessed risk and empirical risk, we improved performance on out-of-sample data and were able to train with less data. EAML provides a platform for automated creation of problem-specific priors, which help build robust and dependable machine-learning models in critical applications.

Proton Pump Inhibitor Use Before ICU Admission Is Not Associated With Mortality of Critically Ill Patients

Some studies have shown that the long-term use of proton pump inhibitors (PPIs) is associated with many adverse events that may increase mortality; however, the relationship between premorbid PPI use and in-hospital mortality has yet to be validated in critically ill patients. Therefore, we performed this study to determine whether the preadmission use of PPIs is associated with mortality in patients admitted to the intensive care unit. This was a retrospective study with a large and freely accessible database in critical-care medicine (the Multiparameter Intelligent Monitoring in Intensive Care III project). The clinical data and outcomes of 17 473 patients, consisting of 1895 in the PPI group, 514 in the H₂ -receptor antagonist group, and 15 064 control subjects, were collected during their hospital stay. The study outcome was in-hospital mortality. A total of 17 473 patients were included in our study. PPI use was associated with significantly increased in-hospital mortality in the original model without adjustment for any parameters (odds ratio 1.19; 95%CI 1.03-1.38; P = .02). However, after adjustments had been made for age, sex, Elixhauser score, Simplified Acute Physiology Score, laboratory results, vasopressor use, ventilator use, and other parameters, PPIs were not associated with significantly increased in-hospital mortality (odds ratio 1.04; 95%CI 0.87-1.26; P = .614). In the subgroup analysis among patients with renal or liver disease, we still found that PPIs were not associated with a significant increase in in-hospital mortality. We found no association between PPI use before ICU admission and increased in-hospital mortality in critically ill patients compared with control subjects.

Learning unsupervised contextual representations for medical synonym discovery

Objectives

An important component of processing medical texts is the identification of synonymous words or phrases. Synonyms can inform learned representations of patients or improve linking mentioned concepts to medical ontologies. However, medical synonyms can be lexically similar (“dilated RA” and “dilated RV”) or dissimilar (“cerebrovascular accident” and “stroke”); contextual information can determine if 2 strings are synonymous. Medical professionals utilize extensive variation of medical terminology, often not evidenced in structured medical resources. Therefore, the ability to discover synonyms, especially without reliance on training data, is an important component in processing training notes. The ability to discover synonyms from models trained on large amounts of unannotated data removes the need to rely on annotated pairs of similar words. Models relying solely on non-annotated data can be trained on a wider variety of texts without the cost of annotation, and thus may capture a broader variety of language.

Materials and Methods

Recent contextualized deep learning representation models, such as ELMo (Peters et al., 2019) and BERT, (Devlin et al. 2019) have shown strong improvements over previous approaches in a broad variety of tasks. We leverage these contextualized deep learning models to build representations of synonyms, which integrate the context of surrounding sentence and use character-level models to alleviate out-of-vocabulary issues. Using these models, we perform unsupervised discovery of likely synonym matches, which reduces the reliance on expensive training data.

Results

We use the ShARe/CLEF eHealth Evaluation Lab 2013 Task 1b data to evaluate our synonym discovery method. Comparing our proposed contextualized deep learning representations to previous non-neural representations, we find that the contextualized representations show consistent improvement over non-contextualized models in all metrics.

Conclusions

Our results show that contextualized models produce effective representations for synonym discovery. We expect that the use of these representations in other tasks would produce similar gains in performance.

A classification model for prediction of clinical severity level using qSOFA medical score

Purpose

The purpose of this study is to develop an efficient prediction model using vital signs and standard medical score systems, which predicts the clinical severity level of the patient in advance based on the quick sequential organ failure assessment (qSOFA) medical score method.

Design/methodology/approach

To predict the clinical severity level of the patient in advance, the authors have formulated a training dataset that is constructed based on the qSOFA medical score method. Further, along with the multiple vital signs, different standard medical scores and their correlation features are used to build and improve the accuracy of the prediction model. It is made sure that the constructed training set is suitable for the severity level prediction because the formulated dataset has different clusters each corresponding to different severity levels according to qSOFA score.

Findings

From the experimental result, it is found that the inclusion of the standard medical scores and their correlation along with multiple vital signs improves the accuracy of the clinical severity level prediction model. In addition, the authors showed that the training dataset formulated from the temporal data (which includes vital signs and medical scores) based on the qSOFA medical scoring system has the clusters which correspond to each severity level in qSOFA score. Finally, it is found that RAndom k-labELsets multi-label classification performs better prediction of severity level compared to neural network-based multi-label classification.

Originality/value

This paper helps in identifying patient' clinical status.

Early prediction of ICU readmissions using classification algorithms

CONTEXT

Determining which patients are ready for discharge from an Intensive Care Unit (ICU) presents a huge challenge, as ICU readmissions are associated with several negative outcomes such as increased mortality, length of stay, and cost compared to those patients who are not readmitted during their hospital stay. For these reasons, enhancing risk stratification in order to identify patients at high risk of clinical deterioration might benefit and improve the outcomes of critically ill hospitalized patients. Existing work on predicting ICU readmissions relies on information available at the time of discharge, however, in order to be more useful and to prevent complications, predictions need to be made earlier.

GOALS

In this work, we investigate the hypothesis that the basal characteristics and information collected at the time of the patient's admission can enable accurate predictions of ICU readmission.

MATERIALS AND METHODS

We analyzed an anonymized dataset of 11,805 adult patients from three ICUs in a Brazilian university hospital. After excluding 1879 patients who died during their first ICU admission, our final dataset contained 9,926 patients. Of these, 658 patients (6.6%) had been readmitted to the ICU. The original dataset had 185 attributes, including demographics, length of stay prior to ICU admission, comorbidities, severity indexes, interventions, organ support care during ICU stay and laboratory results. The problem of predicting ICU readmissions was modeled as a binary classification task. We tested eight classification algorithms (including Bayesian algorithms, decision trees, rule-based, and ensemble methods) over different sets of attributes and evaluated their results based on six metrics.

RESULTS

Predictions made solely based on the attributes collected at the admission are highly accurate. Their quality in terms of prediction is no different from predictions made using the complete set of attributes for our dataset and for a subset of attributes selected by a feature selection method. Furthermore, our AUROC score of 0.91 (95% CI [0.89,0.92]) is higher than existing results published in the literature for other datasets.

DISCUSSION AND CONCLUSION

The results confirm our hypothesis. Our findings suggest that early markers can be used to anticipate patients at high risk of clinical deterioration after ICU discharge.

Brief introduction of medical database and data mining technology in big data era

Data mining technology can search for potentially valuable knowledge from a large amount of data, mainly divided into data preparation and data mining, and expression and analysis of results. It is a mature information processing technology and applies database technology. Database technology is a software science that researches manages, and applies databases. The data in the database are processed and analyzed by studying the underlying theory and implementation methods of the structure, storage, design, management, and application of the database. We have introduced several databases and data mining techniques to help a wide range of clinical researchers better understand and apply database technology.

Annotating and detecting phenotypic information for chronic obstructive pulmonary disease

Objectives

Chronic obstructive pulmonary disease (COPD) phenotypes cover a range of lung abnormalities. To allow text mining methods to identify pertinent and potentially complex information about these phenotypes from textual data, we have developed a novel annotated corpus, which we use to train a neural network-based named entity recognizer to detect fine-grained COPD phenotypic information.

Materials and methods

Since COPD phenotype descriptions often mention other concepts within them (proteins, treatments, etc.), our corpus annotations include both outermost phenotype descriptions and concepts nested within them. Our neural layered bidirectional long short-term memory conditional random field (BiLSTM-CRF) network firstly recognizes nested mentions, which are fed into subsequent BiLSTM-CRF layers, to help to recognize enclosing phenotype mentions.

Results

Our corpus of 30 full papers (available at: http://www.nactem.ac.uk/COPD) is annotated by experts with 27 030 phenotype-related concept mentions, most of which are automatically linked to UMLS Metathesaurus concepts. When trained using the corpus, our BiLSTM-CRF network outperforms other popular approaches in recognizing detailed phenotypic information.

Discussion

Information extracted by our method can facilitate efficient location and exploration of detailed information about phenotypes, for example, those specifically concerning reactions to treatments.

Conclusion

The importance of our corpus for developing methods to extract fine-grained information about COPD phenotypes is demonstrated through its successful use to train a layered BiLSTM-CRF network to extract phenotypic information at various levels of granularity. The minimal human intervention needed for training should permit ready adaption to extracting phenotypic information about other diseases.

Estimation of the Blood Pressure Waveform using Electrocardiography

This work presents a modelling approach to accurately predict the blood pressure (BP) waveform time series from a single input signal. A nonlinear autoregressive model with exogenous input (NARX) is implemented using artificial neural networks and trained on Electrocardiography (ECG) signals to predict the BP waveform. The efficacy of the model is demonstrated using the MIMIC II database. The proposed method can accurately estimate systolic and diastolic BP. The NARX model together with ECG measurement allows continuous monitoring of BP, enables the estimation of other physiological measurements, such as the cardiac output, and provides more insight on the patient health condition.

Forecasting Hypotension during Vasopressor Infusion via Time Series Analysis

For optimal management of hypotension during continuous vasopressor infusion, this study investigated two forecasting models, logistic regression (LR) and auto-regressive (AR) models, to predict sustained hypotension episodes (SHEs) in the ICU, before the SHE occurred. Two investigational models were compared to a simple threshold detector, which alerts whenever the BP is less than the specific hypotension threshold. Datasets were collected from 207 patients treated for a variety of clinical indications in two different hospitals (Hospital 1 & 2). For the 60 mmHg hypotension threshold, LR model predicted SHEs an average of 7.0 min before (Hospital 1) and 2.5 min before (Hospital 2), and the AR model predicted SHEs 10.5 min and 2.0 min before (Hospital 1 and 2 respectively). Both were significantly better than the threshold method and without higher false alarm rates. The AR model offered the flexibility to predict for different hypotension thresholds.

PIC, a paediatric-specific intensive care database

PIC (Paediatric Intensive Care) is a large paediatric-specific, single-centre, bilingual database comprising information relating to children admitted to critical care units at a large children's hospital in China. The database is deidentified and includes vital sign measurements, medications, laboratory measurements, fluid balance, diagnostic codes, length of hospital stays, survival data, and more. The data are publicly available after registration, which includes completion of a training course on research with human subjects and signing of a data use agreement mandating responsible handling of the data and adherence to the principle of collaborative research. Although the PIC can be considered an extension of the widely used MIMIC (Medical Information Mart for Intensive Care) database in the field of paediatric critical care, it has many unique characteristics and can support database-based academic and industrial applications such as machine learning algorithms, clinical decision support tools, quality improvement initiatives, and international data sharing.

A Predictive Model for Acute Respiratory Distress Syndrome Mortality Using Red Cell Distribution Width

Methods

This observational retrospective cohort study includes 318 ARDS patients extracted from an ICU database between the years of 2001 and 2008. Clinical factors including age, gender, comorbidity score, Sequential Organ Failure Assessment (SOFA) score, and PaO₂/FiO₂ ratio were chosen for the base model to predict ICU mortality. The RDW value at the time of ARDS diagnosis was added to the base model to determine if it improved its predictive ability.

Results

318 subjects were included; 113 (36%) died in the ICU. AUC for the base model without RDW was 0.76, and 0.78 following the addition of RDW [p=0.048]. The NRI was 0.46 (p=0.001), indicating that, in 46% of patients, the predictive probability of the model was improved by the inclusion of RDW.

Conclusions

Adding RDW at time of ARDS diagnosis improved discrimination in a model using 4 clinical factors to predict ICU mortality.

A nomogram for predicting the risk of sepsis in patients with acute cholangitis

OBJECTIVE

Sepsis is a serious complication of acute cholangitis. We aimed to establish a nomogram for predicting the probability of sepsis in patients with acute cholangitis.

METHODS

Subjects were patients with acute cholangitis in the Medical Information Mart for Intensive Care database. Extraneous variables were excluded based on stepwise regression. The nomogram was established using logistic regression.

RESULTS

The predictive model comprised five variables: age (odds ratio [OR]: 1.03, 95% confidence interval [CI]: 1.01–1.04), ventilator-support time (OR: 1.004, 95% CI: 1.001–1.008), diabetes (OR: 10.74, 95% CI: 2.80–70.57), coagulopathy (OR: 2.92, 95% CI: 1.83–4.73) and systolic blood pressure (OR: 0.62, 95% CI: 0.41–0.93). The areas under the receiver operating characteristic curve of the nomogram for the training and validation sets were 0.700 and 0.647, respectively. The Hosmer–Lemeshow goodness-of-fit test revealed high concordance between the predicted and observed probabilities for both the training and validation sets. The calibration plot also demonstrated good agreement between the predicted and observed outcomes for both the training and validation sets.

CONCLUSIONS

We developed and validated a risk-prediction model for sepsis in patients with acute cholangitis. Our results will be helpful for preventing sepsis in patients with acute cholangitis.

Energy-Efficient ECG Signal Compression for User Data Input in Cyber-Physical Systems by Leveraging Empirical Mode Decomposition

Human physiological data are naturalistic and objective user data inputs for a great number of cyber-physical systems (CPS). Electrocardiogram (ECG) as a widely used physiological golden indicator for certain human state and disease diagnosis is often used as user data input for various CPS such as medical CPS and human–machine interaction. Wireless transmission and wearable technology enable long-term continuous ECG data acquisition for human–CPS interaction; however, these emerging technologies bring challenges of storing and wireless transmitting huge amounts of ECG data, leading to energy efficiency issue of wearable sensors. ECG signal compression technique provides a promising solution for these challenges by decreasing ECG data size. In this study, we develop the first scheme of leveraging empirical mode decomposition (EMD) on ECG signals for sparse feature modeling and compression and further propose a new ECG signal compression framework based on EMD constructed feature dictionary. The proposed method features in compressing ECG signals using a very limited number of feature bases with low computation cost, which significantly improves the compression performance and energy efficiency. Our method is validated with the ECG data from MIT-BIH arrhythmia database and compared with existing methods. The results show that our method achieves the compression ratio (CR) of up to 164 with the root mean square error (RMSE) of 3.48% and the average CR of 88.08 with the RMSE of 5.66%, which is more than twice of the average CR of the state-of-the-art methods with similar recovering error rate of around 5%. For diagnostic distortion perspective, our method achieves high QRS detection performance with the sensitivity (SE) of 99.8% and the specificity (SP) of 99.6%, which shows that our ECG compression method can preserve almost all the QRS features and have no impact on the diagnosis process. In addition, the energy consumption of our method is only 30% of that of other methods when compared under the same recovering error rate.

Multiple Physiological Signals Fusion Techniques for Improving Heartbeat Detection: A Review

This paper presents a review of the techniques found in the literature that aim to achieve a robust heartbeat detection from fusing multi-modal physiological signals (e.g., electrocardiogram (ECG), blood pressure (BP), artificial blood pressure (ABP), stroke volume (SV), photoplethysmogram (PPG), electroencephalogram (EEG), electromyogram (EMG), and electrooculogram (EOG), among others). Techniques typically employ ECG, BP, and ABP, of which usage has been shown to obtain the best performance under challenging conditions. SV, PPG, EMG, EEG, and EOG signals can help increase performance when included within the fusion. Filtering, signal normalization, and resampling are common preprocessing steps. Delay correction between the heartbeats obtained over some of the physiological signals must also be considered, and signal-quality assessment to retain the best signal/s must be considered as well. Fusion is usually accomplished by exploiting regularities in the RR intervals; by selecting the most promising signal for the detection at every moment; by a voting process; or by performing simultaneous detection and fusion using Bayesian techniques, hidden Markov models, or neural networks. Based on the results of the review, guidelines to facilitate future comparison of the performance of the different proposals are given and promising future lines of research are pointed out.

[Construction of multi-parameter emergency database and preliminary application research]

The analysis of big data in medical field cannot be isolated from the high quality clinical database, and the construction of first aid database in our country is still in the early stage of exploration. This paper introduces the idea and key technology of the construction of multi-parameter first aid database. By combining emergency business flow with information flow, an emergency data integration model was designed with reference to the architecture of the Medical Information Mart for Intensive Care III (MIMIC-III), created by Computational Physiology Laboratory of Massachusetts Institute of Technology (MIT), and a high-quality first-aid database was built. The database currently covers 22 941 medical records for 19 814 different patients from May 2015 to October 2017, including relatively complete information on physiology, biochemistry, treatment, examination, nursing, etc. And based on the database, the first First-Aid Big Data Datathon event, which 13 teams from all over the country participated in, was launched. The First-Aid database provides a reference for the construction and application of clinical database in China. And it could provide powerful data support for scientific research, clinical decision making and the improvement of medical quality, which will further promote secondary analysis of clinical data in our country.

LiSep LSTM: A Machine Learning Algorithm for Early Detection of Septic Shock

Sepsis is a major health concern with global estimates of 31.5 million cases per year. Case fatality rates are still unacceptably high, and early detection and treatment is vital since it significantly reduces mortality rates for this condition. Appropriately designed automated detection tools have the potential to reduce the morbidity and mortality of sepsis by providing early and accurate identification of patients who are at risk of developing sepsis. In this paper, we present “LiSep LSTM”; a Long Short-Term Memory neural network designed for early identification of septic shock. LSTM networks are typically well-suited for detecting long-term dependencies in time series data. LiSep LSTM was developed using the machine learning framework Keras with a Google TensorFlow back end. The model was trained with data from the Medical Information Mart for Intensive Care database which contains vital signs, laboratory data, and journal entries from approximately 59,000 ICU patients. We show that LiSep LSTM can outperform a less complex model, using the same features and targets, with an AUROC 0.8306 (95% confidence interval: 0.8236, 0.8376) and median offsets between prediction and septic shock onset up to 40 hours (interquartile range, 20 to 135 hours). Moreover, we discuss how our classifier performs at specific offsets before septic shock onset, and compare it with five state-of-the-art machine learning algorithms for early detection of sepsis.

A Cascaded Convolutional Neural Network for Assessing Signal Quality of Dynamic ECG

Motion artifacts and myoelectrical noise are common issues complicating the collection and processing of dynamic electrocardiogram (ECG) signals. Recent signal quality studies have utilized a binary classification metric in which ECG samples are determined to either be clean or noisy. However, the clinical use of dynamic ECGs requires specific noise level classification for varying applications. Conventional signal processing methods, including waveform discrimination, are limited in their ability to remove motion artifacts and myoelectrical noise from dynamic ECGs. As such, a novel cascaded convolutional neural network (CNN) is proposed and demonstrated for application to the five-classification problem (low interference, mild motion artifacts, mild myoelectrical noise, severe motion artifacts, and severe myoelectrical noise). Specifically, this study finally categorizes dynamic ECG signals into three levels (low, mild, and severe) using the proposed CNN to meet clinical requirements. The network includes two components, the first of which was used to distinguish signal interference types, while the second was used to distinguish signal interference levels. This model does not require feature engineering, includes powerful nonlinear mapping capabilities, and is robust to varying noise types. Experimental data are composed of private dataset and public dataset, which were acquired from 90,000 four-second dynamic ECG signals and MIT-BIH Arrhythmia database, respectively. Experimental results produced an overall recognition rate of 92.7% on private dataset and 91.8% on public dataset. These results suggest the proposed technique to be a valuable new tool for dynamic ECG auxiliary diagnosis.

The Univariate Flagging Algorithm (UFA): An interpretable approach for predictive modeling

In many data classification problems, a number of methods will give similar accuracy. However, when working with people who are not experts in data science such as doctors, lawyers, and judges among others, finding interpretable algorithms can be a critical success factor. Practitioners have a deep understanding of the individual input variables but far less insight into how they interact with each other. For example, there may be ranges of an input variable for which the observed outcome is significantly more or less likely. This paper describes an algorithm for automatic detection of such thresholds, called the Univariate Flagging Algorithm (UFA). The algorithm searches for a separation that optimizes the difference between separated areas while obtaining a high level of support. We evaluate its performance using six sample datasets and demonstrate that thresholds identified by the algorithm align well with published results and known physiological boundaries. We also introduce two classification approaches that use UFA and show that the performance attained on unseen test data is comparable to or better than traditional classifiers when confidence intervals are considered. We identify conditions under which UFA performs well, including applications with large amounts of missing or noisy data, applications with a large number of inputs relative to observations, and applications where incidence of the target is low. We argue that ease of explanation of the results, robustness to missing data and noise, and detection of low incidence adverse outcomes are desirable features for clinical applications that can be achieved with relatively simple classifier, like UFA.

Preoperative Blood Pressure Complexity Indices as a Marker for Frailty in Patients Undergoing Cardiac Surgery

OBJECTIVE

Frailty, a state of decreased physiological reserve, increases the risk of adverse outcomes. There is no standard tool for frailty during perioperative period. Autonomic dysfunction, an underlying process in frailty, could result in hemodynamic fluctuations. Complexity, the physiological adaptability of a system can quantify these fluctuations. The authors hypothesized that complexity could be a marker for frailty and explored their relationship in cardiac surgical patients.

DESIGN

Prospective, observational study.

SETTING

Single-center teaching hospital.

PARTICIPANTS

Three hundred and sixty-four adult patients undergoing cardiac surgery.

INTERVENTION

None.

MEASUREMENTS AND MAIN RESULTS

Preoperative beat-to-beat systolic arterial pressure (SAP) and mean arterial pressure (MAP) time series were obtained. Complexity indices were calculated using multiscale entropy (MSE) analysis. Frailty was assessed from: age >70 years, body mass index <18.5, hematocrit <35%, albumin <3.4 g/dL, and creatinine >2.0 mg/dL. The association between complexity indices and frailty was explored by logistic regression and predictive ability by C-statistics. In total, 190 (52%) patients had frailty. The complexity index (MSE_Σ) median (quartile 1, quartile 3) of SAP and MAP time series decreased significantly in frail patients (SAP: 8.32 [7.27, 9.24] v 9.13 [8.00, 9.72], p < 0.001 and MAP: 8.56 [7.56; 9.27] v 9.18 [8.26; 9.83], p < 0.001). MSE _Σ demonstrated a fair predictive ability of frailty (C-statistic: SAP 0.62 and MAP 0.64).

CONCLUSION

Preoperative BP complexity indices correlate and predict frailty. Impaired autonomic control is the underlying mechanism to explain this finding. A simple automated measure of preoperative BP complexity in the surgeon's office has the potential to reliably assess frailty.