Deep-Learning based Sleep Apnea Detection using SpO2 and Pulse Rate.

This work presents automated apnea event de-tection using blood oxygen saturation (SpO2) and pulse rate (PR), conveniently recorded with a pulse oximeter. A large, diverse cohort of patients (n=8068, age≥40 years) from the sleep heart health study dataset with annotated sleep events have been employed in this study. A deep-learning model is trained to detect apnea in successive 30 s epochs and performances are assessed on two independent sub-cohorts of test data. The proposed algorithm showcases the highest test performance of 90.4 % area under the receiver operating characteristic curve and 58.9% area under the precision-recall curve for epoch-based apnea detection. Additionally, the model consistently performs well across various apnea subtypes, with the highest sensitivity of 93.4 % for obstructive apnea detection followed by 90.5 % for central apnea and 89.1 % for desaturation associated hypopnea. Overall, the proposed algorithm provides a robust and sensitive approach for sleep apnea event detection using a noninvasive pulse oximeter sensor. Clinical Relevance - The study establishes high sensitivity for automated epoch-based apnea detection across a diverse study cohort with various comorbidities using simply a pulse oximeter. This highly cost-effective approach could also enable convenient sleep and health monitoring over long-term.

Towards Remote Continuous Monitoring of Cytokine Release Syndrome.

Cytokine release syndrome (CRS) is a noninfec-tious systemic inflammatory response syndrome condition and a principle severe adverse event common in oncology patients treated with immunotherapies. Accurate monitoring and timely prediction of CRS severity remain a challenge. This study presents an XGBoost-based machine learning algorithm for forecasting CRS severity (no CRS, mild- and severe-CRS classes) in the 24 hours following the time of prediction utilizing the common vital signs and Glasgow coma scale (GCS) questionnaire inputs. The CRS algorithm was developed and evaluated on a cohort of patients (n=1,139) surgically treated for neoplasm with no ICD9 codes for infection or sepsis during a collective 9,892 patient-days of monitoring in ICU settings. Different models were trained with unique feature sets to mimic practical monitoring environments where different types of data availability will exist. The CRS models that incorporated all time series features up to the prediction time showcased a micro-average area under curve (AUC) statistic for the receiver operating characteristic curve (ROC) of 0.94 for the 3 classes of CRS grades. Models developed on a second cohort requiring data within the 24 hours preceding prediction time showcased a relatively lower 0.88 micro-average AUROC as these models did not benefit from implicit information in the data availability. Systematic removal of blood pressure and/or GCS inputs revealed significant decreases (p<0.05) in model performances that confirm the importance of such features for CRS prediction. Accurate CRS prediction and timely intervention can reverse CRS adverse events and maximize the benefit of immunotherapies in oncology patients.

Pulse Rate Guided Oxygen Saturation Monitoring Using a Wearable Armband Sensor.

Continuous clinical grade measurement of SpO2 in out-of-hospital settings remains a challenge despite the widespread use of photoplethysmography (PPG) based wearable devices for health and wellness applications. This article presents two SpO2 algorithms: PRR (pulse rate derived ratio-of-ratios) and GPDR (green-assisted peak detection ratio-of-ratios), that utilize unique pulse rate frequency estimations to isolate the pulsatile (AC) component of red and infrared PPG signals and derive SpO2 measurements. The performance of the proposed SpO2 algorithms are evaluated using an upper-arm wearable device derived green, red, and infrared PPG signals, recorded in both controlled laboratory settings involving healthy subjects (n=36) and an uncontrolled clinic application involving COVID-19 patients (n=52). GPDR exhibits the lowest root mean square error (RMSE) of 1.6±0.6% for a respiratory exercise test, 3.6 ±1.0% for a standard hypoxia test, and 2.2±1.3% for an uncontrolled clinic use-case. In contrast, PRR provides relatively higher error but with greater coverage overall. Mean error across all combined datasets were 0.2±2.8% and 0.3±2.4% for PRR and GPDR respectively. Both SpO2 algorithms achieve great performance of low error with high coverage on both uncontrolled clinic and controlled laboratory conditions.

Influence of Study Composition on the Efficacy of Sleep Detection Using Actigraphy.

Wearable actigraphy sensors have been useful tools for unobtrusive monitoring of sleep. The influence of the composition and characteristics of study groups such as normal sleep versus sleep disorders affecting the efficacy of sleep assessment using actigraphy has not been fully examined. In this study, we present multi-variate sleep models using actigraphy features obtained from wrist-worn sensors and evaluate the efficacy of sleep detection compared to the overnight polysomnography from two unique datasets: overnight actigraphy recordings in a control population of young healthy individuals (n=31) and 24-hour actigraphy recordings in a more heterogeneous population (n=27) comprised of normal and abnormal sleepers. We evaluate the performance of actigraphy derived logistic regression (LR) and random forest (RF) sleep models for both intra-dataset and inter-dataset training and cross-validation. Both the LR and RF sleep models for the healthy sleep dataset show an area under the receiver operating characteristic (AUROC) of 0.85±0.02 in the control sleep dataset among 50 random splits of training and testing evaluations. We find the AUROC performance from the heterogeneous sleep dataset involving sleep disorders to be relatively lower as 0.74±0.05 and 0.80±0.03 for LR and RF sleep models, respectively. Optimal sleep models trained using heterogeneous datasets perform very well when tested with the normal sleep dataset producing accuracy of ∼92%. Our study supports that using a more diverse training set benefits the sleep classifier model to be more generalizable for both healthy and abnormal sleepers.

Novel Dynamic Prediction of Daily Patient Discharge in Acute and Critical Care.

Determining when a patient can be discharged from a care setting is critical to optimize the utilization and delivery of timely care. Furthermore, timely discharge can lead to better clinical outcomes by effectively mitigating the prolonged length of stay in a care environment. This paper presents a novel algorithm for the prediction of likelihood of patient discharge within the next 24 or 48 hours from acute or critical care environments on a daily basis. Continuous patient monitoring and health data obtained from acute hospital at home environment (n=303 patients) and a critical care unit environment (n=9,520 patients) are retrospectively used to train, validate and test numerous machine learning models for dynamic daily predictions of patients discharge. In the acute hospital at home environment, the area under the receiver operating characteristic (AUROC) curve performance of a top XGBoost model was 0.816 ± 0.025 and 0.758 ± 0.029 for daily discharge prediction within 24 hours and 48 hours respectively. Similar independent prediction models from the critical care environment resulted in relatively a lower AUROC for likewise predicting daily patient discharge. Overall, the results demonstrate the efficacy and utility of our novel algorithm for dynamic predictions of daily patient discharge in both acute- and critical care healthcare settings.

Novel COVID-19 Screening Using Cough Recordings of A Mobile Patient Monitoring System.

Since the COVID-19 pandemic began, research has shown promises in building COVID-19 screening tools using cough recordings as a convenient and inexpensive alternative to current testing techniques. In this paper, we present a novel and fully automated algorithm framework for cough extraction and COVID-19 detection using a combination of signal processing and machine learning techniques. It involves extracting cough episodes from audios of a diverse real-world noisy conditions and then screening for the COVID-19 infection based on the cough characteristics. The proposed algorithm was developed and evaluated using self-recorded cough audios collected from COVID-19 patients monitored by Biovitals® Sentinel remote patient management platform and publicly available datasets of various sound recordings. The proposed algorithm achieves a duration Area Under Receiver Operating Characteristic curve (AUROC) of 98.6% in the cough extraction task and a mean cross-validation AUROC of 98.1% in the COVID-19 classification task. These results demonstrate high accuracy and robustness of the proposed algorithm as a fast and easily accessible COVID-19 screening tool and its potential to be used for other cough analysis applications.

Novel Continuous Respiratory Rate Monitoring Using an Armband Wearable Sensor.

Photoplethysmography (PPG) and accelerometer (ACC) are commonly integrated into wearable devices for continuous unobtrusive pulse rate and activity monitoring of individuals during daily life. However, obtaining continuous and clinically accurate respiratory rate measurements using such wearable sensors remains a challenge. This article presents a novel algorithm for estimation of respiration rate (RR) using an upper-arm worn wearable device by deriving multiple respiratory surrogate signals from PPG and ACC sensing. This RR algorithm is retrospectively evaluated on a controlled respiratory clinical testing dataset from 38 subjects with simultaneously recorded wearable sensor data and a standard capnography monitor as an RR reference. The proposed RR method shows great performance and robustness in determining RR measurements over a wide range of 4-59 brpm with an overall bias of -1.3 brpm, mean absolute error (MAE) of 2.7±1.6 brpm, and a meager outage of 0.3±1.2%, while a standard PPG Smart Fusion method produces a bias of -3.6 brpm, an MAE of 5.5±3.1 brpm, and an outage of 0.7±2.5% for direct comparison. In addition, the proposed algorithm showed no significant differences (p=0.63) in accurately determining RR values in subjects with darker skin tones, while the RR performance of the PPG Smart Fusion method is significantly (P<0.001) affected by the darker skin pigmentation. This study demonstrates a highly accurate RR algorithm for unobtrusive continuous RR monitoring using an armband wearable device.

Automatic and Robust Identification of Spontaneous Coughs from COVID-19 Patients.

Cough is one of the most common symptoms of COVID-19. It is easily recorded using a smartphone for further analysis. This makes it a great way to track and possibly identify patients with COVID. In this paper, we present a deep learning-based algorithm to identify whether a patient's audio recording contains a cough for subsequent COVID screening. More generally, cough identification is valuable for the remote monitoring and tracking of infections and chronic conditions. Our algorithm is validated on our novel dataset in which COVID-19 patients were instructed to volunteer natural coughs. The validation dataset consists of real patient cough and no cough audio. It was supplemented by files without cough from publicly available datasets that had cough-like sounds including: throat clearing, snoring, etc. Our algorithm had an area under receiver operating characteristic curve statistic of 0.977 on a validation set when making a cough/no cough determination. The specificity and sensitivity of the model on a reserved test set, at a threshold set by the validation set, was 0.845 and 0.976. This algorithm serves as a fundamental step in a larger cascading process to monitor, extract, and analyze COVID-19 patient coughs to detect the patient's health status, symptoms, and potential for deterioration.

Observational study on wearable biosensors and machine learning-based remote monitoring of COVID-19 patients.

Patients infected with SARS-CoV-2 may deteriorate rapidly and therefore continuous monitoring is necessary. We conducted an observational study involving patients with mild COVID-19 to explore the potentials of wearable biosensors and machine learning-based analysis of physiology parameters to detect clinical deterioration. Thirty-four patients (median age: 32 years; male: 52.9%) with mild COVID-19 from Queen Mary Hospital were recruited. The mean National Early Warning Score 2 (NEWS2) were 0.59 ± 0.7. 1231 manual measurement of physiology parameters were performed during hospital stay (median 15 days). Physiology parameters obtained from wearable biosensors correlated well with manual measurement including pulse rate (r = 0.96, p < 0.0001) and oxygen saturation (r = 0.87, p < 0.0001). A machine learning-derived index reflecting overall health status, Biovitals Index (BI), was generated by autonomous analysis of physiology parameters, symptoms, and other medical data. Daily BI was linearly associated with respiratory tract viral load (p < 0.0001) and NEWS2 (r = 0.75, p < 0.001). BI was superior to NEWS2 in predicting clinical worsening events (sensitivity 94.1% and specificity 88.9%) and prolonged hospitalization (sensitivity 66.7% and specificity 72.7%). Wearable biosensors coupled with machine learning-derived health index allowed automated detection of clinical deterioration.

Compressive Sampling Based Multi-Spectrum Deep Learning for Sub-Nyquist Pacemaker ECG Analysis.

Automatic electrocardiogram (ECG) analysis for pacemaker patients is crucial for monitoring cardiac conditions and the effectiveness of cardiac resynchronization treatment. However, under the condition of energy-saving remote monitoring, the low-sampling-rate issue of an ECG device can lead to the miss detection of pacemaker spikes as well as incorrect analysis on paced rhythm and non-paced arrhythmias. To solve the issue, this paper proposed a novel system that applies the compressive sampling (CS) framework to sub-Nyquist acquire and reconstruct ECG, and then uses multi-dimensional feature-based deep learning to identify paced rhythm and non-paced arrhythmias. Simulation testing results on ECG databases and comparison with existing approaches demonstrate its effectiveness and outstanding performance for pacemaker ECG analysis.

Feature matching based ECG generative network for arrhythmia event augmentation.

Recent developments in the field of deep learning has shown a rise in its use for clinical applications such as electrocardiogram (ECG) analysis and cardiac arrhythmia classification. Such systems are essential in the early detection and management of cardiovascular diseases. However, due to privacy concerns and also the lack of resources, there is a gap in the data available to run such powerful and data-intensive models. To address the lack of annotated, high-quality ECG data for heart disease research, ECG data generation from a small set of ECG to obtain huge annotated data is seen as an effective solution. Generative Feature Matching Network (GFMN) was shown to resolve few drawbacks of commonly used generative adversarial networks (GAN). Based on this, we developed a deep learning model to generate ECGs that resembles real ECG by feature matching with the existing data.Clinical relevance- This work addresses the lack of a large quantity of good quality, publicly available annotated ECG data required to build deep learning models for cardiac signal processing research. We can use the model presented in this paper to generate ECG signals of a target rhythm pattern and also subject-specific ECG morphology that could improve their cardiac health monitoring while maintaining privacy.

Practical Machine Learning-Based Sepsis Prediction.

Sepsis is a life-threatening clinical syndrome and one of the most expensive conditions treated in hospitals. It is challenging to detect due to the nonspecific clinical signs and the absence of gold standard diagnostics. However, early recognition of sepsis and optimal treatments for sepsis are of paramount importance to improve the condition's management and patient outcomes. This paper aims to delineate key aspects of current sepsis detection systems, including their dependency on clinical expert and laboratory biometric features requiring ongoing critical care intervention, the efficacy of vital sign measures, and the effect of the study population with respect to the precision of sepsis prediction. The AUROC performances of XGBoost models trained on a heterogenous ICU patient group (n=3932) showed significant degradations (p<0.05) as the expert and laboratory biomarker features are removed systematically and vital sign features taken in ICU settings are left. The performance of XGBoost models trained only with vital sign features on a more homogeneous group of ICU patients (n=1927) had a significantly (P<0.05) improved AUPRC to moderate level. The presented results highlight the importance of making a practical machine learning system for sepsis prediction by considering the availability of dominant features as well as personalizing sepsis prediction by configuring it to the specific demographics of a targeted population.

Artificial intelligence mobile health platform for early detection of COVID-19 in quarantine subjects using a wearable biosensor: protocol for a randomised controlled trial.

INTRODUCTION: There is an outbreak of COVID-19 worldwide. As there is no effective therapy or vaccine yet, rigorous implementation of traditional public health measures such as isolation and quarantine remains the most effective tool to control the outbreak. When an asymptomatic individual with COVID-19 exposure is being quarantined, it is necessary to perform temperature and symptom surveillance. As such surveillance is intermittent in nature and highly dependent on self-discipline, it has limited effectiveness. Advances in biosensor technologies made it possible to continuously monitor physiological parameters using wearable biosensors with a variety of form factors. OBJECTIVE: To explore the potential of using wearable biosensors to continuously monitor multidimensional physiological parameters for early detection of COVID-19 clinical progression. METHOD: This randomised controlled open-labelled trial will involve 200-1000 asymptomatic subjects with close COVID-19 contact under mandatory quarantine at designated facilities in Hong Kong. Subjects will be randomised to receive a remote monitoring strategy (intervention group) or standard strategy (control group) in a 1:1 ratio during the 14 day-quarantine period. In addition to fever and symptom surveillance in the control group, subjects in the intervention group will wear wearable biosensors on their arms to continuously monitor skin temperature, respiratory rate, blood pressure, pulse rate, blood oxygen saturation and daily activities. These physiological parameters will be transferred in real time to a smartphone application called Biovitals Sentinel. These data will then be processed using a cloud-based multivariate physiology analytics engine called Biovitals to detect subtle physiological changes. The results will be displayed on a web-based dashboard for clinicians' review. The primary outcome is the time to diagnosis of COVID-19. ETHICS AND DISSEMINATION: Ethical approval has been obtained from institutional review boards at the study sites. Results will be published in peer-reviewed journals.

Spectro-Temporal Feature Based Multi-Channel Convolutional Neural Network for ECG Beat Classification.

Automatic classification of abnormal beats in ECG signals is crucial for monitoring cardiac conditions and the performance of the classification will improve the success rate of the treatment. However, under certain circumstances, traditional classifiers cannot be adapted well to the variation of ECG morphologies or variation of different patients due to fixed hand-crafted features selection. Additionally, existing deep learning related solutions reach their limitation because they fail to use the beat-to-beat information together with single-beat morphologies. This paper applies a novel solution which converts one-dimensional ECG signal into spectro-temporal images and use multiple dense convolutional neural network to capture both beat-to-beat and single-beat information for analysis. The results of simulation on the MIT-BIH arrhythmias database demonstrate the effectiveness of the proposed methodology by showing an outstanding detection performance compared to other existing methods.

Biovitals™: A Personalized Multivariate Physiology Analytics Using Continuous Mobile Biosensors.

More than 50% of the whole world lives with chronic diseases leading to a global economic burden of 47 trillion dollars. Healthcare organizations are moving towards managing patients outside hospital, thereby improving patient safety and quality of life. Current at-home ambulatory remote monitoring analytics based on population level thresholds of individuals physiology have shown poor outcomes and high degree of false alarm burden. The personalized multivariate physiology analytics leverages readily-available low-cost wearable biosensors to detect subtle physiology changes precursor of patient's health deterioration. In this paper we present a novel personalized multivariate physiology analytics for remote patient monitoring in an ambulatory setting. We also present our verification and validation results using perturbation testing along with clinical trial results.

FreeWalker: a smart insole for longitudinal gait analysis.

Gait analysis is an important diagnostic measure to investigate the pattern of walking. Traditional gait analysis is generally carried out in a gait lab, with equipped force and body tracking sensors, which needs a trained medical professional to interpret the results. This procedure is tedious, expensive, and unreliable and makes it difficult to track the progress across multiple visits. In this paper, we present a smart insole called FreeWalker, which provides quantitative gait analysis outside the confinement of traditional lab, at low- cost. The insole consists of eight pressure sensors and two motion tracking sensors, i.e. 3-axis accelerometer and 3-axis gyroscope. This enables measurement of under-foot pressure distribution and motion sequences in real-time. The insole is enabled with onboard SD card as well as wireless data transmission, which help in continuous gait-cycle analysis. The data is then sent to a gateway, for analysis and interpretation of data, using a user interface where gait features are graphically displayed. We also present validation result of a subject's left foot, who was asked to perform a specific task. Experiment results show that we could achieve a data-sampling rate of over 1 KHz, transmitting data up to a distance of 20 meter and maintain a battery life of around 24 hours. Taking advantage of these features, FreeWalker can be used in various applications, like medical diagnosis, rehabilitation, sports and entertainment.