Memory Classifiers for Robust ECG Classification against Physiological Noise.

The development of sophisticated machine learning algorithms has made it possible to detect critical health conditions like cardiac arrhythmia, directly from electrocardiogram (ECG) recordings. Large-scale machine learning models, like deep neural networks, are well known to underperform when subjected to small perturbations which would not pose a challenge to physicians. This is a hurdle that needs to be removed to facilitate wide-scale adoption. We find this to be true even for models trained using data-augmentation schemes.In this paper, we show that using memory classifiers it is possible to attain a boost in robustness using expert-informed features. Memory classifiers combine standard deep neural network training with a domain knowledge-guided similarity metric to boost the robustness of classifiers. We evaluate the performance of the models against naturally occurring physiological perturbations, specifically electrode movement, muscle artifact, and baseline wander noise. Our approach demonstrates improved robustness across all evaluated noises for an average improvement in F1 score of 3.13% compared to models using data augmentation techniques.Clinical relevance- This approach improves the robustness of deep learning methods in safety-critical medical applications.

Prolonged Use of an Automated Insulin Delivery System Improves Sleep in Long-Standing Type 1 Diabetes Complicated by Impaired Awareness of Hypoglycemia.

BACKGROUND: This study assessed changes in actigraphy-estimated sleep and glycemic outcomes after initiating automated insulin delivery (AID). METHODS: Ten adults with long-standing type 1 diabetes and impaired awareness of hypoglycemia (IAH) participated in an 18-month clinical trial assessing an AID intervention on hypoglycemia and counter-regulatory mechanisms. Data from eight participants (median age = 58 years) with concurrent wrist actigraph and continuous glucose monitoring (CGM) data were used in the present analyses. Actigraphs and CGM measured sleep and glycemic control at baseline (one week) and months 3, 6, 9, 12, 15, and 18 (three weeks) following AID initiation. HypoCount software integrated actigraphy with CGM data to separate wake and sleep-associated glycemic measures. Paired sample t-tests and Cohen's d effect sizes modeled changes and their magnitude in sleep, glycemic control, IAH (Clarke score), hypoglycemia severity (HYPO score), hypoglycemia exposure (CGM), and glycemic variability (lability index [LI]; CGM coefficient-of-variation [CV]) from baseline to 18 months. RESULTS: Sleep improved from baseline to 18 months (shorter sleep latency [P < .05, d = 1.74], later sleep offset [P < .05, d = 0.90], less wake after sleep onset [P < .01, d = 1.43]). Later sleep onset (d = 0.74) and sleep midpoint (d = 0.77) showed medium effect sizes. Sleep improvements were evident from 12 to 15 months after AID initiation and were preceded by improved hypoglycemia awareness (Clarke score [d = 1.18]), reduced hypoglycemia severity (HYPO score [d = 2.13]), reduced sleep-associated hypoglycemia (percent time glucose was < 54 mg/dL, < 60 mg/dL,< 70 mg/dL; d = 0.66-0.81), and reduced glucose variability (LI, d = 0.86; CV, d = 0.62). CONCLUSION: AID improved sleep initiation and maintenance. Improved awareness of hypoglycemia, reduced hypoglycemia severity, hypoglycemia exposure, and glucose variability preceded sleep improvements.This trial is registered with ClinicalTrials.gov NCT03215914 https://clinicaltrials.gov/ct2/show/NCT03215914.

Automated Insulin Delivery for Hypoglycemia Avoidance and Glucose Counterregulation in Long-Standing Type 1 Diabetes with Hypoglycemia Unawareness.

Objective: Automated insulin delivery (AID) may benefit individuals with long-standing type 1 diabetes where frequent exposure to hypoglycemia impairs counterregulatory responses. This study assessed the effect of 18 months AID on hypoglycemia avoidance and glucose counterregulatory responses to insulin-induced hypoglycemia in long-standing type 1 diabetes complicated by impaired awareness of hypoglycemia. Methods: Ten participants mean ± standard deviation age 49 ± 16 and diabetes duration 34 ± 16 years were initiated on AID. Continuous glucose monitoring was paired with actigraphy to assess awake- and sleep-associated hypoglycemia exposure every 3 months. Hyperinsulinemic hypoglycemic clamp experiments were performed at baseline, 6, and 18 months postintervention. Hypoglycemia exposure was reduced by 3 months, especially during sleep, with effects sustained through 18 months (P ≤ 0.001) together with reduced glucose variability (P < 0.01). Results: Hypoglycemia awareness and severity scores improved (P < 0.01) with severe hypoglycemia events reduced from median (interquartile range) 3 (3-10) at baseline to 0 (0-1) events/person·year postintervention (P = 0.005). During the hypoglycemic clamp experiments, no change was seen in the endogenous glucose production (EGP) response, however, peripheral glucose utilization during hypoglycemia was reduced following intervention [pre: 4.6 ± 0.4, 6 months: 3.8 ± 0.5, 18 months: 3.4 ± 0.3 mg/(kg·min), P < 0.05]. There were increases over time in pancreatic polypeptide (Pre:62 ± 29, 6 months:127 ± 44, 18 months:176 ± 58 pmol/L, P < 0.01), epinephrine (Pre: 199 ± 53, 6 months: 332 ± 91, 18 months: 386 ± 95 pg/mL, P = 0.001), and autonomic symptom (Pre: 6 ± 2, 6 months: 6 ± 2, 18 months: 10 ± 2, P < 0.05) responses. Conclusions: AID led to a sustained reduction of hypoglycemia exposure. EGP in response to insulin-induced hypoglycemia remained defective, however, partial recovery of glucose counterregulation was evidenced by a reduction in peripheral glucose utilization likely mediated by increased epinephrine secretion and, together with improved autonomic symptoms, may contribute to the observed clinical reduction in hypoglycemia.

Derivation and Validation of an Algorithm to Detect Stroke Using Arm Accelerometry Data.

Background Early diagnosis is essential for effective stroke therapy. Strokes in hospitalized patients are associated with worse outcomes compared with strokes in the community. We derived and validated an algorithm to identify strokes by monitoring upper limb movements in hospitalized patients. Methods and Results A prospective case-control study in hospitalized patients evaluated bilateral arm accelerometry from patients with acute stroke with lateralized weakness and controls without stroke. We derived a stroke classifier algorithm from 123 controls and 77 acute stroke cases and then validated the performance in a separate cohort of 167 controls and 33 acute strokes, measuring false alarm rates in nonstroke controls and time to detection in stroke cases. Faster detection time was associated with more false alarms. With a median false alarm rate among nonstroke controls of 3.6 (interquartile range [IQR], 2.1-5.0) alarms per patient per day, the median time to detection was 15.0 (IQR, 8.0-73.5) minutes. A median false alarm rate of 1.1 (IQR. 0-2.2) per patient per day was associated with a median time to stroke detection of 29.0 (IQR, 11.0-58.0) minutes. There were no differences in algorithm performance for subgroups dichotomized by age, sex, race, handedness, nondominant hemisphere involvement, intensive care unit versus ward, or daytime versus nighttime. Conclusions Arm movement data can be used to detect asymmetry indicative of stroke in hospitalized patients with a low false alarm rate. Additional studies are needed to demonstrate clinical usefulness.

Delays in the Identification and Assessment of in-Hospital Stroke Patients.

OBJECTIVES: In-hospital stroke is associated with poor outcomes. Reasons for delays, use of interventions, and presence of large vessel occlusion are not well characterized. MATERIALS AND METHODS: A retrospective single center cohort of 97 patients with in-hospital stroke was analyzed to identify factors associated with delays from last known normal to symptom identification and to stroke team alerting. Stroke interventions and presence of large vessel occlusion were also assessed. RESULTS: Strokes were predominantly on surgery services (70%), ischemic (82%), and severe (median NIHSS 16; interquartile range [IQR] 6-24). There were long delays from last known normal to symptom identification (median 5.1 hours, IQR 1.0-19.7 hours), symptom identification to stroke team alerting (median 2.1 hours, IQR 0.5-9.9 hours), and total time from last known normal to alerting (median 11.4 [IQR 2.7-34.2] hours). In univariable analysis, being on a surgical service, in an ICU, intubated, and higher NIHSS were associated with delays. In multivariable analysis only intubation was independently associated with time from last known normal to symptom identification (coefficient 20 hours, IQR 0.2 - 39.8, p=0.047). Interventions were given to 17/80 (21%) ischemic stroke patients; 3 (4%) received IV tPA and 14 (18%) underwent thrombectomy. Vascular imaging occurred in 57/80 (71%) ischemic stroke patients and 21/57 (37%) had large vessel occlusion. CONCLUSIONS: Hospitalized patients with stroke experience long delays from symptom identification to stroke team alerting. Intubation was strongly associated with delay to symptom identification. Although stroke severity was high and large vessel occlusion common, many patients did not receive acute interventions.

Willingness to use a wearable device capable of detecting and reversing overdose among people who use opioids in Philadelphia.

BACKGROUND: The incidence of opioid-related overdose deaths has been rising for 30 years and has been further exacerbated amidst the COVID-19 pandemic. Naloxone can reverse opioid overdose, lower death rates, and enable a transition to medication for opioid use disorder. Though current formulations for community use of naloxone have been shown to be safe and effective public health interventions, they rely on bystander presence. We sought to understand the preferences and minimum necessary conditions for wearing a device capable of sensing and reversing opioid overdose among people who regularly use opioids. METHODS: We conducted a combined cross-sectional survey and semi-structured interview at a respite center, shelter, and syringe exchange drop-in program in Philadelphia, Pennsylvania, USA, during the COVID-19 pandemic in August and September 2020. The primary aim was to explore the proportion of participants who would use a wearable device to detect and reverse overdose. Preferences regarding designs and functionalities were collected via a questionnaire with items having Likert-based response options and a semi-structured interview intended to elicit feedback on prototype designs. Independent variables included demographics, opioid use habits, and previous experience with overdose. RESULTS: A total of 97 adults with an opioid use history of at least 3 months were interviewed. A majority of survey participants (76%) reported a willingness to use a device capable of detecting an overdose and automatically administering a reversal agent upon initial survey. When reflecting on the prototype, most respondents (75.5%) reported that they would wear the device always or most of the time. Respondents indicated discreetness and comfort as important factors that increased their chance of uptake. Respondents suggested that people experiencing homelessness and those with low tolerance for opioids would be in greatest need of the device. CONCLUSIONS: The majority of people sampled with a history of opioid use in an urban setting were interested in having access to a device capable of detecting and reversing an opioid overdose. Participants emphasized privacy and comfort as the most important factors influencing their willingness to use such a device. TRIAL REGISTRATION: NCT04530591.

Characterizing Glycemic Control and Sleep in Adults with Long-Standing Type 1 Diabetes and Hypoglycemia Unawareness Initiating Hybrid Closed Loop Insulin Delivery.

Nocturnal hypoglycemia is life threatening for individuals with type 1 diabetes (T1D) due to loss of hypoglycemia symptom recognition (hypoglycemia unawareness) and impaired glucose counter regulation. These individuals also show disturbed sleep, which may result from glycemic dysregulation. Whether use of a hybrid closed loop (HCL) insulin delivery system with integrated continuous glucose monitoring (CGM) designed for improving glycemic control, relates to better sleep across time in this population remains unknown. The purpose of this study was to describe long-term changes in glycemic control and objective sleep after initiating hybrid closed loop (HCL) insulin delivery in adults with type 1 diabetes and hypoglycemia unawareness. To accomplish this, six adults (median age = 58 y) participated in an 18-month ongoing trial assessing HCL effectiveness. Glycemic control and sleep were measured using continuous glucose monitoring and wrist accelerometers every 3 months. Paired sample t-tests and Cohen's d effect sizes modeled glycemic and sleep changes and the magnitude of these changes from baseline to 9 months. Reduced hypoglycemia (d = 0.47-0.79), reduced basal insulin requirements (d = 0.48), and a smaller glucose coefficient of variation (d = 0.47) occurred with medium-large effect sizes from baseline to 9 months. Hypoglycemia awareness improved from baseline to 6 months with medium-large effect sizes (Clarke score (d = 0.60), lability index (d = 0.50), HYPO score (d = 1.06)). Shorter sleep onset latency (d = 1.53; p < 0.01), shorter sleep duration (d = 0.79), fewer total activity counts (d = 1.32), shorter average awakening length (d = 0.46), and delays in sleep onset (d = 1.06) and sleep midpoint (d = 0.72) occurred with medium-large effect sizes from baseline to 9 months. HCL led to clinically significant reductions in hypoglycemia and improved hypoglycemia awareness. Sleep showed a delayed onset, reduced awakening length and onset latency, and maintenance of high sleep efficiency after initiating HCL. Our findings add to the limited evidence on the relationships between diabetes therapeutic technologies and sleep health. This trial is registered with ClinicalTrials.gov (NCT03215914).

VitalCore: Analytics and Support Dashboard for Medical Device Integration.

Medical professionals spend extensive time collecting, validating, reviewing, and analyzing medical device data. These devices use vendor-specific applications with lengthy troubleshooting times, causing extended downtimes where medical professionals have to manually document patient data in the electronic health record (EHR). Manual logging of this data creates delays and leaves it vulnerable to errors, manipulation, and omissions. In this paper, we present VitalCore, a medical device integration platform that supports access to medical device data in real-time. We deploy VitalCore in three applications at Penn Medicine: Medical Device Dashboard, Ventilation Alert, and Anomaly Detector. In the Medical Device Dashboard, we reduced, by up to six times, the amount of time required of medical professionals, clinical engineers, and IT analysts by simplifying the troubleshooting workflow, thus decreasing downtimes and increasing clinical productivity. In Ventilation Alert, we demonstrated the ability to assist medical professionals by alerting them to newly ventilated patients. In Anomaly Detector, we showed that we could predict anomalous patterns in our data with 93% accuracy.

The application of continuous glucose monitoring technology to eating disorders research: An idea worth researching.

Continuous glucose monitoring (CGM) devices have revolutionized our capacity to measure blood glucose levels in real time using minimally invasive technology, yet to date there are no studies using CGM in individuals with eating disorders (EDs). Preliminary evidence suggests that eating disorder behaviors (EDBs) have substantial and characteristic impacts on blood glucose levels and glucose-related variables (e.g., binge-eating episodes cause rapid spikes in blood glucose levels, purging causes rapid drops in blood glucose to below normal levels). The aims of this article are to describe the benefits of CGM technology over older methods of measuring blood glucose levels and to discuss several specific ways in which CGM technology can be applied to EDs research to (a) improve our ability to identify and predict engagement in EDBs in real time, (b) identify relationships between blood glucose levels and maintenance factors for EDs, and (c) increase our understanding of the physiological and psychological impacts of disordered eating. We also present preliminary acceptability and feasibility data on the use of CGM devices in individuals with EDs. Overall, the article will describe several applications of CGM technology in EDs research with compelling potential to improve research methodologies.

Intelligent and Dynamic Ransomware Spread Detection and Mitigation in Integrated Clinical Environments.

Medical Cyber-Physical Systems (MCPS) hold the promise of reducing human errors and optimizing healthcare by delivering new ways to monitor, diagnose and treat patients through integrated clinical environments (ICE). Despite the benefits provided by MCPS, many of the ICE medical devices have not been designed to satisfy cybersecurity requirements and, consequently, are vulnerable to recent attacks. Nowadays, ransomware attacks account for 85% of all malware in healthcare, and more than 70% of attacks confirmed data disclosure. With the goal of improving this situation, the main contribution of this paper is an automatic, intelligent and real-time system to detect, classify, and mitigate ransomware in ICE. The proposed solution is fully integrated with the ICE++ architecture, our previous work, and makes use of Machine Learning (ML) techniques to detect and classify the spreading phase of ransomware attacks affecting ICE. Additionally, Network Function Virtualization (NFV) and Software Defined Networking (SDN)paradigms are considered to mitigate the ransomware spreading by isolating and replacing infected devices. Different experiments returned a precision/recall of 92.32%/99.97% in anomaly detection, an accuracy of 99.99% in ransomware classification, and promising detection and mitigation times. Finally, different labelled ransomware datasets in ICE have been created and made publicly available.

Computer Aided Clinical Trials for Implantaule Cardiac Devices.

In this paper we aim to answer the question, "How can modeling and simulation of physiological systems be used to evaluate life-critical implantable medical devices?" Clinical trials for medical devices are becoming increasingly inefficient as they take several years to conduct, at very high cost and suffer from high rates of failure. For example, the Rhythm ID Goes Head-to-head Trial (RIGHT) sought to evaluate the performance of two arrhythmia discriminator algorithms for implantable cardioverter defibrillators, Vitality 2 vs. Medtronic, in terms of time-to-first inappropriate therapy, but concluded with results contrary to the initial hypothesis- after 5 years, 2,000+ patients and at considerble ethical and monetary cost. In this paper, we describe the design and performance of a Computer-aided Clinical Trial (CACT) for Implantable Cardiac Devices where previous trial information, real patient data and closed-loop device models are effectively used to evaluate the trial with high confidence. We formulate the CACT in the context of RIGHT using a Bayesian statistical framework. We define a hierarchical model of the virtual cohort generated from a physiological model which captures the uncertainty in the parameters and allow for the systematic incorporation of information available at the design of the trial. With this formulation, the estimates the inappropriate therapy rate of Vitality 2 compared to Medtronic as 33.22% vs 15.62% $(\mathrm{p}\lt 0.001)$, which is comparable to the original trial. Finally, we relate the outcomes of the computer- aided clinical trial to the primary endpoint of RIGHT.

Reducing Pulse Oximetry False Alarms Without Missing Life-Threatening Events.

Alarm fatigue has been increasingly recognized as one of the most significant problems in the hospital environment. One of the major causes is the excessive number of false physiologic monitor alarms. An underlying problem is the inefficient traditional threshold alarm system for physiologic parameters such as low blood oxygen saturation (SpO2). In this paper, we propose a robust classification procedure based on the AdaBoost algorithm with reject option that can identify and silence false SpO2 alarms, while ensuring zero misclassified clinically significant alarms. Alarms and vital signs related to SpO2 such as heart rate and pulse rate, within monitoring interval are extracted into different numerical features for the classifier. We propose a variant of AdaBoost with reject option by allowing a third decision (i.e., reject) expressing doubt. Weighted outputs of each weak classifier are input to a softmax function optimizing to satisfy a desired false negative rate upper bound while minimizing false positive rate and indecision rate. We evaluate the proposed classifier using a dataset collected from 100 hospitalized children at Children's Hospital of Philadelphia and show that the classifier can silence 23.12% of false SpO2 alarms without missing any clinically significant alarms.

Physiology-Invariant Meal Detection for Type 1 Diabetes.

BACKGROUND: Fully automated artificial pancreas systems require meal detectors to supplement blood glucose level regulation, where false meal detections can cause unnecessary insulin delivery with potentially fatal consequences, and missed detections may cause the patient to experience extreme hyperglycemia. Most existing meal detectors monitor various measures of glucose rate-of-change to detect meals where varying physiology and meal content complicate balancing detector sensitivity versus specificity. METHODS: We developed a novel meal detector based on a minimal glucose-insulin metabolism model and show that the detector is, by design, invariant to patient-specific physiological parameters in the minimal model. Our physiological parameter-invariant (PAIN) detector achieves a near-constant false alarm rate across all individuals and is evaluated against three other major existing meal detectors on a clinical type 1 diabetes data set. RESULTS: In the clinical evaluation, the PAIN-based detector achieves an 86.9% sensitivity for an average false alarm rate of two alarms per day. In addition, for all false alarm rates, the PAIN-based detector performance is significantly better than three other existing meal detectors. In addition, the evaluation results show that the PAIN-based detector uniquely (as compared with the other meal detectors) has low variance in detection and false alarm rates across all patients, without patient-specific personalization. CONCLUSIONS: The PAIN-based meal detector has demonstrated better detection performance than existing meal detectors, and it has the unique strength of achieving a consistent performance across a population with varying physiology without any individual-level parameter tuning or training.

Robust monitoring of hypovolemia in intensive care patients using photoplethysmogram signals.

The paper presents a fingertip photoplethysmography based technique to assess patient fluid status that is robust to waveform artifacts and health variability in the underlying patient population. The technique is intended for use in intensive care units, where patients are at risk for hypovolemia, and signal artifacts and inter-patient variations in health are common. Input signals are preprocessed to remove artifact, then a parameter-invariant statistic is calculated to remove effects of patient-specific physiology. Patient data from the Physionet MIMICII database was used to evaluate the performance of this technique. The proposed method was able to detect hypovolemia within 24 hours of onset in all hypovolemic patients tested, while producing minimal false alarms over non-hypovolemic patients.