Clinical validation of an artificial intelligence algorithm offering cross-platform detection of atrial fibrillation using smart device electrocardiograms.

BACKGROUND: Several smart devices are able to detect atrial fibrillation automatically by recording a single-lead electrocardiogram, and have created a work overload at the hospital level as a result of the need for over-reads by physicians. AIM: To compare the atrial fibrillation detection performances of the manufacturers' algorithms of five smart devices and a novel deep neural network-based algorithm. METHODS: We compared the rate of inconclusive tracings and the diagnostic accuracy for the detection of atrial fibrillation between the manufacturers' algorithms and the deep neural network-based algorithm on five smart devices, using a physician-interpreted 12-lead electrocardiogram as the reference standard. RESULTS: Of the 117 patients (27% female, median age 65 years, atrial fibrillation present at time of recording in 30%) included in the final analysis (resulting in 585 analyzed single-lead electrocardiogram tracings), the deep neural network-based algorithm exhibited a higher conclusive rate relative to the manufacturer algorithm for all five models: 98% vs. 84% for Apple; 99% vs. 81% for Fitbit; 96% vs. 77% for AliveCor; 99% vs. 85% for Samsung; and 97% vs. 74% for Withings (P<0.01, for each model). When applying our deep neural network-based algorithm, sensitivity and specificity to correctly identify atrial fibrillation were not significantly different for all assessed smart devices. CONCLUSION: In this clinical validation, the deep neural network-based algorithm significantly reduced the number of tracings labeled inconclusive, while demonstrating similarly high diagnostic accuracy for the detection of atrial fibrillation, thereby providing a possible solution to the data surge created by these smart devices.

Evaluation of an Ambulatory ECG Analysis Platform Using Deep Neural Networks in Routine Clinical Practice.

Background Holter analysis requires significant clinical resources to achieve a high-quality diagnosis. This study sought to assess whether an artificial intelligence (AI)-based Holter analysis platform using deep neural networks is noninferior to a conventional one used in clinical routine in detecting a major rhythm abnormality. Methods and Results A total of 1000 Holter (24-hour) recordings were collected from 3 tertiary hospitals. Recordings were independently analyzed by cardiologists for the AI-based platform and by electrophysiologists as part of clinical practice for the conventional platform. For each Holter, diagnostic performance was evaluated and compared through the analysis of the presence or absence of 5 predefined cardiac abnormalities: pauses, ventricular tachycardia, atrial fibrillation/flutter/tachycardia, high-grade atrioventricular block, and high burden of premature ventricular complex (>10%). Analysis duration was monitored. The deep neural network-based platform was noninferior to the conventional one in its ability to detect a major rhythm abnormality. There were no statistically significant differences between AI-based and classical platforms regarding the sensitivity and specificity to detect the predefined abnormalities except for atrial fibrillation and ventricular tachycardia (atrial fibrillation, 0.98 versus 0.91 and 0.98 versus 1.00; pause, 0.95 versus 1.00 and 1.00 versus 1. 00; premature ventricular contractions, 0.96 versus 0.87 and 1.00 versus 1.00; ventricular tachycardia, 0.97 versus 0.68 and 0.99 versus 1.00; atrioventricular block, 0.93 versus 0.57 and 0.99 versus 1.00). The AI-based analysis was >25% faster than the conventional one (4.4 versus 6.0 minutes; P<0.001). Conclusions These preliminary findings suggest that an AI-based strategy for the analysis of Holter recordings is faster and at least as accurate as a conventional analysis by electrophysiologists.

Smartwatch Electrocardiogram and Artificial Intelligence for Assessing Cardiac-Rhythm Safety of Drug Therapy in the COVID-19 Pandemic. The QT-logs study.

BACKGROUND: QTc interval monitoring, for the prevention of drug-induced arrhythmias is necessary, especially in the context of coronavirus disease 2019 (COVID-19). For the provision of widespread use, surrogates for 12­lead ECG QTc assessment may be useful. This prospective observational study compared QTc duration assessed by artificial intelligence (AI-QTc) (Cardiologs®, Paris, France) on smartwatch single­lead electrocardiograms (SW-ECGs) with those measured on 12­lead ECGs, in patients with early stage COVID-19 treated with a hydroxychloroquine-azithromycin regimen. METHODS: Consecutive patients with COVID-19 who needed hydroxychloroquine-azithromycin therapy, received a smartwatch (Withings Move ECG®, Withings, France). At baseline, day-6 and day-10, a 12­lead ECG was recorded, and a SW-ECG was transmitted thereafter. Throughout the drug regimen, a SW-ECG was transmitted every morning at rest. Agreement between manual QTc measurement on a 12­lead ECG and AI-QTc on the corresponding SW-ECG was assessed by the Bland-Altman method. RESULTS: 85 patients (30 men, mean age 38.3 ± 12.2 years) were included in the study. Fair agreement between manual and AI-QTc values was observed, particularly at day-10, where the delay between the 12­lead ECG and the SW-ECG was the shortest (-2.6 ± 64.7 min): 407 ± 26 ms on the 12­lead ECG vs 407 ± 22 ms on SW-ECG, bias -1 ms, limits of agreement -46 ms to +45 ms; the difference between the two measures was <50 ms in 98.2% of patients. CONCLUSION: In real-world epidemic conditions, AI-QTc duration measured by SW-ECG is in fair agreement with manual measurements on 12­lead ECGs. Following further validation, AI-assisted SW-ECGs may be suitable for QTc interval monitoring. REGISTRATION: ClinicalTrial.govNCT04371744.

AI Filter Improves Positive Predictive Value of Atrial Fibrillation Detection by an Implantable Loop Recorder.

OBJECTIVES: The purpose of this study was to determine whether incorporation of a 2-part artificial intelligence (AI) filter can improve the positive predictive value (PPV) of implantable loop recorder (ILR)-detected atrial fibrillation (AF) episodes. BACKGROUND: ILRs can detect AF. Devices transmit data daily. It is critical that the PPV of ILR-detected AF events be high. METHODS: In total, 1,500 AF episodes were evaluated from patients with cryptogenic stroke or known AF who underwent ILR implantation (Reveal LINQ, Medtronic, Minneapolis, Minnesota). Each episode was annotated as either a true or false AF episode to determine the PPV. A 2-part AI-based filter (Cardiologs, Paris, France) was then employed using a deep neural network (DNN) for AF detection. The impact of this DNN filter on the PPV was then assessed. RESULTS: The cohort included 425 patients (mean age 69 ± 10 years; 62% men) with an ILR. After excluding 17 (1.1%) uninterpretable electrocardiograms, 800 (53.9%) of the remaining 1,483 episodes were manually adjudicated to represent an actual atrial arrhythmia. The PPV of ILR-detected AF episodes was 53.9% (95% confidence interval (CI): 51.4% to 56.5%), which increased to 74.5% (95% CI: 71.8% to 77.0%; p < 0.001) following use of the DNN filter. The increase was greatest for AF episodes ≤30 min. The most common reason for a false-positive AF event was premature atrial contractions. There was a negligible failure to identify true AF episodes. CONCLUSIONS: Despite currently available ILR programming options, designed to maximize PPV in a given population, false-positive AF episodes remain common. An AI-based solution may significantly reduce the time and effort needed to adjudicate these false-positive events.

A deep neural network for 12-lead electrocardiogram interpretation outperforms a conventional algorithm, and its physician overread, in the diagnosis of atrial fibrillation.

BACKGROUND: Automated electrocardiogram (ECG) interpretations may be erroneous, and lead to erroneous overreads, including for atrial fibrillation (AF). We compared the accuracy of the first version of a new deep neural network 12-Lead ECG algorithm (Cardiologs®) to the conventional Veritas algorithm in interpretation of AF. METHODS: 24,123 consecutive 12-lead ECGs recorded over 6 months were interpreted by 1) the Veritas® algorithm, 2) physicians who overread Veritas® (Veritas®â€¯+ physician), and 3) Cardiologs® algorithm. We randomly selected 500 out of 858 ECGs with a diagnosis of AF according to either algorithm, then compared the algorithms' interpretations, and Veritas®â€¯+ physician, with expert interpretation. To assess sensitivity for AF, we analyzed a separate database of 1473 randomly selected ECGs interpreted by both algorithms and by blinded experts. RESULTS: Among the 500 ECGs selected, 399 had a final classification of AF; 101 (20.2%) had ≥1 false positive automated interpretation. Accuracy of Cardiologs® (91.2%; CI: 82.4-94.4) was higher than Veritas® (80.2%; CI: 76.5-83.5) (p < 0.0001), and equal to Veritas®â€¯+ physician (90.0%, CI:87.1-92.3) (p = 0.12). When Veritas® was incorrect, accuracy of Veritas®â€¯+ physician was only 62% (CI 52-71); among those ECGs, Cardiologs® accuracy was 90% (CI: 82-94; p < 0.0001). The second database had 39 AF cases; sensitivity was 92% vs. 87% (p = 0.46) and specificity was 99.5% vs. 98.7% (p = 0.03) for Cardiologs® and Veritas® respectively. CONCLUSION: Cardiologs® 12-lead ECG algorithm improves the interpretation of atrial fibrillation.

Modeling the Correlated Activity of Neural Populations: A Review.

The principles of neural encoding and computations are inherently collective and usually involve large populations of interacting neurons with highly correlated activities. While theories of neural function have long recognized the importance of collective effects in populations of neurons, only in the past two decades has it become possible to record from many cells simultaneously using advanced experimental techniques with single-spike resolution and to relate these correlations to function and behavior. This review focuses on the modeling and inference approaches that have been recently developed to describe the correlated spiking activity of populations of neurons. We cover a variety of models describing correlations between pairs of neurons, as well as between larger groups, synchronous or delayed in time, with or without the explicit influence of the stimulus, and including or not latent variables. We discuss the advantages and drawbacks or each method, as well as the computational challenges related to their application to recordings of ever larger populations.

Blindfold learning of an accurate neural metric.

The brain has no direct access to physical stimuli but only to the spiking activity evoked in sensory organs. It is unclear how the brain can learn representations of the stimuli based on those noisy, correlated responses alone. Here we show how to build an accurate distance map of responses solely from the structure of the population activity of retinal ganglion cells. We introduce the Temporal Restricted Boltzmann Machine to learn the spatiotemporal structure of the population activity and use this model to define a distance between spike trains. We show that this metric outperforms existing neural distances at discriminating pairs of stimuli that are barely distinguishable. The proposed method provides a generic and biologically plausible way to learn to associate similar stimuli based on their spiking responses, without any other knowledge of these stimuli.

A spike sorting toolbox for up to thousands of electrodes validated with ground truth recordings in vitro and in vivo.

In recent years, multielectrode arrays and large silicon probes have been developed to record simultaneously between hundreds and thousands of electrodes packed with a high density. However, they require novel methods to extract the spiking activity of large ensembles of neurons. Here, we developed a new toolbox to sort spikes from these large-scale extracellular data. To validate our method, we performed simultaneous extracellular and loose patch recordings in rodents to obtain 'ground truth' data, where the solution to this sorting problem is known for one cell. The performance of our algorithm was always close to the best expected performance, over a broad range of signal-to-noise ratios, in vitro and in vivo. The algorithm is entirely parallelized and has been successfully tested on recordings with up to 4225 electrodes. Our toolbox thus offers a generic solution to sort accurately spikes for up to thousands of electrodes.

Identification of Parameters of Host Cell Vulnerability during Salmonella Infection by Quantitative Image Analysis and Modeling.

Salmonella targets and enters epithelial cells at permissive entry sites: some cells are more likely to be infected than others. However, the parameters that lead to host cell heterogeneity are not known. Here, we quantitatively characterized host cell vulnerability to Salmonella infection based on imaged parameters. We performed successive infections of the same host cell population followed by automated high-throughput microscopy and observed that infected cells have a higher probability of being reinfected. Establishing a predictive model, we identified two combined origins of host cell vulnerability: pathogen-induced cellular vulnerability emerging from Salmonella uptake and persisting at later stages of the infection and host cell-inherent vulnerability. We linked the host cell-inherent vulnerability with its morphological attributes, such as local cell crowding, and with host cell cholesterol content. This showed that the probability of Salmonella infection success can be forecast from morphological or molecular host cell parameters.

Closed-Loop Estimation of Retinal Network Sensitivity by Local Empirical Linearization.

Understanding how sensory systems process information depends crucially on identifying which features of the stimulus drive the response of sensory neurons, and which ones leave their response invariant. This task is made difficult by the many nonlinearities that shape sensory processing. Here, we present a novel perturbative approach to understand information processing by sensory neurons, where we linearize their collective response locally in stimulus space. We added small perturbations to reference stimuli and tested if they triggered visible changes in the responses, adapting their amplitude according to the previous responses with closed-loop experiments. We developed a local linear model that accurately predicts the sensitivity of the neural responses to these perturbations. Applying this approach to the rat retina, we estimated the optimal performance of a neural decoder and showed that the nonlinear sensitivity of the retina is consistent with an efficient encoding of stimulus information. Our approach can be used to characterize experimentally the sensitivity of neural systems to external stimuli locally, quantify experimentally the capacity of neural networks to encode sensory information, and relate their activity to behavior.

A Tractable Method for Describing Complex Couplings between Neurons and Population Rate.

Neurons within a population are strongly correlated, but how to simply capture these correlations is still a matter of debate. Recent studies have shown that the activity of each cell is influenced by the population rate, defined as the summed activity of all neurons in the population. However, an explicit, tractable model for these interactions is still lacking. Here we build a probabilistic model of population activity that reproduces the firing rate of each cell, the distribution of the population rate, and the linear coupling between them. This model is tractable, meaning that its parameters can be learned in a few seconds on a standard computer even for large population recordings. We inferred our model for a population of 160 neurons in the salamander retina. In this population, single-cell firing rates depended in unexpected ways on the population rate. In particular, some cells had a preferred population rate at which they were most likely to fire. These complex dependencies could not be explained by a linear coupling between the cell and the population rate. We designed a more general, still tractable model that could fully account for these nonlinear dependencies. We thus provide a simple and computationally tractable way to learn models that reproduce the dependence of each neuron on the population rate.