Linear ECG-lead transformations (LELTs) are used to estimate unrecorded target leads by applying a number of recorded basis leads to a LELT matrix. Such LELT matrices are commonly developed using training datasets that are composed of ECGs that belong to different diagnostic classes (DCs). The aim of our research was to assess the influence of the training set composition on the estimation performance of LELTs that estimate target leads V1, V3, V4 and V6 from basis leads I, II, V2 and V5 of the 12-lead ECG. Our assessment was performed using ECGs from the three DCs left ventricular hypertrophy, right bundle branch block and normal (ECGs without abnormalities). Training sets with different DC compositions were used for the development of LELT matrices. These matrices were used to estimate the target leads of different test sets. The estimation performance of the developed matrices was quantified using root mean square error values calculated between derived and recorded target leads. Our findings indicate that unbalanced training sets can lead to LELTs that show large estimation performance variability across different DCs. Balanced training sets were found to produce LELTs that performed well across multiple DCs. We recommend balanced training sets for the development of LELTs.
Background: Electrocardiogram (ECG) signals are often contaminated by noise. Manual review of large ECG databases to identify noisy signals is time-consuming. Traditional signal quality assessment algorithms often do not generalize well or are computationally expensive. This study developed a Temporal Convolutional Neural Network (TCNN) to estimate the signal-to-noise ratio (SNR) of ECG signals. Method: We trained a TCNN on a proprietary database of 134,019 12-lead ECGs without any machine or human-added noise labels. Assuming that this data had high SNR, we randomly selected a single lead from each ECG and added random Gaussian noise. We then scaled the signals and added noise to give a negatively skewed normal distribution of true SNR values. We trained a TCNN to regress low- and high-frequency pseudo-SNR values from the raw noisy input signals. Results: On the testing dataset, the TCNN achieved a mean error of 0.31±1.80 dB and a Pearson correlation coefficient of 0.96 for low-frequency pseudo-SNR. Similarly, for high-frequency pseudo-SNR, the mean error was 0.29±1.63 dB and the Pearson correlation coefficient was 0.97. Conclusion: A Temporal Convolutional Neural Network can accurately estimate the SNR of unseen ECGs.
This study assessed the performance of a deep neural network (PulseAI, Belfast, United Kingdom) used in conjunction with a dry-electrode ECG sensor device (RhythmPad, D&FT, United Kingdom) to detect AF automatically. Simultaneous pairs of 12-lead ECGs and single-lead dry-electrode ECGs were collected from 622 patients. The 12-lead ECGs were manually overread and used as reference diagnoses. Twenty-two patients were confirmed with AF and had an interpretable 12-lead and single-lead dry-electrode ECG recording. The deep neural network analysed the dry-electrode ECGs, and performance was compared to the 12-lead interpretation. Overall, the deep neural network algorithm yielded a sensitivity of 96% (95% CI, 87%-100%), specificity of 99% (95% CI, 98%-100%) and positive predictive value of 81% (95% CI, 66%-96%) for detection of AF episodes. When coupled with dry-electrode ECG sensors, the PulseAI neural network allows for large-scale and low-cost screening for AF. Widespread implementation of this technology may allow for earlier detection, treatment, and management of patients with AF.
Linear ECG-lead transformations (LELTs) are used to estimate unrecorded leads by applying a number of recorded leads to a LELT matrix. Such LELT matrices are commonly developed using a training dataset and linear regression analysis. An important performance metric of LELTs is the subject-to-subject variability (SSV) of their estimation performance. In this research, we assess the relationship between an increasing training set size (from n=10 to n=370 subjects) and the SSV of LELTs. A total of 200 LELT matrices were developed for each training sets size. The developed LELT matrices and 12-lead ECG data of a testing dataset (n=123 subjects) were used for the estimation of Frank VCGs. Root-mean-squared-error (RMSE) values between recorded and estimated Frank VCG leads were used for the quantification of the estimation performance. The SSV associated with each LELT matrix was quantified as the standard deviation of the corresponding RMSE values. This was followed by an analysis of the relationship between the training set size and the associated SSV values. Increasing the training set size from 10 to 180, to 160 and to 200 subjects, for Frank VCG leads X,Y and Z respectively, was associated with a reduction of the observed SSV. Further increases in training set size were found to only have a marginal effect on the observed SSV.
Background
The application of artificial intelligence to interpret the electrocardiogram (ECG) has predominantly included the use of knowledge engineered rule-based algorithms which have become widely used today in clinical practice. However, over recent decades, there has been a steady increase in the number of research studies that are using machine learning (ML) to read or interrogate ECG data.
Objective
The aim of this study is to review the use of ML with ECG data using a time series approach.
Methods
Papers that address the subject of ML and the ECG were identified by systematically searching databases that archive papers from January 1995 to October 2019. Time series analysis was used to study the changing popularity of the different types of ML algorithms that have been used with ECG data over the past two decades. Finally, a meta-analysis of how various ML techniques performed for various diagnostic classifications was also undertaken.
Results
A total of 757 papers was identified. Based on results, the use of ML with ECG data started to increase sharply (p < 0.001) from 2012. Healthcare applications, especially in heart abnormality classification, were the most common application of ML when using ECG data (p < 0.001). However, many new emerging applications include using ML and the ECG for biometrics and driver drowsiness. The support vector machine was the technique of choice for a decade. However, since 2018, deep learning has been trending upwards and is likely to be the leading technique in the coming few years. Despite the accuracy paradox, accuracy was the most frequently used metric in the studies reviewed, followed by sensitivity, specificity, F1 score and then AUC.
Conclusion
Applying ML using ECG data has shown promise. Data scientists and physicians should collaborate to ensure that clinical knowledge is being applied appropriately and is informing the design of ML algorithms. Data scientists also need to consider knowledge guided feature engineering and the explicability of the ML algorithm as well as being transparent in the algorithm's performance to appropriately calibrate human-AI trust. Future work is required to enhance ML performance in ECG classification.
Deep Convolutional Neural Networks (DCNNs) have been shown to provide improved performance over traditional heuristic algorithms for the detection of arrhythmias from ambulatory ECG recordings. However, these DCNNs have primarily been trained and tested on device-specific databases with standardized electrode positions and uniform sampling frequencies. This work explores the possibility of training a DCNN for Atrial Fibrillation (AF) detection on a database of single‑lead ECG rhythm strips extracted from resting 12‑lead ECGs. We then test the performance of the DCNN on recordings from ambulatory ECG devices with different recording leads and sampling frequencies.
We developed an extensive proprietary resting 12‑lead ECG dataset of 549,211 patients. This dataset was randomly split into a training set of 494,289 patients and a testing set of the remaining 54,922 patients. We trained a 34-layer convolutional DCNN to detect AF and other arrhythmias on this dataset. The DCNN was then validated on two Physionet databases commonly used to benchmark automated ECG algorithms (1) MIT-BIH Arrhythmia Database and (2) MIT-BIH Atrial Fibrillation Database. Validation was performed following the EC57 guidelines, with performance assessed by gross episode and duration sensitivity and positive predictive value (PPV). Finally, validation was also performed on a selection of rhythm strips from an ambulatory ECG patch that a committee of board-certified cardiologists annotated.
On MIT-BIH, The DCNN achieved a sensitivity of 100% and 84% PPV in detecting episodes of AF. and 100% sensitivity and 94% PPV in quantifying AF episode duration. On AFDB, The DCNN achieved a sensitivity of 94% and PPV of 98% in detecting episodes of AF, and 98% sensitivity and 100% PPV in quantifying AF episode duration. On the patch database, the DCNN demonstrated performance that was closely comparable to that of a cardiologist.
The results indicate that DCNN models can learn features that generalize between resting 12‑lead and ambulatory ECG recordings, allowing DCNNs to be device agnostic for detecting arrhythmias from single‑lead ECG recordings and enabling a range of clinical applications.
Linear electrocardiographic lead transformations (LELTs) are used to estimate unrecorded ECG leads by applying a number of recorded leads to a LELT matrix. Such matrices are commonly developed using a training dataset. The size of the training dataset has an influence on the estimation performance of a LELT matrix. However, an estimate of the minimal size required for the development of LELTs has previously not been reported.
The aim of this research was to determine such an estimate. We generated LELT matrices from differently sized (from n = 10 to n = 540 subjects in steps of 10 subjects) training datasets. The LELT matrices and the 12-lead ECG data of a testing dataset (n = 186 subjects) were used for the estimation of Frank VCGs. Root-mean-squared-error values between recorded and estimated Frank leads of the testing dataset were used for the quantification of the estimation performance associated with a given size of the training dataset.
The performance of the LELTs was, after an initial phase of improvement, found to only marginally improve with additional increases in the size of the training dataset. Our findings suggest that the training dataset should have
a minimal size of 170 subjects when developing LELTs that utilise the 12-lead ECG for the estimation of unrecorded
ECG leads.
Introduction: Interpretation of the 12‑lead Electrocardiogram (ECG) is normally assisted with an automated diagnosis (AD), which can facilitate an ‘automation bias’ where interpreters can be anchored. In this paper, we studied, 1) the effect of an incorrect AD on interpretation accuracy and interpreter confidence (a proxy foruncertainty), and 2) whether confidence and other interpreter features can predict interpretation accuracy using machine learning. Methods: This study analysed 9000 ECG interpretations from cardiology and non-cardiology fellows (CFs and non-CFs). One third of the ECGs involved no ADs, one third with ADs (half as incorrect) and one third had multiple ADs. Interpretations were scored and interpreter confidence was recorded for each interpretation and subsequently standardised using sigma scaling. Spearman coefficients were used for correlation analysis and C5.0 decision trees were used for predicting interpretation accuracy using basic interpreter features such as confidence, age, experience and designation. Results: Interpretation accuracies achieved by CFs and non-CFs dropped by 43.20% and 58.95% respectively when an incorrect AD was presented (p b 0.001). Overall correlation between scaled confidence and interpretation accuracy was higher amongst CFs. However, correlation between confidence and interpretation accuracy decreased for both groups when an incorrect AD was presented. We found that an incorrect AD disturbs the reliability of interpreter confidence in predicting accuracy. An incorrect AD has a greater effect on the confidence of nonCFs (although this is not statistically significant it is close to the threshold, p = 0.065). The best C5.0 decision tree achieved an accuracy rate of 64.67% (p b 0.001), however this is only 6.56% greater than the noinformation-rate. Conclusion: Incorrect ADs reduce the interpreter's diagnostic accuracy indicating an automation bias. Non-CFs tend to agree more with the ADs in comparison to CFs, hence less expert physicians are more effected by automation bias. Incorrect ADs reduce the interpreter's confidence and also reduces the predictive power of confidence for predicting accuracy (even more so for non-CFs). Whilst a statistically significant model was developed, it is difficult to predict interpretation accuracy using machine learning on basic features such as interpreter confidence, age, reader experience and designation.
CPR Guideline Chest Compression Depths May Exceed Requirements for Optimal Physiological Response
(2018)
A twelve-animal porcine study dataset was retrospectively analyzed to assess associations between chest compression (CC) depth, systolic blood pressure (SBP) and end-tidal carbon dioxide (EtCO2). Manual CCs were applied for 7 two-minute episodes, at CC depths between 10mm-55mm. A rolling 15s analysis window was applied to the continuous signals. Mean peak values were calculated for each window. Correlation analysis was applied to assess strength of association. Optimal CC depth to achieve physiological targets was determined via cut-off analysis. A total of 672 observations for each variable were available for analysis. Pearson correlations (95% confidence interval; p-value) between CC depth and both SBP and ETCO2 were 0.84 (0.82, 0.86; p < 0.001) and 0.75 (0.71, 0.78; p < 0.001) respectively. Optimal CC depth cutoff (sensitivity, specificity) to achieve SBP ≥ 100mmHg and EtCO2 ≥ 10mmHg was 33 mm (98.29%, 88.94%) and 20mm (95.08%, 78.30%) respectively. A reasonable relationship between CC depth and
physiological response was observed. Optimal SBP and EtCO2 cut-offs were achieved significantly below guideline depths. Furthermore, cut-off analysis suggests a disparity between CC depth and physiological targets.
Body surface potential maps (BSPMs) are typically recorded from a large number of ECG leads that cover the entire thorax. This improves diagnostic accuracy and is required in Electrocardiographic imaging (ECGi). BSPMs recorded in the clinical setting may have some leads that are noisy due to poor skin electrode contact. We analyzed 117 lead BSPMs recorded from 360 subjects. We successively simulated the removal of ECG leads at various locations and tested the ability of our algorithm to accurately reconstruct the missing information. When seven electrodes were removed, the algorithm could reconstruct BSPM patterns from QRS segments with median RMSE of 6.24µV and 12.15µV and CC of 0.999 and 0.997 when Laplacian method and PCA based method were used respectively. This work shows that noisy BSPM leads, which often manifest in the clinical setting, can be more accurately reconstructed using our Laplacian based interpolation
algorithm, when low number of missed electrodes in regions where electrodes are organised in a well distributed and tight mesh.