Objective : Atrial fibrillation is a common type of heart rhythm abnormality caused by a problem with the heart’s electrical system. Early detection of this disease has important implications for stroke prevention and management. Our objective is to construct an intelligent tool that assists cardiologists in identifying automatically cardiac arrhythmias and noise in electrocardiogram (ECG) recordings. Approach : Our base deep classifier combined a convolutional neural network (CNNs) and a sequence of long short-term memory units, with pooling, dropout and normalization techniques to improve their accuracy. The network predicted a classification at every 18th input sample and the final prediction was selected for classification. Ten standalone models that used our base classifier architecture were first cross-validated separately on 90% of the PhysioNet/CinC Challenge 2017 dataset and then tested on 10%. An ensemble classifier selected the label of the best average probability from the ten sub-models to improve prediction quality. Main results : Our original result submitted to the challenge gave a mean F ₁ -measure of 80%. The new proposed method improved the test score to 82%, which was tied for the third-highest score in the follow-up phase of the challenge. Significance : Without employing a time-consuming feature engineering step, the ensemble classifier trained with this architecture provided a robust solution to the problem of detecting cardiac arrhythmia from noisy ECG signals. In addition, interpretation of the classifier by inspection of its network parameters and predictions revealed what aspects of the ECG signal the classifier considered most discriminating.

Objective : Early detection of sleep arousal in polysomnographic (PSG) signals is crucial for monitoring or diagnosing sleep disorders and reducing the risk of further complications, including heart disease and blood pressure fluctuations. Approach : In this paper, we present a new automatic detector of non-apnea arousal regions in multichannel PSG recordings. This detector cascades four different modules: a second-order scattering transform (ST) with Morlet wavelets; depthwise-separable convolutional layers; bidirectional long short-term memory (BiLSTM) layers; and dense layers. While the first two are shared across all channels, the latter two operate in a multichannel formulation. Following a deep learning paradigm, the whole architecture is trained in an end-to-end fashion in order to optimize two objectives: the detection of arousal onset and offset, and the classification of the type of arousal. Main results and Significance : The novelty of the approach is three-fold: it is the first use of a hybrid ST-BiLSTM network with biomedical signals; it captures frequency information lower (0.1 Hz) than the detection sampling rate (0.5 Hz); and it requires no explicit mechanism to overcome class imbalance in the data. In the follow-up phase of the 2018 PhysioNet/CinC Challenge the proposed architecture achieved a state-of-the-art area under the precision-recall curve (AUPRC) of 0.50 on the hidden test data, tied for the second-highest official result overall.

We describe an automatic classifier of arrhythmias based on 12-lead and reduced-lead electrocardiograms. Our classifier comprises four modules: scattering transform (ST), phase harmonic correlation (PHC), depthwise separable convolutions (DSC), and a long short-term memory (LSTM) network. It is trained on PhysioNet/Computing in Cardiology Challenge 2021 data. The ST captures short-term temporal ECG modulations while the PHC characterizes the phase dependence of coherent ECG components. Both reduce the sampling rate to a few samples per typical heart beat. We pass the output of the ST and PHC to a depthwise-separable convolution layer (DSC) which combines lead responses separately for each ST or PHC coefficient and then combines resulting values across all coefficients. At a deeper level, two LSTM layers integrate local variations of the input over long time scales. We train in an end-to-end fashion as a multilabel classification problem with a normal and 25 arrhythmia classes. Lastly, we use canonical correlation analysis (CCA) for transfer learning from 12-lead ST and PHC representations to reduced-lead ones. After local cross-validation on the public data from the challenge, our team ‘BitScattered’ achieved the following results: 0.682 ± 0.0095 for 12-lead; 0.666 ± 0.0257 for 6-lead; 0.674 ± 0.0185 for 4-lead; 0.661 ± 0.0098 for 3-lead; and 0.662 ± 0.0151 for 2-lead.

Neonatal hypoxic-ischemic encephalopathy (HIE) is caused by sustained hypoxemia near birth. Clinical assessment using cardiotocography (CTG), which measures the fetal heart rate (FHR) and maternal uterine pressure (UP), aims to identify infants at increased risk of HIE. Although CTG is nonstationary, current automated methods for its analysis use time invariant discrimination rules. Our objective was to examine the association between features of CTG and the development of HIE to determine if accounting for the time to delivery (TTD) would strengthen these associations. We analyzed 88 features extracted from FHR and UP signals from 25,197 vaginally delivered infants for whom blood gas measurements were available. All infants were categorized according to their blood gas exams into three mutually exclusive groups: 167 HIE, 1,912 acidosis - a precursor to HIE, and 22,903 healthy cases. We evaluated CTG features during the last twelve hours of labor to explore the associations between 1) CTG features and TTD, 2) CTG features and the development of HIE, and 3) the conditional association between CTG features and the development of HIE given TTD. These associations were quantified using the normalized mutual information. We found that all CTG features varied with TTD. Furthermore, 48 out of 88 features were not significantly associated with the outcome of labor and might not be useful in classification studies. We also found that 40 out of 88 features had significant associations with the development of HIE; accounting for TTD increased the association for 26 of these features. Therefore, automated methods for prediction of infants at risk of HIE should focus on this set of CTG features and account for their time-varying properties.

Recording of maternal uterine pressure (UP) and fetal heart rate (FHR) during labour and delivery is a procedure referred to as cardiotocography (CTG). We model this as an input-output system to estimate its dynamics in terms of an impulse response function (IRF). CTG data is very noisy and missing data are common. In this paper, we identify the models using subspace methods, which incorporate noise-suppression and permit the use of non-contiguous data. Using contiguous data, the subspace method performed better than linear regression; more of the 57 CTG pathological records in our database were modelled (30 vs. 26). Allowing non-contiguous data, even more pathological records were modelled with this approach (49). Furthermore, the models were discriminating; compared to linear regression, the IRF gain showed statistically significant differences more often between normal and pathological records (in 15/18 vs. 10/18 epochs) over the final three hours of labour.

We examined the use of bivariate mutual information (MI) and its conditional variant transfer entropy (TE) to address synchronization of perinatal uterine pressure (UP) and fetal heart rate (FHR). We used a nearest-neighbour based Kraskov entropy estimator, suitable to the non-Gaussian distributions of the UP and FHR signals. Moreover, the estimates were robust to noise by use of surrogate data testing. Estimating degree of synchronicity and UP-FHR delay length is useful since they are physiological correlates to fetal hypoxia. Mutual information of the UP-FHR discriminated normal and pathological fetuses early (160 min before delivery) and discriminated normal and metabolic acidotic fetuses slightly later (110 min before delivery), with higher mutual information for progressively pathological classes. The delay in mutual information transfer was also discriminating in the last 50 min of labour. Transfer entropy discriminated normal and pathological cases 110 min before delivery with lower TE values and longer information transfer delays in pathological cases, to our knowledge, the first report of this phenomena in the literature.