Considering the significant burden to patients and healthcare systems globally related to atrial fibrillation (AF) complications, the early AF diagnosis is of crucial importance. In the view of prominent perspectives for fast and accurate point-of-care arrhythmia detection, our study optimizes an artificial neural network (NN) classifier and ranks the importance of enhanced 137 diagnostic ECG features computed from time and frequency ECG signal representations of short single-lead strips available in 2017 Physionet/CinC Challenge database. Based on hyperparameters’ grid search of densely connected NN layers, we derive the optimal topology with three layers and 128, 32, 4 neurons per layer (DenseNet-3@128-32-4), which presents maximal F1-scores for classification of Normal rhythms (0.883, 5076 strips), AF (0.825, 758 strips), Other rhythms (0.705, 2415 strips), Noise (0.618, 279 strips) and total F1 relevant to the CinC Challenge of 0.804, derived by five-fold cross-validation. DenseNet-3@128-32-4 performs equally well with 137 to 32 features and presents tolerable reduction by about 0.03 to 0.06 points for limited input sets, including 8 and 16 features, respectively. The feature reduction is linked to effective application of a comprehensive method for computation of the feature map importance based on the weights of the activated neurons through the total path from input to specific output in DenseNet. The detailed analysis of 20 top-ranked ECG features with greatest importance to the detection of each rhythm and overall of all rhythms reveals DenseNet decision-making process, noticeably corresponding to the cardiologists’ diagnostic point of view.

GPU-based deep neural networks (DNNs) are powerful for electrocardiogram (ECG) processing and rhythm classification. Although questions often arise about their practical application in embedded systems with low computational resources, few studies have investigated the associated challenges. This study aims to show a useful workflow for deploying a pre-trained DNN model from a GPU-based development platform to two popular ARM-based microcontrollers: Raspberry Pi 4 and ARM Cortex-M7. Specifically, a five-layer convolutional neural network pre-trained in TensorFlow (TF) for the detection of ventricular fibrillation is converted to Lite Runtime (LiteRT) format and subjected to post-training quantization to reduce model size and computational complexity. Using a test dataset of 7482 10 s cardiac arrest ECGs, the inference of LiteRT DNN in Raspberry Pi 4 takes about 1 ms with a sensitivity of 98.6% and specificity of 99.5%, reproducing the TF DNN performance. An optimization study with 1300 representative datasets (RDSs), including 10 to 4000 calibration ECG signals selected by random, rhythm, or amplitude-based criteria, showed that choosing a random RDS with a relatively small size of 80 resulted in a quantized integer LiteRT DNN with minimal quantization error. The inference of both non-quantized and quantized LiteRT DNNs on a low-resource ARM Cortex-M7 microcontroller (STM32F7) shows rhythm accuracy deviation of <0.4%. Quantization reduces internal computation latency from 4.8 s to 0.6 s, flash memory usage from 40 kB to 20 kB, and energy consumption by 7.85 times. This study ensures that DNN models retain their functionality while being optimized for real-time execution on resource-constrained hardware, demonstrating application in automated external defibrillators.

The InterCriteria Analysis (ICrA) is based on the mathematical formalisms of index matrices and intuitionistic fuzzy sets. It has been elaborated to discern possible similarities in the behavior of criteria pairs when multiple objects are considered, allowing also the accounting of information uncertainty. The focus of this study is to validate the applicability of ICrA over a large set of ECG criteria extracted for arrhythmia analysis and to evaluate its ability to support the pre-selection of criteria that could be further involved in decision making procedures. ICrA is applied over 88 ECG criteria (resulting in 3828 criteria pairs) calculated for 8528 ECGs from PhysioNet/CinC Challenge 2017 database. Three criteria pairs show strong positive consonance, another 26—positive consonance, while another 15 are in negative consonance. ICrA also reveals lack of dependencies in 98 criteria pairs. The correspondence between our observations (high degrees of agreement/disagreement and lack of dependencies) and our expectations based on knowledge of the principles involved in the computation of the ECG criteria validates the application of ICrA for reliable evaluation of the relation between different criteria. This potential of ICrA to highlight useful relations between ECG criteria makes it suitable in the ECG pre-processing stage for criteria pre-selection. Thus, optimization of the feature space could be achieved together with minimization of the computations’ complexity.

Background/Objectives: The timely and accurate detection of atrial fibrillation (AF) is critical from a clinical perspective. Detecting short or transient AF events is challenging in 24–72 h Holter ECG recordings, especially when symptoms are infrequent. This study aims to explore the potential of deep transfer learning with ImageNet deep neural networks (DNNs) to improve the interpretation of short-term ECHOView images for the presence of AF. Methods: Thirty-second ECHOView images, composed of stacked heartbeat amplitudes, were rescaled to fit the input of 18 pretrained ImageNet DNNs with the top layers modified for binary classification (AF, non-AF). Transfer learning provided both retrained DNNs by training only the top layers (513–2048 trainable parameters) and fine-tuned DNNs by slowly training retrained DNNs (0.38–23.48 M parameters). Results: Transfer learning used 13,536 training and 6624 validation samples from the two leads in the IRIDIA-AF Holter ECG database, evenly split between AF and non-AF cases. The top-ranked DNNs evaluated on 11,400 test samples from independent records are the retrained EfficientNetV2B1 (96.3% accuracy with minimal inter-patient (1%) and inter-lead (0.3%) drops), and fine-tuned EfficientNetV2B1 and DenseNet-121, -169, -201 (97.2–97.6% accuracy with inter-patient (1.4–1.6%) and inter-lead (0.5–1.2%) drops). These models can process shorter ECG episodes with a tolerable accuracy drop of up to 0.6% for 20 s and 4–15% for 10 s. Case studies present the GradCAM heatmaps of retrained EfficientNetV2B1 overlaid on raw ECG and ECHOView images to illustrate model interpretability. Conclusions: In an extended deep transfer learning study, we validate that ImageNet DNNs applied to short-term ECHOView images through retraining and fine-tuning can significantly enhance automated AF diagnoses. GradCAM heatmaps provide meaningful model interpretability, highlighting ECG regions of interest aligned with cardiologist focus.

The aim of this study is to address the challenge of 12-lead ECG delineation by different encoder–decoder architectures of deep neural networks (DNNs). This study compares four concepts for encoder–decoders based on a fully convolutional architecture (CED-Net) and its modifications with a recurrent layer (CED-LSTM-Net), residual connections between symmetrical encoder and decoder feature maps (CED-U-Net), and sequential residual blocks (CED-Res-Net). All DNNs transform 12-lead representative beats to three diagnostic ECG intervals (P-wave, QRS-complex, QT-interval) used for the global delineation of the representative beat (P-onset, P-offset, QRS-onset, QRS-offset, T-offset). All DNNs were trained and optimized using the large PhysioNet ECG database (PTB-XL) under identical conditions, applying an advanced approach for machine-based supervised learning with a reference algorithm for ECG delineation (ETM, Schiller AG, Baar, Switzerland). The test results indicate that all DNN architectures are equally capable of reproducing the reference delineation algorithm’s measurements in the diagnostic PTB database with an average P-wave detection accuracy (96.6%) and time and duration errors: mean values (−2.6 to 2.4 ms) and standard deviations (2.9 to 11.4 ms). The validation according to the standard-based evaluation practices of diagnostic electrocardiographs with the CSE database outlines a CED-Net model, which measures P-duration (2.6 ± 11.0 ms), PQ-interval (0.9 ± 5.8 ms), QRS-duration (−2.4 ± 5.4 ms), and QT-interval (−0.7 ± 10.3 ms), which meet all standard tolerances. Noise tests with high-frequency, low-frequency, and power-line frequency noise (50/60 Hz) confirm that CED-Net, CED-Res-Net, and CED-LSTM-Net are robust to all types of noise, mostly presenting a mean duration error < 2.5 ms when compared to measurements without noise. Reduced noise immunity is observed for the U-net architecture. Comparative analysis with other published studies scores this research within the lower range of time errors, highlighting its competitive performance.

Gait analysis is a non-invasive, cost-effective method for detecting subtle motor changes in neurodegenerative disorders. This study uses an exploratory approach to identify temporal–kinetic gait feature relationships specific to amyotrophic lateral sclerosis (ALS) and Huntington (HUNT) and Parkinson (PARK) disease versus healthy controls (CONTROL) using recent advances in InterCriteria Analysis (ICrA). The novelty lies in the (i) comprehensive temporal–kinetic feature set, (ii) use of ICrA to characterize inter-feature coordination patterns at population and disease-group levels and (iii) interpretation in a neuromechanical context. Forty-one temporal/kinetic features were extracted from left/right leg ground reaction force and rate-of-force-development signals, considering laterality, gait phase (stance, swing, double support), magnitudes, waveform correlations, and inter-/intra-limb asymmetries. The analysis included 14,580 steps from 64 recordings in the Gait in Neurodegenerative Disease Database: 16 CONTROL (4054 steps), 13 ALS (2465), 20 HUNT (4730), 15 PARK (3331). Sensitivity analysis identified strict consonance thresholds (μ ≥ 0.75, ν ≤ 0.25), selecting <5% strongest inter-feature relations from 820 feature pairs: population level (16 positive, 14 negative), group-level (15–25 positive, 9–14 negative). ICrA identified group-specific consonances—present in one group but absent in others—highlighting disease-related alterations in gait coordination: ALS (15/11 positive/negative, disrupted bilateral stride coordination, prolonged stance/double-support, decoupled stride/cadence, desynchronized force-generation patterns—reflecting compensatory adaptations to muscle weakness and instability), HUNT (11/7, severe temporal–kinetic breakdown consistent with gait instability—loss of bilateral coordination, reduced swing time, slowed force development), PARK (1/2, subtle localized disruptions—prolonged stance and double-support intervals, reduced force during weight transfer, overall coordination remained largely preserved). Benchmarking vs. Pearson correlation showed strong linear agreement (R² = 0.847, p < 0.001), confirming that ICrA captures dominant dependencies while moderating the correlation via uncertainty. These results demonstrate that ICrA provides a quantitative, interpretable framework for characterizing gait coordination patterns and can guide principled feature selection in future predictive modeling.