Enhancing Cardiovascular Risk Stratification: Radiomics of Coronary Plaque and Perivascular Adipose Tissue - Current Insights and Future Perspectives.

Radiomics, the quantitative extraction and mining of features from radiological images, has recently emerged as a promising source of non-invasive image-based cardiovascular biomarkers, potentially revolutionizing diagnostics and risk assessment. This review explores its application within coronary plaques and pericoronary adipose tissue, particularly focusing on plaque characterization and cardiac events prediction. By shedding light on the current state-of-the-art, achievements, and prospective avenues, this review contributes to a deeper understanding of the evolving landscape of radiomics in the context of coronary arteries. Finally, open challenges and existing gaps are emphasized to underscore the need for future efforts aimed at ensuring the robustness and reliability of radiomics studies, facilitating their clinical translation.

Effect of variability of mechanical properties on the predictive capabilities of vulnerable coronary plaques.

BACKGROUND AND OBJECTIVE: Coronary plaque rupture is a precipitating event responsible for two thirds of myocardial infarctions. Currently, the risk of plaque rupture is computed based on demographic, clinical, and image-based adverse features. However, using these features the absolute event rate per single higher-risk lesion remains low. This work studies the power of a novel framework based on biomechanical markers accounting for material uncertainty to stratify vulnerable and non-vulnerable coronary plaques. METHODS: Virtual histology intravascular ultrasounds from 55 patients, 29 affected by acute coronary syndrome and 26 affected by stable angina pectoris, were included in this study. Two-dimensional vessel cross-sections for finite element modeling (10 sections per plaque) incorporating plaque structure (medial tissue, loose matrix, lipid core and calcification) were reconstructed. A Montecarlo finite element analysis was performed on each section to account for material variability on three biomechanical markers: peak plaque structural stress at diastolic and systolic pressure, and peak plaque stress difference between systolic and diastolic pressures, together with the luminal pressure. Machine learning decision tree classifiers were trained on 75% of the dataset and tested on the remaining 25% with a combination of feature selection techniques. Performance against classification trees based on geometric markers (i.e., luminal, external elastic membrane and plaque areas) was also performed. RESULTS: Our results indicate that the plaque structural stress outperforms the classification capacity of the combined geometric markers only (0.82 vs 0.51 area under curve) when accounting for uncertainty in material parameters. Furthermore, the results suggest that the combination of the peak plaque structural stress at diastolic and systolic pressures with the maximum plaque structural stress difference between systolic and diastolic pressures together with the systolic pressure and the diastolic to systolic pressure gradient is a robust classifier for coronary plaques when the intrinsic variability in material parameters is considered (area under curve equal to [0.91-0.93]). CONCLUSION: In summary, our results emphasize that peak plaque structural stress in combination with the patient's luminal pressure is a potential classifier of plaque vulnerability as it independently considers stress in all directions and incorporates total geometric and compositional features of atherosclerotic plaques.

Detection of Non-Sustained Supraventricular Tachycardia in Atrial Fibrillation Screening.

OBJECTIVE: Non-sustained supraventricular tachycardia (nsSVT) is associated with a higher risk of developing atrial fibrillation (AF), and, therefore, detection of nsSVT can improve AF screening efficiency. However, the detection is challenged by the lower signal quality of ECGs recorded using handheld devices and the presence of ectopic beats which may mimic the rhythm characteristics of nsSVT. METHODS: The present study introduces a new nsSVT detector for use in single-lead, 30-s ECGs, based on the assumption that beats in an nsSVT episode exhibits similar morphology, implying that episodes with beats of deviating morphology, either due to ectopic beats or noise/artifacts, are excluded. A support vector machine is used to classify successive 5-beat sequences in a sliding window with respect to similar morphology. Due to the lack of adequate training data, the classifier is trained using simulated ECGs with varying signal-to-noise ratio. In a subsequent step, a set of rhythm criteria is applied to similar beat sequences to ensure that episode duration and heart rate is acceptable. RESULTS: The performance of the proposed detector is evaluated using the StrokeStop II database, resulting in sensitivity, specificity, and positive predictive value of 84.6%, 99.4%, and 18.5%, respectively. CONCLUSION: The results show that a significant reduction in expert review burden (factor of 6) can be achieved using the proposed detector.Clinical and Translational Impact: The reduction in the expert review burden shows that nsSVT detection in AF screening can be made considerably more efficiently.

DEep LearnIng-based QuaNtification of epicardial adipose tissue predicts MACE in patients undergoing stress CMR.

BACKGROUND AND AIMS: This study investigated the additional prognostic value of epicardial adipose tissue (EAT) volume for major adverse cardiovascular events (MACE) in patients undergoing stress cardiac magnetic resonance (CMR) imaging. METHODS: 730 consecutive patients [mean age: 63 ± 10 years; 616 men] who underwent stress CMR for known or suspected coronary artery disease were randomly divided into derivation (n = 365) and validation (n = 365) cohorts. MACE was defined as non-fatal myocardial infarction and cardiac deaths. A deep learning algorithm was developed and trained to quantify EAT volume from CMR. EAT volume was adjusted for height (EAT volume index). A composite CMR-based risk score by Cox analysis of the risk of MACE was created. RESULTS: In the derivation cohort, 32 patients (8.7 %) developed MACE during a follow-up of 2103 days. Left ventricular ejection fraction (LVEF) < 35 % (HR 4.407 [95 % CI 1.903-10.202]; p<0.001), stress perfusion defect (HR 3.550 [95 % CI 1.765-7.138]; p<0.001), late gadolinium enhancement (LGE) (HR 4.428 [95%CI 1.822-10.759]; p = 0.001) and EAT volume index (HR 1.082 [95 % CI 1.045-1.120]; p<0.001) were independent predictors of MACE. In a multivariate Cox regression analysis, adding EAT volume index to a composite risk score including LVEF, stress perfusion defect and LGE provided additional value in MACE prediction, with a net reclassification improvement of 0.683 (95%CI, 0.336-1.03; p<0.001). The combined evaluation of risk score and EAT volume index showed a higher Harrel C statistic as compared to risk score (0.85 vs. 0.76; p<0.001) and EAT volume index alone (0.85 vs.0.74; p<0.001). These findings were confirmed in the validation cohort. CONCLUSIONS: In patients with clinically indicated stress CMR, fully automated EAT volume measured by deep learning can provide additional prognostic information on top of standard clinical and imaging parameters.

MRI radiomics in head and neck cancer from reproducibility to combined approaches.

The clinical applicability of radiomics in oncology depends on its transferability to real-world settings. However, the absence of standardized radiomics pipelines combined with methodological variability and insufficient reporting may hamper the reproducibility of radiomic analyses, impeding its translation to clinics. This study aimed to identify and replicate published, reproducible radiomic signatures based on magnetic resonance imaging (MRI), for prognosis of overall survival in head and neck squamous cell carcinoma (HNSCC) patients. Seven signatures were identified and reproduced on 58 HNSCC patients from the DB2Decide Project. The analysis focused on: assessing the signatures' reproducibility and replicating them by addressing the insufficient reporting; evaluating their relationship and performances; and proposing a cluster-based approach to combine radiomic signatures, enhancing the prognostic performance. The analysis revealed key insights: (1) despite the signatures were based on different features, high correlations among signatures and features suggested consistency in the description of lesion properties; (2) although the uncertainties in reproducing the signatures, they exhibited a moderate prognostic capability on an external dataset; (3) clustering approaches improved prognostic performance compared to individual signatures. Thus, transparent methodology not only facilitates replication on external datasets but also advances the field, refining prognostic models for potential personalized medicine applications.

ECG-based data-driven solutions for diagnosis and prognosis of cardiovascular diseases: A systematic review.

Cardiovascular diseases (CVD) are a leading cause of death globally, and result in significant morbidity and reduced quality of life. The electrocardiogram (ECG) plays a crucial role in CVD diagnosis, prognosis, and prevention; however, different challenges still remain, such as an increasing unmet demand for skilled cardiologists capable of accurately interpreting ECG. This leads to higher workload and potential diagnostic inaccuracies. Data-driven approaches, such as machine learning (ML) and deep learning (DL) have emerged to improve existing computer-assisted solutions and enhance physicians' ECG interpretation of the complex mechanisms underlying CVD. However, many ML and DL models used to detect ECG-based CVD suffer from a lack of explainability, bias, as well as ethical, legal, and societal implications (ELSI). Despite the critical importance of these Trustworthy Artificial Intelligence (AI) aspects, there is a lack of comprehensive literature reviews that examine the current trends in ECG-based solutions for CVD diagnosis or prognosis that use ML and DL models and address the Trustworthy AI requirements. This review aims to bridge this knowledge gap by providing a systematic review to undertake a holistic analysis across multiple dimensions of these data-driven models such as type of CVD addressed, dataset characteristics, data input modalities, ML and DL algorithms (with a focus on DL), and aspects of Trustworthy AI like explainability, bias and ethical considerations. Additionally, within the analyzed dimensions, various challenges are identified. To these, we provide concrete recommendations, equipping other researchers with valuable insights to understand the current state of the field comprehensively.

Developing a robust two-step machine learning multiclassification pipeline to predict primary site in head and neck carcinoma from lymph nodes.

This study aimed to develop a robust multiclassification pipeline to determine the primary tumor location in patients with head and neck carcinoma of unknown primary using radiomics and machine learning techniques. The dataset included 400 head and neck cancer patients with primary tumor in oropharynx, OPC (n = 162), nasopharynx, NPC (n = 137), oral cavity, OC (n = 63), larynx and hypopharynx, HL (n = 38). Two radiomic-based multiclassification pipelines (P1 and P2) were developed. P1 consisted in a direct identification of the primary sites, whereas P2 was based on a two-step approach: in the first step, the number of classes was reduced by merging the two minority classes which were reclassified in the second step. Diverse correlation thresholds (0.75, 0.80, 0.85), feature selection methods (sequential forwards/backwards selection, sequential floating forward selection, neighborhood component analysis and minimum redundancy maximum relevance), and classification models (neural network, decision tree, naïve Bayes, bagged trees and support vector machine) were assessed. P2 outperformed P1, with the best results obtained with the support vector machine classifier including radiomic and clinical features (accuracies of 75.3 % (HL), 75.4 % (OC), 71.3 % (OPC), 92.9 % (NPC)). These results indicate that the two-step multiclassification pipeline integrating radiomics and clinical information is a promising approach to predict the tumor site of unknown primary.

AI-based hip prosthesis failure prediction through evolutional radiological indices.

INTRODUCTION: This study aimed to develop artificial intelligence models for predicting hip implant failure from radiological features. Analyzing the evolution of the periprosthetic bone and implant's position throughout the entire follow-up period has shown the potential to be more relevant in outcome prediction than simply considering the latest radiographic images. Thus, we investigated an AI-based model employing a small set of evolutional parameters derived from conventional radiological features to predict hip prosthesis failure. MATERIALS AND METHODS: One hundred sixty-nine radiological features were annotated from historical anteroposterior and lateral radiographs for 162 total hip arthroplasty patients, 32 of which later underwent implant failure. Linear regression on each patient's chronologically sorted radiological features was employed to derive 169 corresponding evolutional parameters per image. Three sets of machine learning predictors were developed: one employing the original features (standard model), one the evolutional ones (evolutional model), and the last their union (hybrid model). Each set included a model employing all the available features (full model) and a model employing the few most predictive ones according to Gini importance (minimal model). RESULTS: The evolutional and hybrid predictors resulted highly effective (area under the ROC curve (AUC) of full models = 0.94), outperforming the standard one, whose AUC was only 0.82. The minimal hybrid model, employing just four features, three of which evolutional, scored an AUC of 0.95, proving even more accurate than the full one, exploiting 173 features. This tool could be shaped to be either highly specific (sensitivity: 80%, specificity: 98.6%) or highly sensitive (sensitivity: 90%, specificity: 92.4%). CONCLUSION: The proposed predictor may represent a highly sensitive screening tool for clinicians, capable to predict THA failure with an advance between a few months and more than a year through only four radiological parameters, considering either their value at the latest visit or their evolution through time.

A robust radiomic-based machine learning approach to detect cardiac amyloidosis using cardiac computed tomography.

Introduction: Cardiac amyloidosis (CA) shares similar clinical and imaging characteristics (e.g., hypertrophic phenotype) with aortic stenosis (AS), but its prognosis is generally worse than severe AS alone. Recent studies suggest that the presence of CA is frequent (1 out of 8 patients) in patients with severe AS. The coexistence of the two diseases complicates the prognosis and therapeutic management of both conditions. Thus, there is an urgent need to standardize and optimize the diagnostic process of CA and AS. The aim of this study is to develop a robust and reliable radiomics-based pipeline to differentiate the two pathologies. Methods: Thirty patients were included in the study, equally divided between CA and AS. For each patient, a cardiac computed tomography (CCT) was analyzed by extracting 107 radiomics features from the LV wall. Feature robustness was evaluated by means of geometrical transformations to the ROIs and intra-class correlation coefficient (ICC) computation. Various correlation thresholds (0.80, 0.85, 0.90, 0.95, 1), feature selection methods [p-value, least absolute shrinkage and selection operator (LASSO), semi-supervised LASSO, principal component analysis (PCA), semi-supervised PCA, sequential forwards selection] and machine learning classifiers (k-nearest neighbors, support vector machine, decision tree, logistic regression and gradient boosting) were assessed using a leave-one-out cross-validation. Data augmentation was performed using the synthetic minority oversampling technique. Finally, explainability analysis was performed by using the SHapley Additive exPlanations (SHAP) method. Results: Ninety-two radiomic features were selected as robust and used in the further steps. Best performances of classification were obtained using a correlation threshold of 0.95, PCA (keeping 95% of the variance, corresponding to 9 PCs) and support vector machine classifier reaching an accuracy, sensitivity and specificity of 0.93. Four PCs were found to be mainly dependent on textural features, two on first-order statistics and three on shape and size features. Conclusion: These preliminary results show that radiomics might be used as non-invasive tool able to differentiate CA from AS using clinical routine available images.

MRI-based radiomic prognostic signature for locally advanced oral cavity squamous cell carcinoma: development, testing and comparison with genomic prognostic signatures.

BACKGROUND: . At present, the prognostic prediction in advanced oral cavity squamous cell carcinoma (OCSCC) is based on the tumor-node-metastasis (TNM) staging system, and the most used imaging modality in these patients is magnetic resonance image (MRI). With the aim to improve the prediction, we developed an MRI-based radiomic signature as a prognostic marker for overall survival (OS) in OCSCC patients and compared it with published gene expression signatures for prognosis of OS in head and neck cancer patients, replicated herein on our OCSCC dataset. METHODS: For each patient, 1072 radiomic features were extracted from T1 and T2-weighted MRI (T1w and T2w). Features selection was performed, and an optimal set of five of them was used to fit a Cox proportional hazard regression model for OS. The radiomic signature was developed on a multi-centric locally advanced OCSCC retrospective dataset (n = 123) and validated on a prospective cohort (n = 108). RESULTS: The performance of the signature was evaluated in terms of C-index (0.68 (IQR 0.66-0.70)), hazard ratio (HR 2.64 (95% CI 1.62-4.31)), and high/low risk group stratification (log-rank p < 0.001, Kaplan-Meier curves). When tested on a multi-centric prospective cohort (n = 108), the signature had a C-index of 0.62 (IQR 0.58-0.64) and outperformed the clinical and pathologic TNM stage and six out of seven gene expression prognostic signatures. In addition, the significant difference of the radiomic signature between stages III and IVa/b in patients receiving surgery suggests a potential association of MRI features with the pathologic stage. CONCLUSIONS: Overall, the present study suggests that MRI signatures, containing non-invasive and cost-effective remarkable information, could be exploited as prognostic tools.

Combining deep learning and machine learning for the automatic identification of hip prosthesis failure: Development, validation and explainability analysis.

AIM: Revision hip arthroplasty has a less favorable outcome than primary total hip arthroplasty and an understanding of the timing of total hip arthroplasty failure may be helpful. The aim of this study is to develop a combined deep learning (DL) and machine learning (ML) approach to automatically detect hip prosthetic failure from conventional plain radiographs. METHODS: Two cohorts of patients (of 280 and 352 patients) were included in the study, for model development and validation, respectively. The analysis was based on one antero-posterior and one lateral radiographic view obtained from each patient during routine post-surgery follow-up. After pre-processing, three images were obtained: the original image, the acetabulum image and the stem image. These images were analyzed through convolutional neural networks aiming to predict prosthesis failure. Deep features of the three images were extracted for each model and two feature-based pipelines were developed: one utilizing only the features of the original image (original image pipeline) and the other concatenating the features of the three images (3-image pipeline). The obtained features were either used directly or reduced through principal component analysis. Both support vector machine (SVM) and random forest (RF) classifiers were considered for each pipeline. RESULTS: The SVM applied to the 3-image pipeline provided the best performance, with an accuracy of 0.958 ± 0.006 in the internal validation and an F1-score of 0.874 in the external validation set. The explainability analysis, besides identifying the features of the complete original images as the major contributor, highlighted the role of the acetabulum and stem images on the prediction. CONCLUSIONS: This study demonstrated the potentialities of the developed DL-ML procedure based on plain radiographs in the detection of the failure of the hip prosthesis.

Prognostic radiomic signature for head and neck cancer: Development and validation on a multi-centric MRI dataset.

BACKGROUND AND PURPOSE: Prognosis in locally advanced head and neck cancer (HNC) is currently based on TNM staging system and tumor subsite. However, quantitative imaging features (i.e., radiomic features) from magnetic resonance imaging (MRI) may provide additional prognostic info. The aim of this work is to develop and validate an MRI-based prognostic radiomic signature for locally advanced HNC. MATERIALS AND METHODS: Radiomic features were extracted from T1- and T2-weighted MRI (T1w and T2w) using the segmentation of the primary tumor as mask. In total 1072 features (536 per image type) were extracted for each tumor. A retrospective multi-centric dataset (n = 285) was used for features selection and model training. The selected features were used to fit a Cox proportional hazard regression model for overall survival (OS) that outputs the radiomic signature. The signature was then validated on a prospective multi-centric dataset (n = 234). Prognostic performance for OS and disease-free survival (DFS) was evaluated using C-index. Additional prognostic value of the radiomic signature was explored. RESULTS: The radiomic signature had C-index = 0.64 for OS and C-index = 0.60 for DFS in the validation set. The addition of the radiomic signature to other clinical features (TNM staging and tumor subsite) increased prognostic ability for both OS (HPV- C-index 0.63 to 0.65; HPV+ C-index 0.75 to 0.80) and DFS (HPV- C-index 0.58 to 0.61; HPV+ C-index 0.64 to 0.65). CONCLUSION: An MRI-based prognostic radiomic signature was developed and prospectively validated. Such signature can successfully integrate clinical factors in both HPV+ and HPV- tumors.

Heart rate characteristic based modelling of atrial fibrillatory rate using implanted cardiac monitor data.

Objective. The objective of the present study is to investigate the feasibility of using heart rate characteristics to estimate atrial fibrillatory rate (AFR) in a cohort of atrial fibrillation (AF) patients continuously monitored with an implantable cardiac monitor. We will use a mixed model approach to investigate population effect and patient specific effects of heart rate characteristics on AFR, and will correct for the effect of previous ablations, episode duration, and onset date and time.Approach. The f-wave signals, from which AFR is estimated, were extracted using a QRST cancellation process of the AF episodes in a cohort of 99 patients (67% male; 57 ± 12 years) monitored for 9.2(0.2-24.3) months as median(min-max). The AFR from 2453 f-wave signals included in the analysis was estimated using a model-based approach. The association between AFR and heart rate characteristics, prior ablations, and episode-related features were modelled using fixed-effect and mixed-effect modelling approaches.Main results. The mixed-effect models had a better fit to the data than fixed-effect models showing h.c. of determination (R2 = 0.49 versusR2 = 0.04) when relating the variations of AFR to the heart rate features. However, when correcting for the other factors, the mixed-effect model showed the best fit (R2 = 0.04). AFR was found to be significantly affected by previous catheter ablations (p< 0.05), episode duration (p< 0.05), and irregularity of theRRinterval series (p< 0.05).Significance. Mixed-effect models are more suitable for AFR modelling. AFR was shown to be faster in episodes with longer duration, less organizedRRintervals and after several ablation procedures.

Atrial fibrillation episode patterns as predictor of clinical outcome of catheter ablation.

Methods for characterization of atrial fibrillation (AF) episode patterns have been introduced without establishing clinical significance. This study investigates, for the first time, whether post-ablation recurrence of AF can be predicted by evaluating episode patterns. The dataset comprises of 54 patients (age 56 ± 11 years; 67% men), with an implantable cardiac monitor, before undergoing the first AF catheter ablation. Two parameters of the alternating bivariate Hawkes model were used to characterize the pattern: AF dominance during the monitoring period (log(mu)) and temporal aggregation of episodes (beta1). Moreover, AF burden and AF density, a parameter characterizing aggregation of AF burden, were studied. The four parameters were computed from an average of 29 AF episodes before ablation. The risk of AF recurrence after catheter ablation using the Hawkes parameters log(mu) and beta1, AF burden, and AF density was evaluated. While the combination of AF burden and AF density is related to a non-significant hazard ratio, the combination of log(mu) and beta1 is related to a hazard ratio of 1.95 (1.03-3.70; p < 0.05). The Hawkes parameters showed increased risk of AF recurrence within 1 year after the procedure for patients with high AF dominance and high episode aggregation and may be used for pre-ablation risk assessment.

Repeatability and reproducibility of MRI-radiomic features: A phantom experiment on a 1.5 T scanner.

PURPOSE: Aim of this study is to assess the repeatability of radiomic features on magnetic resonance images (MRI) and their stability to variations in time of repetition (TR), time of echo (TE), slice thickness (ST), and pixel spacing (PS) using vegetable phantoms. METHODS: The organic phantom was realized using two cucumbers placed inside a cylindrical container, and the analysis was performed using T1-weighted (T1w), T2-weighted (T2w), and diffusion-weighted images. One dataset was used to test the repeatability of the radiomic features, whereas other four datasets were used to test the sensitivity of the different MRI sequences to image acquisition parameters (TR, TE, ST, and PS). Four regions of interest (ROIs) were segmented: two for the central part of each cucumber and two for the external parts. Radiomic features were extracted from each ROI using Pyradiomics. To assess the effect of preprocessing on the reduction of variability, features were extracted both before and after the preprocessing. The coefficient of variation (CV) and intra-class correlation coefficient (ICC) were used to evaluate variability. RESULTS: The use of intensity standardization increased the stability for the first-order statistics features. Shape and size features were always stable for all the analyses. Textural features were particularly sensitive to changes in ST and PS, although some increase in stability could be obtained by voxel size resampling. When images underwent image preprocessing, the number of stable features (ICC > 0.75 and mean absolute CV < 0.3) was 33 for apparent diffusion coefficient (ADC), 52 for T1w, and 73 for T2w. CONCLUSIONS: The most critical source of variability is related to changes in voxel size (either caused by changes in ST or PS). Preprocessing increases features stability to both test-retest and variation of the image acquisition parameters for all the types of analyzed MRI (T1w, T2w, and ADC), except for ST.

3D vs. 2D MRI radiomics in skeletal Ewing sarcoma: Feature reproducibility and preliminary machine learning analysis on neoadjuvant chemotherapy response prediction.

Objective: The extent of response to neoadjuvant chemotherapy predicts survival in Ewing sarcoma. This study focuses on MRI radiomics of skeletal Ewing sarcoma and aims to investigate feature reproducibility and machine learning prediction of response to neoadjuvant chemotherapy. Materials and methods: This retrospective study included thirty patients with biopsy-proven skeletal Ewing sarcoma, who were treated with neoadjuvant chemotherapy before surgery at two tertiary sarcoma centres. 7 patients were poor responders and 23 were good responders based on pathological assessment of the surgical specimen. On pre-treatment T1-weighted and T2-weighted MRI, 2D and 3D tumour segmentations were manually performed. Features were extracted from original and wavelet-transformed images. Feature reproducibility was assessed through small geometrical transformations of the regions of interest mimicking multiple manual delineations, and intraclass correlation coefficient >0.75 defined feature reproducibility. Feature selection also consisted of collinearity and significance analysis. After class balancing in the training cohort, three machine learning classifiers were trained and tested on unseen data using hold-out cross-validation. Results: 1303 (77%) 3D and 620 (65%) 2D radiomic features were reproducible. 4 3D and 4 2D features passed feature selection. Logistic regression built upon 3D features achieved the best performance with 85% accuracy (AUC=0.9) in predicting response to neoadjuvant chemotherapy. Conclusion: Compared to 2D approach, 3D MRI radiomics of Ewing sarcoma had superior reproducibility and higher accuracy in predicting response to neoadjuvant chemotherapy, particularly when using logistic regression classifier.

Ensemble classification combining ResNet and handcrafted features with three-steps training.

Objective.This work presents an ECG classifier for variable leads as a contribution to the Computing in Cardiology Challenge/CinC Challenge 2021. It aims to integrate deep and classic machine learning features into a single model, exploring the proper structure and training procedure.Approach.From the initial 88 253 signals, only 84 210 were included. Low quality and unscored recordings were excluded. Three different database subsets of 40 365 recording each were created by dividing in three normal sinus rhythm and sinus bradycardia recordings. Each subset was used to train a different model with shared architecture integrated as an ensemble to provide the final classification through major voting. Models contained a deep branch composed of a modified ResNet with dilation convolutional layers and squeeze and excitation Block that took as input windowed ECG signals. This was concatenated with a wide branch that integrated 20 cardiac rhythm features into a fully connected 3-layered network. Three different training steps were studied: just the deep branch (D), wide integration and training (D+W), and a final fine tuning of the deep branch posterior to wide training (D+W+D).Main Results.Results obtained in a local test set formed by a stratified 12.5% split of the given full dataset were presented for 2-lead and 12-lead models. The best training method was the 3-step D + W + D procedure obtaining a challenge metric of 0.709 and 0.677 for 12 and 2-lead models respectively.Significance.Integration of handcrafted features and deep learning model not only may increase the generalization capacity of the network but also provide a path to add explicit information into the classification decision process. To the best of our knowledge this is the first work studying the training procedure to properly integrate both types of information for ECG signals classification.