Macro F1 Score Research Articles

Explainable vision transformer for automatic visual sleep staging on multimodal PSG signals

Polysomnography (PSG) is crucial for diagnosing sleep disorders, but manual scoring of PSG is time-consuming and subjective, leading to high variability. While machine-learning models have improved PSG scoring, their clinical use is hindered by the ‘black-box’ nature. In this study, we present SleepXViT, an automatic sleep staging system using Vision Transformer (ViT) that provides intuitive, consistent explanations by mimicking human ‘visual scoring’. Tested on KISS–a PSG image dataset from 7745 patients across four hospitals–SleepXViT achieved a Macro F1 score of 81.94%, outperforming baseline models and showing robust performances on public datasets SHHS1 and SHHS2. Furthermore, SleepXViT offers well-calibrated confidence scores, enabling expert review for low-confidence predictions, alongside high-resolution heatmaps highlighting essential features and relevance scores for adjacent epochs’ influence on sleep stage predictions. Together, these explanations reinforce the scoring consistency of SleepXViT, making it both reliable and interpretable, thereby facilitating the synergy between the AI model and human scorers in clinical settings.

FldtMatch: Improving Unbalanced Data Classification via Deep Semi-Supervised Learning with Self-Adaptive Dynamic Threshold

Among the many methods of deep semi-supervised learning (DSSL), the holistic method combines ideas from other methods, such as consistency regularization and pseudo-labeling, with great success. This method typically introduces a threshold to utilize unlabeled data. If the highest predictive value from unlabeled data exceeds the threshold, the associated class is designated as the data’s pseudo-label. However, current methods utilize fixed or dynamic thresholds, disregarding the varying learning difficulties across categories in unbalanced datasets. To overcome these issues, in this paper, we first designed Cumulative Effective Labeling (CEL) to reflect a particular class’s learning difficulty. This approach differs from previous methods because it uses effective pseudo-labels and ground truth, collectively influencing the model’s capacity to acquire category knowledge. In addition, based on CEL, we propose a simple but effective way to compute the threshold, Self-adaptive Dynamic Threshold (SDT). It requires a single hyperparameter to adjust to various scenarios, eliminating the necessity for a unique threshold modification approach for each case. SDT utilizes a clever mapping function that can solve the problem of differential learning difficulty of various categories in an unbalanced image dataset that adversely affects dynamic thresholding. Finally, we propose a deep semi-supervised method with SDT called FldtMatch. Through theoretical analysis and extensive experiments, we have fully proven that FldtMatch can overcome the negative impact of unbalanced data. Regardless of the choice of the backbone network, our method achieves the best results on multiple datasets. The maximum improvement of the macro F1-Score metric is about 5.6% in DFUC2021 and 2.2% in ISIC2018.

P0567 Use of a convolutional neural network to quantify and predict ulcerative colitis disease endoscopic activity and histological outcomes

Abstract Background There is considerable inter-user variability in endoscopic and histologic scoring of the severity of ulcerative colitis (UC). The aim of this study was to create and train an artificial intelligence algorithm to predict the severity of UC using endoscopic videos. Methods This prospective cohort study recruited UC patients undergoing flexible sigmoidoscopy or colonoscopy between October 2023 and September 2024. The endoscopic examination was recorded from the splenic flexure to the anus. The videos were segmented into 3 second tubelets, and poor quality videos excluded. Objective disease activity was quantified by an experienced IBD gastroenterologist (by determining individual endoscopic Mayo Scores- eMS), and corresponding biopsies were scored according to the Nancy Histologic Index (NHI) by one of three gastrointestinal histopathologists. The dataset was divided into a training dataset (38 patients, 88% of video data) and a test dataset (6 patients, 12% of data), stratified by disease severity and equipment source. Two video vision transformer models were trained to recognise Mayo and NHI respectively, and their performance in differentiating disease activity from remission, and accuracy in determining endoscopic and histologic severity scores were evaluated. Standard diagnostic accuracy measures were used to calculate the models’ ability to detect and quantitate UC disease activity and included area under the receiver operating characteristic curve (AUROC), binary and multi-class accuracy, sensitivity, specificity, and precision. Results 53 patients were recruited, and after exclusions a total of 44 were included in the final analysis. 30 patients had endoscopic activity and 30 patients had histological activity. The resultant eMS and NHI showed moderate correlation (R=0.72). For eMS, the accuracy of the model was excellent and had an AUROC of 0.98 in identifying disease activity (eMS≥1), with accuracy, sensitivity, specificity, and precision of 92%, 86%, 96%, 93% respectively. Multiclass accuracy was 87%, with a macro F1 score of 0.85. For NHI, the model was highly accurate in detecting histopathologic activity (NHI≥1), with AUROC of 0.95, accuracy, sensitivity, specificity, precision at 89%, 92%, 86%, 85% respectively. However, it was less robust in predicting specific NHI scores, with multiclass accuracy at 69%, and a macro F1 score of 0.41. Conclusion We developed an AI algorithm that is able to reliably predict endoscopic and histologic disease activity for patients with UC, and accurately differentiate those patients from those in remission. This may reduce clinician reliance on biopsies to determine disease activity, and have far-reaching implications in UC disease assessment. References Hashash JG, Yu Ci Ng F, Farraye FA, Wang Y, Colucci DR, Baxi S, Muneer S, Reddan M, Shingru P, Melmed GY. Inter- and Intraobserver Variability on Endoscopic Scoring Systems in Crohn's Disease and Ulcerative Colitis: A Systematic Review and Meta-Analysis. Inflamm Bowel Dis. 2024 Mar 28:izae051. doi: 10.1093/ibd/izae051 Marchal-Bressenot A, Salleron J, Boulagnon-Rombi C, et al. Development and validation of the Nancy histological index for UC. Gut 2017;66:43-49.

Prediction on Mobile Phone Price Range using Classification Models

Function and pricing of mobile phones have always been a noticeable question to customers. Countless studies have been done for phone brands to set their standard on pricing new released phones. This paper analyzes the mobile price classification dataset on Kaggle dataset, by using three models: logistic regression, k-nearest neighbors, and support vector machine. Feature selection by considering the correlations between features are used as comparison to the models without feature selection. The performance of models is shown by accuracy and macro F1 score. The performance of models on training dataset and testing dataset are then compared. The conclusion by this paper is that two models of logistic regression have the best performance, followed by two models of k-nearest neighbors which has almost the same performance. Support vector machine with feature selection has a good performance, while it performs poorly without feature selection. It is concluded that feature selection improves the performance of the model significantly. The performance of models on training dataset and testing dataset are consistent.

Comparative Evaluation of IndoBERT, IndoBERTweet, and mBERT for Multilabel Student Feedback Classification

Student feedback plays a crucial role in enhancing the quality of educational programs, yet analyzing this feedback, especially in informal contexts, remains challenging. In Indonesia, where student comments often include colloquial language and vary widely in content, effective multilabel classification is essential to accurately identify the aspects of courses being critiqued. Despite the development of several BERT-based models, the effectiveness of these models for classifying informal Indonesian text remains underexplored. Here we evaluate the performance of three BERT variants—IndoBERT, IndoBERTweet, and mBERT—on the task of multilabel classification of student feedback. Our experiments investigate the impact of different sequence lengths and truncation strategies on model performance. We find that IndoBERTweet, with a macro F1-score of 0.8462, outperforms IndoBERT (0.8243) and mBERT (0.8230) when using a sequence length of 64 tokens and truncation at the end. These findings suggest that IndoBERTweet is well-suited for handling the informal, abbreviated text common in Indonesian student feedback, providing a robust tool for educational institutions aiming for actionable insights from student comments.

CARE-SD: classifier-based analysis for recognizing provider stigmatizing and doubt marker labels in electronic health records: model development and validation.

To detect and classify features of stigmatizing and biased language in intensive care electronic health records (EHRs) using natural language processing techniques. We first created a lexicon and regular expression lists from literature-driven stem words for linguistic features of stigmatizing patient labels, doubt markers, and scare quotes within EHRs. The lexicon was further extended using Word2Vec and GPT 3.5, and refined through human evaluation. These lexicons were used to search for matches across 18 million sentences from the de-identified Medical Information Mart for Intensive Care-III (MIMIC-III) dataset. For each linguistic bias feature, 1000 sentence matches were sampled, labeled by expert clinical and public health annotators, and used to supervised learning classifiers. Lexicon development from expanded literature stem-word lists resulted in a doubt marker lexicon containing 58 expressions, and a stigmatizing labels lexicon containing 127 expressions. Classifiers for doubt markers and stigmatizing labels had the highest performance, with macro F1-scores of 0.84 and 0.79, positive-label recall and precision values ranging from 0.71 to 0.86, and accuracies aligning closely with human annotator agreement (0.87). This study demonstrated the feasibility of supervised classifiers in automatically identifying stigmatizing labels and doubt markers in medical text and identified trends in stigmatizing language use in an EHR setting. Additional labeled data may help improve lower scare quote model performance. Classifiers developed in this study showed high model performance and can be applied to identify patterns and target interventions to reduce stigmatizing labels and doubt markers in healthcare systems.

Reduced Forgetfulness in Continual Learning for Named Entity Recognition Through Confident Soft-Label Imitation

Continual Learning for Named Entity Recognition (CL-NER) is a crucial task in recognizing emerging concepts when constructing real-world natural language processing applications. It involves sequentially updating an existing NER model with new entity types while retaining previously learned information. However, current CL methods are struggling with a major challenge called catastrophic forgetting. Owing to the semantic shift of the non-entity type, the issue is further intensified in NER. Most existing CL-NER methods rely on knowledge distillation through the output probabilities of previously learned entities, resulting in excessive stability (recognition of old entities) at the expense of plasticity (recognition of new entities). Some recent works further extend these methods by improving the distinction between old entities and non-entity types. Although these methods result in overall performance improvements, the preserved knowledge does not necessarily ensure the retention of task-related information for the oldest entities, which can lead to significant performance drops. To address this issue while maintaining overall performance, we propose a method called Confident Soft-Label Imitation (ConSOLI) for continual learning in NER. Inspired by methods that balance stability and plasticity, ConSOLI incorporates a soft-label distillation process and confident soft-label imitation learning. The former helps to gather the task-related knowledge in the old model and the latter further preserves the knowledge from diluting in the step-wise continual learning process. Moreover, ConSOLI demonstrates significant improvements in recognizing the oldest entity types, achieving Micro-F1 and Macro-F1 scores of up to 8.72 and 9.72, respectively, thus addressing the challenge of catastrophic forgetting in CL-NER.

BioElectra-BiLSTM-Dual Attention classifier for optimizing multilabel scientific literature classification

Abstract Scientific literature is growing in volume with time. The number of papers published each year by 28 100 journals is 2.5 million. The citation indexes and search engines are used extensively to find these publications. An individual receives many documents in response to a query, but only a few are relevant. The final documents lack structure due to inadequate indexing. Many systems index research papers using keywords instead of subject hierarchies. In the scientific literature classification paradigm, various multilabel classification methods have been proposed based on metadata features. The existing metadata-driven statistical measures use bag of words and traditional embedding techniques, like Word2Vec and BERT, which cannot quantify textual properties effectively. In this paper, we try to solve the limitations of existing classification techniques by unveiling the semantic context of the words using an advanced transformer-based recurrent neural networks (RNN) approach incorporating Dual Attention and layer-wise learning rate to enhance the classification performance. We propose a novel model, BioElectra-BiLSTM-Dual Attention that extracts the semantic features from the titles and abstracts of the research articles using BioElectra-encoder and then BILSTM layer along with Dual Attention label embeddings their correlation matrix and layer-wise learning rate strategy employed for performance enhancement. We evaluated the performance of the proposed model on the multilabel scientific literature LitCovid dataset and the results suggest that it significantly improves the macro-F1 and micro-F1 score as compared to the state-of-the-art baselines (ML-Net, Binary Bert, and LitMCBert).

End-to-end deep learning patient level classification of affected territory of ischemic stroke patients in DW-MRI.

To develop an end-to-end DL model for automated classification of affected territory in DWI of stroke patients. In this retrospective multicenter study, brain DWI studies from January 2017 to April 2020 from Center 1, from June 2020 to December 2020 from Center 2, and from November 2019 to April 2020 from Center 3 were included. Four radiologists labeled images into five classes: anterior cerebral artery (ACA), middle cerebral artery (MCA), posterior circulation (PC), and watershed (WS) regions, as well as normal images. Additionally, for Center 1, clinical information was encoded as a domain knowledge vector to incorporate into image embeddings. 3D convolutional neural network (CNN) and attention gate integrated versions for direct 3D encoding, long short-term memory (LSTM-CNN), and time-distributed layer for slice-based encoding were employed. Balanced classification accuracy, macro averaged f1 score, AUC, and interrater Cohen's kappa were calculated. Overall, 624 DWI MRIs from 3 centers were utilized (mean age, interval: 66.89 years, 29-95 years; 345 male) with 439 patients in the training, 103 in the validation, and 82 in the test sets. The best model was a slice-based parallel encoding model with 0.88 balanced accuracy, 0.80 macro-f1 score, and an AUC of 0.98. Clinical domain knowledge integration improved the performance with 0.93 best overall accuracy with parallel stream model embeddings and support vector machine classifiers. The mean kappa value for interrater agreement was 0.87. Developed end-to-end deep learning models performed well in classifying affected regions from stroke in DWI. The end-to-end deep learning model with a parallel stream encoding strategy for classifying stroke regions in DWI has performed comparably with radiologists.

Surrounding Rock Squeezing Classification in Underground Engineering Using a Hybrid Paradigm of Generative Artificial Intelligence and Deep Ensemble Learning

Surrounding rock squeezing is a common geological disaster in underground excavation projects (e.g., TBM tunneling and deep mining), which has adverse effects on construction safety, schedule, and property. To predict the squeezing of the surrounding rock accurately and quickly, this study proposes a hybrid machine learning paradigm that integrates generative artificial intelligence and deep ensemble learning. Specifically, conditional tabular generative adversarial network is devised to solve the problems of data shortage and class imbalance for data augmentation at the data level, and the deep random forest is built based on the augmented data for subsequent squeezing classification. A total of 139 historical squeezing cases are collected worldwide to validate the efficacy of the proposed modeling paradigm. The results reveal that this paradigm achieves a prediction accuracy of 92.86% and a macro F1-score of 0.9292. In particular, the individual F1-scores on strong squeezing and extremely strong squeezing are more than 0.9, with excellent prediction reliability for high-intensity squeezing. Finally, a comparative analysis with traditional machine learning techniques is conducted and the superiority of this paradigm is further verified. This study provides a valuable reference for surrounding rock squeezing classification under a limited data environment.

Fusion of residual network and t-SNE-CS for 2D visualization of open set recognition

Open set recognition (OSR) is a technique employed to ascertain whether unknown data belongs to a class in a database when the training class is incomplete. In addressing the OSR challenge associated with ADS-B leading pulse signals, this paper proposes a two-dimensional visualization of open set recognition (VOSR) approach that encompasses the stages of feature extraction, feature selection, and feature learning levels. At the feature extraction level, the I/Q features and phase features of the signal are selected; at the feature selection level, feature similarity analysis and mean decrease impurity-based random forest are employed; at the feature learning level, the framework of fusion residual network and the t-distributed stochastic neighbor embedding and circular surfaces (t-SNE-CS) strategy is constructed, and experiments are carried out on the close set data containing 20 classes of 10,229 samples and open set data containing 10 classes of 1,688 samples. Results show that the accuracy of the optimal combination of the residual network and the constructed features is 94.63% for the test set for the close set classification task. For the VOSR task, the accuracy of the test set is 93.69%, the open set recognition accuracy is 53.97% and Macro-F1 scores is 91.8%.

Sleep Stage Classification Via Multi-View Based Self-Supervised Contrastive Learning of EEG.

Self-supervised learning (SSL) is a challenging task in sleep stage classification (SSC) that is capable of mining valuable representations from unlabeled data. However, traditional SSL methods typically focus on single-view learning and do not fully exploit the interactions among information across multiple views. In this study, we focused on a multi-domain view of the same EEG signal and developed a self-supervised multi-view representation learning framework via time series and time-frequency contrasting (MV-TTFC). In the MV-TTFC framework, we built-in a cross-domain view contrastive learning prediction task to establish connections between the temporal view and time-frequency (TF) view, thereby enhancing the information exchange between multiple views. In addition, to improve the quality of the TF view inputs, we introduced an enhanced multisynchrosqueezing transform, which can create high energy concentration TF image views to compensate for the inaccurate representations in traditional TF processing techniques. Finally, integrating temporal, TF, and fusion space contrastive learning effectively captured the latent features in EEG signals. We evaluated MV-TTFC based on two real-world SSC datasets (SleepEDF-78 and SHHS) and compared it with baseline methods in downstream tasks. Our method exhibited state-of-the-art performance, achieving accuracies of 78.64% and 81.45% with SleepEDF-78 and SHHS, respectively, and macro F1-scores of 70.39% with SleepEDF-78 and 70.47% with SHHS.

Biomedical document-level relation extraction with thematic capture and localized entity pooling

In contrast to sentence-level relational extraction, document-level relation extraction poses greater challenges as a document typically contains multiple entities, and one entity may be associated with multiple other entities. Existing methods often rely on graph structures to capture path representations between entity pairs. However, this paper introduces a novel approach called local entity pooling that solely relies on the pre-training model to identify the bridge entity related to the current entity pair and generate the reasoning path representation. This technique effectively mitigates the multi-entity problem. Additionally, the model leverages the multi-entity and multi-label characteristics of the document to acquire the document’s thematic representation, thereby enhancing the document-level relation extraction task. Experimental evaluations conducted on two biomedical datasets, CDR and GDA. Our TCLEP (Thematic Capture and Localized Entity Pooling) model achieved the Macro-F1 scores of 71.7% and 85.3%, respectively. Simultaneously, we incorporated local entity pooling and thematic capture modules into the state-of-the-art model, resulting in performance improvements of 1.5% and 0.2% on the respective datasets. These results highlight the advanced performance of our proposed approach.

Monitoring Anthropogenically Disturbed Parcels with Soil Erosion Dynamics Change Based on an Improved SegFormer

Amidst burgeoning socioeconomic development, anthropogenic activities have exacerbated soil erosion. This erosion, characterized by its brief duration, high frequency, and considerable environmental degradation, presents a major challenge to ecological systems. Therefore, it is imperative to regulate and remediate erosion–prone, anthropogenically disturbed parcels, with dynamic change detection (CD) playing a crucial role in enhancing management efficiency. Currently, traditional methods for change detection, such as field surveys and visual interpretation, suffer from time inefficiencies, complexity, and high resource consumption. Meanwhile, despite advancements in remote sensing technology that have improved the temporal and spatial resolution of images, the complexity and heterogeneity of terrestrial cover types continue to limit large–scale dynamic monitoring of anthropogenically disturbed soil erosion parcels (ADPSE) using remote sensing techniques. To address this, we propose a novel ISegFormer model, which integrates the SegFormer network with a pseudo–residual multilayer perceptron (PR–MLP), cross–scale boundary constraint module (CSBC), and multiscale feature fusion module (MSFF). The PR–MLP module improves feature extraction by capturing spatial contextual information, while the CSBC module enhances boundary prediction through high– and low–level semantic guidance. The MSFF module fuses multiscale features with attention mechanisms, boosting segmentation precision for diverse change types. Model performance is evaluated using metrics, such as precision, recall, F1–score, intersection over union (IOU), and mean intersection over union (mIOU). The results demonstrate that our improved model performs exceptionally well in dynamic monitoring tasks for ADPSE. Compared to five other models, our model achieved an mIOU of 72.34% and a Macro–F1 score of 83.55% across twelve types of ADPSE changes, surpassing the other models by 1.52–2.48% in mIOU and 2.25–3.64% in Macro–F1 score. This work provides a theoretical and methodological foundation for policy–making in soil and water conservation departments.

Recognition analysis of spiral and straight-line drawings in tremor assessment.

No standard, objective diagnostic procedure exists for most neurological diseases causing tremors. Therefore, drawing tests have been widely analyzed to support diagnostic procedures. In this study, we examine the comparison of Archimedean spiral and line drawings, the possibilities of their joint application, and the relevance of displaying pressure on the drawings to recognize Parkinsonism and cerebellar dysfunction. We further attempted to use an automatic processing and evaluation system. Digital images were developed from raw data by adding or omitting pressure data. Pre-trained (MobileNet, Xception, ResNet50) models and a Baseline (from scratch) model were applied for binary classification with a fold cross-validation procedure. Predictions were analyzed separately by drawing tasks and in combination. The neurological diseases presented here can be recognized with a significantly higher macro f1 score from the spiral drawing task (up to 95.7 %) than lines (up to 84.3 %). A significant improvement can be achieved if the spiral is supplemented with line drawing. The pressure inclusion in the images did not result in significant information gain. The spiral drawing has a robust recognition power and can be supplemented with a line drawing task to increase the correct recognition. Moreover, X and Y coordinates appeared sufficient without pressure with this methodology.

Emoji multimodal microblog sentiment analysis based on mutual attention mechanism.

Emojis, utilizing visual means, mimic human facial expressions and postures to convey emotions and opinions. They are widely used in social media platforms such as Sina Weibo, and have become a crucial feature for sentiment analysis. However, existing approaches often treat emojis as special symbols or convert them into text labels, thereby neglecting the rich visual information of emojis. We propose a novel multimodal information integration model for emoji microblog sentiment analysis. To effectively leverage the emoji visual information, the model employs a text-emoji visual mutual attention mechanism. Experiments on a manually annotated microblog dataset show that compared to the baseline models without incorporating emoji visual information, the proposed model achieves improvements of 1.37% in macro F1 score and 2.30% in accuracy, respectively. To facilitate the related research, our corpus will be publicly available at https://github.com/yx100/Emojis/blob/main/weibo-emojis-annotation .

R3Design: deep tertiary structure-based RNA sequence design and beyond.

The rational design of Ribonucleic acid (RNA) molecules is crucial for advancing therapeutic applications, synthetic biology, and understanding the fundamental principles of life. Traditional RNA design methods have predominantly focused on secondary structure-based sequence design, often neglecting the intricate and essential tertiary interactions. We introduce R3Design, a tertiary structure-based RNA sequence design method that shifts the paradigm to prioritize tertiary structure in the RNA sequence design. R3Design significantly enhances sequence design on native RNA backbones, achieving high sequence recovery and Macro-F1 score, and outperforming traditional secondary structure-based approaches by substantial margins. We demonstrate that R3Design can design RNA sequences that fold into the desired tertiary structures by validating these predictions using advanced structure prediction models. This method, which is available through standalone software, provides a comprehensive toolkit for designing, folding, and evaluating RNA at the tertiary level. Our findings demonstrate R3Design's superior capability in designing RNA sequences, which achieves around $44\%$ in terms of both recovery score and Macro-F1 score in multiple datasets. This not only denotes the accuracy and fairness of the model but also underscores its potential to drive forward the development of innovative RNA-based therapeutics and to deepen our understanding of RNA biology.

Automatic TNM staging of colorectal cancer radiology reports using pre-trained language models

Background and Objective:Colorectal cancer is one of the major causes of cancer death worldwide. Essential for prognosis and treatment planning, TNM staging offers critical insights into the advancement of colorectal cancer. However, manual TNM staging from colon magnetic resonance imaging (MRI) is a laborious and error prone process. This study introduces an automated text classification system for TNM staging of colon MRI images in Spanish. Methods:A dataset of 1319 Spanish colon MRI reports was collected and manually labeled with TNM staging. In order to automate the task of TNM staging, a multimodal system was proposed. The system is based on RoBERTa language model pre-trained on a combination of biomedical and clinical Spanish language corpora and uses Natural Language Processing (NLP) techniques to extract relevant categorical and numerical features from MRI reports. Results:The performance of the system was evaluated using different metrics and the results obtained are very promising: the best performance among the proposed systems reached 0.7464, 0.8792 and 0.6776 of macro F1-score for T, N and M respectively. Conclusions:This study demonstrates the feasibility of using a language model for automatic TNM staging based on Spanish clinical reports of colorectal cancer patients. The proposed system can be a useful tool to improve the efficiency and accuracy of colorectal cancer diagnosis.

Empowering crisis information extraction through actionability event schemata and domain-adaptive pre-training

One of the persistent challenges in crisis detection is inferring actionable information to support emergency response. Existing methods focus on situational awareness but often lack actionable insights. This study proposes a holistic approach to implementing an actionability extraction system on social media, including requirement gathering, selection of machine learning tasks, data preparation, and integration with existing resources, providing guidance for governments, civil services, emergency workers, and researchers on supplementing existing channels with actionable information from social media. Our solution leverages an actionability schema and domain-adaptive pre-training, improving upon the state-of-the-art model by 5.5 % and 10.1 % in micro and macro F1 scores.

A Multi-Class ECG Signal Classifier Using a Binarized Depthwise Separable CNN with the Merged Convolution-Pooling Method.

Binarized convolutional neural networks (bCNNs) are favored for the design of low-storage, low-power cardiac arrhythmia classifiers owing to their high weight compression rate. However, multi-class classification of ECG signals based on bCNNs is challenging due to the accuracy loss introduced by the binarization operation. In this paper, an effective multi-classifier system is proposed for electrocardiogram (ECG) signals using a binarized depthwise separable convolutional neural network (bDSCNN) with the merged convolution-pooling (MCP) method. The binarized depthwise separable convolution layer is adopted to reduce the increased number of parameters in multi-classification systems. Instead of operating convolution and pooling sequentially as in a traditional convolutional neural network (CNN), the MCP method merges pooling together with convolution layers to reduce the number of computations. To further reduce hardware resources, this work employs blockwise incremental calculation to eliminate redundant storage with computations. In addition, the R peak interval data are integrated with P-QRS-T features to improve the classification accuracy. The proposed bDSCNN model is evaluated on an Intel DE1-SoC field-programmable gate array (FPGA), and the experimental results demonstrate that the proposed system achieves a five-class classification accuracy of 96.61% and a macro-F1 score of 89.08%, along with a dynamic power dissipation of 20 μW for five-category ECG signal classification. The hardware resource usage of BRAM and LUTs plus REGs is reduced by at least 2.94 and 1.74 times, respectively, compared with existing ECG classifiers using bCNN methods.