Activation Mapping Research Articles

Enhancing the Diagnostic Accuracy of Sacroiliitis: A Machine Learning Approach Applied to Computed Tomography Imaging.

Aims/Background Sacroiliitis is a challenging condition to diagnose accurately due to the subtle nature of its presentation in imaging studies. This study aims to improve the diagnostic accuracy of sacroiliitis by applying advanced machine learning techniques to computed tomography (CT) images. Methods We employed five convolutional neural network (CNN) models-Visual Geometry Group 16-layer Network (VGG16), ResNet101, DenseNet, Inception-v4, and ResNeXt-50-to analyze a dataset of 830 CT images, including both sacroiliitis and non-sacroiliitis cases. Each model's performance was evaluated using metrics such as accuracy, precision, recall, F1 score, Receiver Operating Characteristic (ROC), and Area Under the Curve (AUC). The interpretability of the models' decisions was enhanced using Gradient-weighted Class Activation Mapping (Grad-CAM) visualization. Results The ResNeXt-50 and Inception-v4 models demonstrated superior performance, achieving the highest accuracy and F1 scores among the tested models. Grad-CAM visualizations offered insights into the decision-making processes, highlighting the models' focus on relevant anatomical features critical for accurate diagnosis. Conclusion The use of CNN models, particularly ResNeXt-50 and Inception-v4, significantly improves the diagnosis of sacroiliitis from CT images. These models not only provide high diagnostic accuracy but also offer transparency in their decision-making processes, aiding clinicians in understanding and trusting Artificial Intelligence (AI)-driven diagnostics.

Evaluation of highly sensitive vibration states of nanomechanical resonators in liquid using a convolutional neural network

Nanomechanical resonators can detect various small physical quantities with high sensitivity using changes in resonant properties. However, viscous damping in liquids significantly reduces the measurement sensitivity. This study proposes convolutional neural network (CNN) vibration spectrum analysis to evaluate the highly sensitive vibration states of nanomechanical resonators, which are useful for in-liquid measurements. This research was carried out through the measurement of acetone concentration. First, we compared the concentration classification ability between the proposed and conventional methods and determined that the proposed method of analyzing vibration spectral changes using the CNN model can provide higher measurement sensitivity than the conventional measurement method of observing resonance properties changes and comparing the values for each measurement condition. This result shows that CNN-based spectral analysis is effective for the vibration spectra of in-liquid measurements. Next, gradient-weighted class activation mapping (Grad-CAM) was applied to verify which frequency bands are important for concentration classification in CNN model decision-making. The vibration states in these frequency bands were analyzed in terms of oscillation modes. This analysis revealed significant oscillation modes of the nanomechanical resonator in the liquid environment. Notably, in addition to the resonance states utilized in the conventional method, several other oscillation modes were found to be significant for measurements. This finding suggests that these oscillation modes may be highly sensitive for measurements in liquid environments. Among these oscillation modes, the mode with very small amplitude is highly promising for achieving unprecedented levels of sensitivity in sensing technologies.

Natural disasters detection using explainable deep learning

Deep learning applications have far-reaching implications in people’s daily lives. Disaster management professionals are becoming increasingly interested in applying deep learning to prepare for and respond to natural disasters. In this paper, we aim to assist natural disaster management professionals in preparing for disasters by developing a framework that can accurately classify natural disasters and interpret the results using a combination of a deep learning model and an XAI method to ensure reliability and ease of interpretation without a technical background. Two main aspects categorize the novelty of our work. The first is utilizing pre-trained Models such as VGGNet19, ResNet50, and ViT for accurate classification of natural disaster images. The second is implementing three explainable AI techniques-Gradient-weighted Class Activation Mapping (Grad-CAM), Grad CAM++, and Local Interpretable Model-agnostic Explanations (LIME) to ensure the interpretability of the model’s predictions, making the decision-making process transparent and reliable. Experiments on the Natural disaster datasets (Niloy et al. 2021) and MEDIC with a ViT-B-32 model achieved a high accuracy of 95.23%. Additionally, explainable artificial intelligence techniques such as LIME, Grad-CAM, and Grad-CAM++ are used to evaluate model performance and visualize decision-making. Our code is available at.11https://github.com/tariqshaban/disaster-classification-with-xai.

ConvNext as a Basis for Interpretability in Coffee Leaf Rust Classification

The increasing complexity of deep learning models can make it difficult to interpret and fit models beyond a purely accuracy-focused evaluation. This is where interpretable and [ eXplainable Artificial Intelligence (XAI)explainable Artificial Intelligence (AI) come into play to facilitate an understanding of the inner workings of models. Consequently, alternatives have emerged, such as [class activation mapping (CAM) techniquesClass Activation Maps (CAM) aimed at identifying regions of importance for an image classification model. However, the behavior of such models can be highly dependent on the type of architecture and the different variants of convolutional neural networks. Accordingly, this paper evaluates three Convolutional Neural Network (CNN) architectures (VGG16, ResNet50, ConvNext-T) against seven CAM models (GradCAM, XGradCAM, HiResCAM, LayerCAM, GradCAM++, GradCAMElementWise, and EigenCAM), indicating that the CAM maps obtained with ConvNext models show less variability among them, i.e., they are less dependent on the selected CAM approach. This study was performed on an image dataset for the classification of coffee leaf rust and evaluated using the RemOve And Debias (ROAD) metric.

Deep Learning-Based System Combining Chest X-Ray and Computerized Tomography Images for COVID-19 Diagnosis.

Aims/Background: The coronavirus disease 2019 (COVID-19) pandemic has highlighted the need for accurate and efficient diagnostic methods. This study aims to improve COVID-19 detection by integrating chest X-ray (CXR) and computerized tomography (CT) images using deep learning techniques, further improving diagnostic accuracy by using a combined imaging approach. Methods: The study used two publicly accessible databases, COVID-19 Questionnaires for Understanding the Exposure (COVID-QU-Ex) and Integrated Clinical and Translational Cancer Foundation (iCTCF), containing CXR and CT images, respectively. The proposed system employed convolutional neural networks (CNNs) for classification, specifically EfficientNet and ResNet architectures. The data underwent preprocessing steps, including image resizing, Gaussian noise addition, and data augmentation. The dataset was divided into training, validation, and test sets. Gradient-weighted Class Activation Mapping (Grad-CAM) was used for model interpretability. Results: The EfficientNet-based models outperformed the ResNet-based models across all metrics. The highest accuracy achieved was 99.44% for CXR images and 99.81% for CT images with EfficientNetB5. The models also demonstrated high precision, recall, and F1 scores. For statistical significance, the p-values were less than 0.05, indicating that the results are significant. Conclusion: Integrating CXR and CT images using deep learning significantly improves the accuracy of COVID-19 diagnosis. The EfficientNet-based models, with their superior feature extraction capabilities, show better performance than ResNet models. Grad-CAM Visualizations provide insights into the model's decision-making process, potentially reducing diagnostic errors and accelerating diagnosis processes. This approach can improve patient care and support healthcare systems in managing the pandemic more effectively.

Approaching Ventricular Tachycardia Ablation in 2024: An Update on Mapping and Ablation Strategies, Timing, and Future Directions.

This review provides insights into mapping and ablation strategies for VT, offering a comprehensive overview of contemporary approaches and future perspectives in the field. The strengths and limitations of classical mapping strategies, namely activation mapping, pace mapping, entrainment mapping, and substrate mapping, are deeply discussed. The increasing pivotal relevance of CMR and MDCT in substrate definition is highlighted, particularly in defining the border zone, tissue channels, and fat. The integration of CMR and MDCT images with EAM is explored, with a special focus on their role in enhancing effectiveness and procedure safety. The abstract concludes by illustrating the Pisa workflow for the VT ablation procedure.

Complementary branch fusing class and semantic knowledge for robust weakly supervised semantic segmentation

Leveraging semantically precise pseudo masks derived from image-level class knowledge for segmentation, namely image-level Weakly Supervised Semantic Segmentation (WSSS), remains challenging. Class Activation Maps (CAMs) using CNNs enhance WSSS by focusing on specific class parts like only the face of a human, whereas Vision Transformers (ViT) capture broader semantic parts but often miss complete class-specific details, such as human bodies with nearby objects like dogs. In this work, we propose a Complementary Branch (CoBra), a novel dual-branch framework consisting of two distinct architectures that provide valuable complementary knowledge of class (from CNN) and semantics (from ViT). In particular, we learn Class-Aware Projection (CAP) for the CNN branch and Semantic-Aware Projection (SAP) for the ViT branch, combining their insights to facilitate new patch-level supervision and create effective pseudo masks integrating class and semantic information. Extensive experiments qualitatively and quantitatively investigate how each branch complements the other, showing a significant result. Project Page and code are available: https://micv-yonsei.github.io/cobra2024/.

An explainable weakly supervised model for multi-disease detection and localization from thoracic X-rays

Thoracic diseases are a major source of mortality, often requiring diagnosis from plain chest X-rays. However, differentiating between complex conditions based on subtle radiographic patterns poses challenges even for experts. Recently, deep learning methods have shown promise in automating thoracic disease detection from chest radiographs. Many existing approaches focus on the diseased organs in the radiographs by utilizing spatial regions that contribute significantly to the model’s prediction. Expert radiologists, on the other hand, first identify the prominent region before determining whether those regions are abnormal or not. Therefore, incorporating localization information through deep learning models could result in significant improvements in automatic disease classification. Motivated by this, we have proposed a generalized weakly supervised Confidence-Aware Probabilistic Class Activation Map (CAPCAM) classification model that localizes anomalies for thoracic disease. The CAPCAM used CX-Ultranet as the backbone with the combination of Confidence Aware Network (CAN) and Anomaly Detection Network (ADN) without having any localization labeling. This learning from the backbone helps the model to utilize all components of the feature extracted and, therefore eliminating the need to train them individually reducing the time taken. We have experimentally shown that the proposed CAPCAM method sets a new state-of-the-art benchmark by achieving accuracy in terms of Intersection of bounding box (IoBB) in the range of 85 % - 94 %, and Dice scores in the range of 88 %-90 % for all thirteen diseases on two publicly available large-scale CXR datasets–NIH, Stanford and CheXpert. Testing across different noise levels and different levels of blurred level assessed real-world viability. We have also added a layer of explainability to show how the image is processed. This study demonstrates deep learning’s potential to augment radiologists’ decision-making by providing fast, accurate automated aids for thoracic disease diagnosis. The proposed CAPCAM model could be readily translatable to improve clinical workflows.

Prediction of Contact Resistance of 4H–SiC by Machine Learning Using Optical Microscope Images after Laser Doping

In this study, machine learning (ML) was employed to predict the electrical properties of finished devices, specifically focusing on the state of the contacts at the electrodes. The predictions are based on optical microscope images of the surface conditions, which were captured immediately following the laser doping of nitrogen atoms into 4H-SiC. The laser doping process involved varying the laser fluence from 0.4 to 4.0 J/cm2 and using number of laser irradiation to 5, 10, 20, and 100 shots. The ML prediction was carried out in two steps. In STEP1, we classified the contact status into three types.: 1) Schottky junctions (insufficient doping), 2) Ohmic contact (good contact), and 3) Not ohmic (damage caused by laser irradiation). In STEP2, contact resistance prediction (numerical regression) was performed using the dataset predicted as an ohmic contact. As a result, we found that the three classifications in STEP1 could be predicted with a high accuracy of over 88%. Furthermore, the contact resistance prediction in STEP2 could be made with an accuracy (RMSPE: root mean square percent error) of 27.2%. Visualizing the prediction basis of numerical regression using modulus-reweighted grad-regression activation mapping (MoRAM) revealed that the ML model focused on the inside of the laser-irradiated area in the optical microscope image. The results of the scanning electron microscopy observation of the laser-irradiated area showed that ablation and residuals were generated during laser doping in that area. Consequently, it was concluded that our ML model predicted the contact resistance of the finished device taking into consideration these surface conditions. Even highly-skilled laser doping technicians have difficulty predicting the resistance values arising from the ablation and residue conditions. Based on above results, we conclude that our ML model is capable of predicting the electrical characteristics of a finished device, a task that is often considered challenging for humans.

Accurate and visualiable discrimination of Chenpi age using 2D-CNN and Grad-CAM++ based on infrared spectral images

Dried tangerine peel (“Chenpi”), has numerous clinical and nutritional benefits, with its quality being significantly influenced by its storage age, referred to as “Chen Jiu Zhe Liang” in Chinese. Concequently, the rapid and accurate identification of Chenpi's age is important for consumers. In this study, Fourier transform infrared spectroscopy (FTIR) was employed to capture spectral images of Chenpi. These FTIR images were then analyzed using a two-dimensional convolutional neural networks (2D-CNN) model, achieving a discrimination accuracy of 97.92%. To address the “black box” nature of the 2D-CNN, Gradient-weighted Class Activation Mapping Plus Plus (Grad-CAM++) was utilized to highlight the important regions contributing to the model's performance. Additionally, six other machine learning models were developped using features identified by the 2D-CNN to validate their effectiveness. The results demonstrated that the combination of FTIR spectral images and 2D-CNN provides a highly effective method for accurately determining the age of Chenpi.

Distributed fiber optic sensing signal recognition based on class-incremental learning

Distributed fiber optic sensing (DFOS) based on phase-sensitive optical time-domain reflectance (φ-OTDR) technology has outstanding performance in pipeline safety monitoring and perimeter security detection. Accurate identification of new events remains challenging due to environmental variability and emerging forms of intrusions. In order to solve the problem of failing to accurately identify new events due to the inability to obtain all samples at once in real-time monitoring, this paper proposes an incremental learning network framework for distributed fiber-optic sensing signal recognition. This framework integrates an optimized Learning without Memorizing (LwM) algorithm with an improved ConvNeXt network for dynamic training of new events. An improved Efficient Channel Attention (HECA) is incorporated to thoroughly extract the spatio-temporal features of the intrusion signals collected by the DFOS. The forgetting problem is mitigated during incremental learning using knowledge distillation and optimized Gradient Weighted Class Activation Mapping to generate an attention map. A linear correction layer is added after the output layer to correct the bias towards new classes by rebalancing the information between new and old classes. Experimental comparisons show that the recognition rate for 10 different intrusion signals exceeds 93 %, while the forgetting rate is reduced from a peak of 41.44 % to 5.25 %. The time required to process and train incremental learning for 1000 samples in real time on an edge device (NVIDIA 3050 GPU) is approximately 1060 s, its ability to demonstrating its suitability for deployment in resource-constrained.

Estimation of the Global Equivalence Ratio of a Swirl Combustor from a Small Number of Absorption Spectra Using Machine Learning.

A new optical diagnostic method that predicts the global fuel-air equivalence ratio of a swirl combustor using absorption spectra from only three optical paths is proposed here. Under normal operation, the global equivalence ratio and total flow rate determine the temperature and concentration fields of the combustor, which subsequently determine the absorption spectra of any combustion species. Therefore, spectra, as the fingerprint for a produced combustion field, were employed to predict the global equivalence ratio, one of the key operational parameters, in this study. Specifically, absorption spectra of water vapor at wavenumbers around 7444.36, 7185.6, and 6805.6 cm-1 measured at three different downstream locations of the combustor were used to predict the global equivalence ratio. As it is difficult to find analytical relationships between the spectra and produced combustion fields, a predictive model was a data-driven acquisition. The absorption spectra as an input were first feature-extracted through stacked convolutional autoencoders and then a dense neural network was used for regression prediction between the feature scores and the global equivalence ratio. The model could predict the equivalence ratio with an absolute error of ±0.025 with a probability of 96%, and a gradient-weighted regression activation mapping analysis revealed that the model leverages not only the peak intensities but also the variations in the shape of absorption lines for its predictions.

Automatic Classification of Nodules from 2D Ultrasound Images Using Deep Learning Networks.

In clinical practice, thyroid nodules are typically visually evaluated by expert physicians using 2D ultrasound images. Based on their assessment, a fine needle aspiration (FNA) may be recommended. However, visually classifying thyroid nodules from ultrasound images may lead to unnecessary fine needle aspirations for patients. The aim of this study is to develop an automatic thyroid ultrasound image classification system to prevent unnecessary FNAs. An automatic computer-aided artificial intelligence system is proposed for classifying thyroid nodules using a fine-tuned deep learning model based on the DenseNet architecture, which incorporates an attention module. The dataset comprises 591 thyroid nodule images categorized based on the Bethesda score. Thyroid nodules are classified as either requiring FNA or not. The challenges encountered in this task include managing variability in image quality, addressing the presence of artifacts in ultrasound image datasets, tackling class imbalance, and ensuring model interpretability. We employed techniques such as data augmentation, class weighting, and gradient-weighted class activation maps (Grad-CAM) to enhance model performance and provide insights into decision making. Our approach achieved excellent results with an average accuracy of 0.94, F1-score of 0.93, and sensitivity of 0.96. The use of Grad-CAM gives insights on the decision making and then reinforce the reliability of the binary classification for the end-user perspective. We propose a deep learning architecture that effectively classifies thyroid nodules as requiring FNA or not from ultrasound images. Despite challenges related to image variability, class imbalance, and interpretability, our method demonstrated a high classification accuracy with minimal false negatives, showing its potential to reduce unnecessary FNAs in clinical settings.

Sensing noise exposure and its inequality based on noise complaint data through vision-language hybrid method

This study seeks to reveal urban noise exposure patterns and inequalities using noise complaint data and vision-language hybrid method. By applying a natural language processing model to 17,243 noise complaint records, we uncovered distinct patterns of traffic, industrial, and living noise exposures across residential communities. Our analysis of street view images near complaint locations, utilizing a Residual Network (ResNet) model and Class Activation Mapping (CAM), identified the key environmental elements of different noise sources. Notably, our assessment of noise exposure inequality across 9791 communities yielded a counterintuitive finding: contrary to previous studies in Western contexts, rich communities in China experience higher and more unequal noise exposure compared to average communities, with Gini coefficients exceeding 0.8. This unexpected result likely stems from China's unique rapid urbanization process. Our use of crowdsourced complaint data aligns more closely with human subjective perceptions of noise, offering a novel perspective on noise exposure inequality. These findings challenge existing assumptions about the relationship between socioeconomic status and environmental quality in urban China, and have significant implications for urban planning and noise management strategies in rapidly developing cities.

Improving flood forecast accuracy based on explainable convolutional neural network by Grad-CAM method

The advent of deep learning techniques has shown promising improvement in flood forecasting accuracy which are crucial for operating reservoir and mitigating flood damages. However, the practical application of typical deep learning models, such as the Long Short-Term Memory network (LSTM), is hindered by their high complexity, computational cost, and lack of interpretability. This study introduces an Explainable Convolutional Neural Network (ECNN) model that can effectively capture inter-variable relationships and temporal dynamics to address these challenges. The proposed ECNN model was applied to the Lushui basin in China and compared with benchmark models. The interpretability of the ECNN model was analyzed using the Gradient-weighted Class Activation Mapping (Grad-CAM) method. Our findings demonstrated that the ECNN model exhibits exceptional performance in flood forecast tasks, particularly for long-term forecast horizons. Furthermore, precipitation and inflow play a decisive role in flood forecasting with the watershed-specific flood characteristics. The ECNN model is highly promising in real-time flood forecasting with ingenious parallel structure, remarkable accuracy, and interpretability.

Artificial intelligence-empowered assessment of bile duct stone removal challenges

The aim of this investigation was to unravel the complexities inherent in endoscopic retrograde cholangiopancreatography (ERCP) procedures for bile duct stone removal by leveraging advanced artificial intelligence (AI) methodologies to support clinical decision-making and optimize patient outcomes. We introduced the Assessment of Stone Removal Challenges (ACRS) system, a novel integration of Data-efficient Image Transformers (DeiT) for image data analysis and eXtreme Gradient Boosting (XGBoost) for clinical data interpretation. Our study included a patient cohort of 2,129 individuals, focusing on training the ACRS system to achieve high diagnostic precision. Using logistic regression, we identified pivotal predictors affecting the complexity of bile duct stone removal. These findings were visually represented through forest plots and nomograms. Analytical and visualization processes were conducted using the Python and R programming languages, adhering to a p value significance threshold of less than 0.05. By utilizing DeiT enhanced by transfer learning and XGBoost for clinical data interpretation, the system achieved an accuracy of 0.83 and a perfect recall rate on a test set of 2,129 patients. Gradient-weighted class activation mapping (Grad-CAM) and SHapley Additive exPlanations (SHAP) provided in-depth diagnostic insights. Logistic regression was applied to identify crucial clinical predictors for stone removal difficulty, as visualized through forest plots and nomograms. These tools facilitated measurable assessments of procedural complexity, making significant strides in gastroenterology diagnostics and decision-making. By adopting causal inference techniques, the system effectively quantifies the influence of various clinical entities on stone removal difficulty, thereby augmenting both diagnostic precision and procedural strategies in ERCP.

Three-Dimensional Electrodes of Liquid Metals for Long-Term, Wireless Cardiac Analysis and Modulation.

Cardiovascular disease is a major public health issue, and smart diagnostic approaches play an important role in the analysis of electrocardiograms. Here, we present three-dimensional, soft electrodes of liquid metals that can be conformably attached to the surfaces of the heart and skin for long-term cardiac analysis. The fine micropillar structures of biocompatible liquid metals enable precise targeting to small tissue areas, allowing for spatiotemporal mapping and modulation of cardiac electrical activity with high resolution. The low mechanical modulus of these liquid-metal electrodes not only helps avoid inflammatory responses triggered by modulus mismatch between the tissue and electrodes, but also minimizes pain when embedded in biological tissues such as the skin and heart. Furthermore, in vivo experiments with animal models and human trials demonstrate long-term and accurate monitoring of electrocardiograms over a period of 30 days.

Whole-brain mapping in adult zebrafish and identification of a novel tank test functional connectome.

Identifying general principles of brain function requires the study of structure-function relationships in a variety of species. Zebrafish have recently gained prominence as a model organism in neuroscience, yielding important insights into vertebrate brain function. Although methods have been developed for mapping neural activity in larval animals, we lack similar techniques for adult zebrafish that have the advantage of a fully developed neuroanatomy and larger behavioral repertoire. Here, we describe a pipeline built around open-source tools for whole-brain activity mapping in freely swimming adult zebrafish. Our pipeline combines recent advances in histology, microscopy, and machine learning to capture cfos activity across the entirety of the adult brain. Images captured using light-sheet microscopy are registered to the recently created adult zebrafish brain atlas (AZBA) for automated segmentation using advanced normalization tools (ANTs). We used our pipeline to measure brain activity after zebrafish were subject to the novel tank test. We found that cfos levels peaked 15 minutes following behavior and that several regions containing serotoninergic, dopaminergic, noradrenergic, and cholinergic neurons were active during exploration. Finally, we generated a novel tank test functional connectome. Functional network analysis revealed that several regions of the medial ventral telencephalon form a cohesive sub-network during exploration. We also found that the anterior portion of the parvocellular preoptic nucleus (PPa) serves as a key connection between the ventral telencephalon and many other parts of the brain. Taken together, our work enables whole-brain activity mapping in adult zebrafish for the first time while providing insight into neural basis for the novel tank test.

Interpretability in Video-based Human Action Recognition: Saliency Maps and GradCAM in 3D Convolutional Neural Networks

Interpretability plays a vital role in understanding complex deep learning models by providing transparency and insights. It addresses the black-box nature of these models, aids in human-in-the-loop systems, enhances model development, and supports education. However, existing interpretability algorithms in computer vision primarily target images, leaving a gap for video applications. In this article, we emphasize the importance of interpretability in video-based Human Action Recognition (HAR). We extend existing 2D interpretability techniques to the 3D domain, specifically focusing on saliency maps and gradient-weighted class activation maps. The proposed interpretability system is then employed in the analysis of a well-known HAR dataset to better understand action recognition in videos.

Heartbeat-related spectral perturbation of electroencephalogram reflects dynamic interoceptive attention states in the trial-by-trial classification analysis

Attending to heartbeats for interoceptive awareness initiates distinct electrophysiological responses synchronized with the R-peaks of an electrocardiogram (ECG), such as the heartbeat-evoked potential (HEP). Beyond HEP, this study proposes heartbeat-related spectral perturbation (HRSP), a time–frequency map of the R-peak locked electroencephalogram (EEG), and explores its characteristics in identifying interoceptive attention states using a classification approach. HRSPs of EEG brain components specified by independent component analysis (ICA) were used for the offline and online classification of interoceptive states. A convolutional neural network (CNN) designed specifically for HRSP was applied to publicly available data from a binary-state experiment (attending to self-heartbeats and white noise) and data from our four-state classification experiment (attending to self-heartbeats, white noise, time passage, and toe) with diverse input feature conditions of HRSP. From the dynamic state perspective, we evaluated the primary frequency bands of HRSP and the minimal number of averaging epochs required to reflect changing interoceptive attention states without compromising accuracy. We also assessed the utility of group ICA and models for classifying HRSP in new participants. The CNN for trial-by-trial HRSP with actual R-peaks demonstrated significantly higher classification accuracy than HRSP with sham, i.e., randomly positioned, R-peaks. Gradient-weighted class activation mapping highlighted the prominent role of theta and alpha bands between 200–600 ms post-R-peak—features absent in classifications using sham HRSPs. Online classification benefits from employing a group ICA and classification model, ensuring reliable accuracy without individual EEG precollection. These results suggest HRSP’s potential to reflect interoceptive attention states, proposing transformative implications for clinical applications.