Pub Date : 2025-04-07DOI: 10.1016/j.bspc.2025.107856
Himashree Kalita, Samarendra Dandapat, Prabin Kumar Bora
Age-related macular degeneration (AMD) is a retinal disease that can impair the central vision permanently. Accurate delineation of the retinal pigment epithelium (RPE) and Bruch’s membrane (BM) in optical coherence tomography (OCT) B-scans is crucial for diagnosing and monitoring AMD. While automated segmentation methods exist for early AMD stages, late-stage AMD remains a challenging area due to the pronounced disruption of the RPE and BM. To ensure spatial contiguity in the boundary delineation of RPE and BM, both the global and local contextual information must be learned. In this context, we propose a generative adversarial network (GAN) to segment these significant retinal interfaces in OCT B-scans from AMD patients. A UNet++ model with its deep supervision is trained using a hybrid loss function combining adversarial loss and multi-class cross-entropy (CE) segmentation loss. The CE loss learns the local features by optimizing the per-pixel accuracy, while the adversarial loss captures a broader context by learning overall layer label statistics. This loss combination allows the model to capture fine details in the ordered retinal layer structure and guide layer boundaries along discontinuities in the RPE and BM in severe AMD cases. Additionally, a graph search algorithm refines boundary delineations from predicted segmentation maps. The model’s effectiveness is validated on the DUEIA and AROI datasets, which include OCT B-scans from both AMD-affected and healthy individuals. The proposed approach achieves Mean Absolute Errors (MAE) of 0.45 and 1.19 on the respective datasets, demonstrating its capability to handle boundary segmentation in severe AMD cases.
{"title":"A generative adversarial network for delineation of retinal interfaces in OCT B-scans with age-related macular degeneration","authors":"Himashree Kalita, Samarendra Dandapat, Prabin Kumar Bora","doi":"10.1016/j.bspc.2025.107856","DOIUrl":"10.1016/j.bspc.2025.107856","url":null,"abstract":"<div><div>Age-related macular degeneration (AMD) is a retinal disease that can impair the central vision permanently. Accurate delineation of the retinal pigment epithelium (RPE) and Bruch’s membrane (BM) in optical coherence tomography (OCT) B-scans is crucial for diagnosing and monitoring AMD. While automated segmentation methods exist for early AMD stages, late-stage AMD remains a challenging area due to the pronounced disruption of the RPE and BM. To ensure spatial contiguity in the boundary delineation of RPE and BM, both the global and local contextual information must be learned. In this context, we propose a generative adversarial network (GAN) to segment these significant retinal interfaces in OCT B-scans from AMD patients. A UNet++ model with its deep supervision is trained using a hybrid loss function combining adversarial loss and multi-class cross-entropy (CE) segmentation loss. The CE loss learns the local features by optimizing the per-pixel accuracy, while the adversarial loss captures a broader context by learning overall layer label statistics. This loss combination allows the model to capture fine details in the ordered retinal layer structure and guide layer boundaries along discontinuities in the RPE and BM in severe AMD cases. Additionally, a graph search algorithm refines boundary delineations from predicted segmentation maps. The model’s effectiveness is validated on the DUEIA and AROI datasets, which include OCT B-scans from both AMD-affected and healthy individuals. The proposed approach achieves Mean Absolute Errors (MAE) of 0.45 and 1.19 on the respective datasets, demonstrating its capability to handle boundary segmentation in severe AMD cases.</div></div>","PeriodicalId":55362,"journal":{"name":"Biomedical Signal Processing and Control","volume":"108 ","pages":"Article 107856"},"PeriodicalIF":4.9,"publicationDate":"2025-04-07","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"143785643","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"医学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2025-04-07DOI: 10.1016/j.bspc.2025.107828
Mucong Zhuang , Yulin Li , Liying Hu , Zhiling Hong , Lifei Chen
Convolutional neural networks with U-shaped architectures are widely used in medical image segmentation. However, their performance is often limited by imbalanced regional attention caused by interference from irrelevant features within localized receptive fields. To overcome this limitation, FDU-Net is proposed as a novel U-Net-based model that incorporates a feature decorrelation strategy. Specifically, FDU-Net introduces a feature decorrelation method that extracts multiple groups of features from the encoder and optimizes sample weights to reduce internal feature correlations, thereby minimizing the interference from irrelevant features. Comprehensive experiments on diverse medical imaging datasets show that FDU-Net achieves superior evaluation scores and finer segmentation results, outperforming state-of-the-art methods.
{"title":"Narrowing the regional attention imbalance in medical image segmentation via feature decorrelation","authors":"Mucong Zhuang , Yulin Li , Liying Hu , Zhiling Hong , Lifei Chen","doi":"10.1016/j.bspc.2025.107828","DOIUrl":"10.1016/j.bspc.2025.107828","url":null,"abstract":"<div><div>Convolutional neural networks with U-shaped architectures are widely used in medical image segmentation. However, their performance is often limited by imbalanced regional attention caused by interference from irrelevant features within localized receptive fields. To overcome this limitation, FDU-Net is proposed as a novel U-Net-based model that incorporates a feature decorrelation strategy. Specifically, FDU-Net introduces a feature decorrelation method that extracts multiple groups of features from the encoder and optimizes sample weights to reduce internal feature correlations, thereby minimizing the interference from irrelevant features. Comprehensive experiments on diverse medical imaging datasets show that FDU-Net achieves superior evaluation scores and finer segmentation results, outperforming state-of-the-art methods.</div></div>","PeriodicalId":55362,"journal":{"name":"Biomedical Signal Processing and Control","volume":"108 ","pages":"Article 107828"},"PeriodicalIF":4.9,"publicationDate":"2025-04-07","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"143785642","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"医学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2025-04-06DOI: 10.1016/j.bspc.2025.107824
Mrinalini Bhagawati , Siddharth Gupta , Sudip Paul , Laura Mantella , Amer M. Johri , John R. Laird , Ekta Tiwari , Narendra N. Khanna , Andrew Nicolaides , Rajesh Singh , Mustafa Al-Maini , Luca Saba , Jasjit S. Suri
Background
Carotid plaque can be used to predict the risk of cardiovascular disease (CVD). Earlier machine learning solutions were not reliable, or accurate. The authors hypothesize that (i) attention-based unidirectional or bidirectional hybrid deep learning (HDL) is superior to non-attention-based unidirectional or bidirectional hybrid deep learning and (ii) attention-based bidirectional hybrid deep learning models are superior to attention-based unidirectional HDL paradigms. The proposed design, AtheroEdge™ 3.0att-HDL (AtheroPoint™, Roseville, CA, USA), shows how effectively characteristics of the carotid plaque in attention-based hybrid deep learning systems predict the risk of CVD more accurately and reliably.
Methodology
The study involved 500 participants who underwent targeted carotid B-mode ultrasonography along with coronary angiography. Six hybrid models (four attention types) were used, totaling 6x4 = 24 models. These were benchmarked against the machine learning models. Mann-Whitney U test, Wilcoxon test, and paired T-test were used for the statistical and reliability tests. The scientific validation was performed using the unseen data. The area-under-the-curve and p-values were used for the performance evaluation of AtheroEdge™ 3.0att-HDL.
Results
The best attention-based bidirectional HDL model showed a mean improvement of 36.11 %, 5.37 %, 5.37 %, and 1.04 % over Random Forest, unidirectional LSTM, bidirectional LSTM, and best attention-based unidirectional HDL models, respectively. As per the reliability and statistical test findings, the bidirectional AtheroEdge™ 3.0att-HDL had a p-value of less than 0.001, while the unidirectional AtheroEdge™ 3.0att-HDL also complied with regulations having a p-value < 0.005.
Conclusions
The hypothesis was scientifically validated, assessed for reliability and stability, and deemed suitable for clinical application.
{"title":"Attention-based hybrid deep learning models and its scientific validation for cardiovascular disease risk stratification","authors":"Mrinalini Bhagawati , Siddharth Gupta , Sudip Paul , Laura Mantella , Amer M. Johri , John R. Laird , Ekta Tiwari , Narendra N. Khanna , Andrew Nicolaides , Rajesh Singh , Mustafa Al-Maini , Luca Saba , Jasjit S. Suri","doi":"10.1016/j.bspc.2025.107824","DOIUrl":"10.1016/j.bspc.2025.107824","url":null,"abstract":"<div><h3>Background</h3><div>Carotid plaque can be used to predict the risk of cardiovascular disease (CVD). Earlier machine learning solutions were not reliable, or accurate. The authors hypothesize that (i) attention-based unidirectional or bidirectional hybrid deep learning (HDL) is superior to non-attention-based unidirectional or bidirectional hybrid deep learning and (ii) attention-based bidirectional hybrid deep learning models are superior to attention-based unidirectional HDL paradigms. The proposed design, AtheroEdge™ 3.0<sub>att-HDL</sub> (AtheroPoint™, Roseville, CA, USA), shows how effectively characteristics of the carotid plaque in attention-based hybrid deep learning systems predict the risk of CVD more accurately and reliably.</div></div><div><h3>Methodology</h3><div>The study involved 500 participants who underwent targeted carotid B-mode ultrasonography along with coronary angiography. Six hybrid models (four attention types) were used, totaling 6x4 = 24 models. These were benchmarked against the machine learning models. Mann-Whitney <em>U</em> test, Wilcoxon test, and paired <em>T</em>-test were used for the statistical and reliability tests. The scientific validation was performed using the unseen data. The area-under-the-curve and p-values were used for the performance evaluation of AtheroEdge™ 3.0<sub>att-HDL</sub>.</div></div><div><h3>Results</h3><div>The best attention-based bidirectional HDL model showed a mean improvement of <strong>36.11 %</strong>, <strong>5.37 %</strong>, <strong>5.37 %</strong>, and <strong>1.04 %</strong> over Random Forest, unidirectional LSTM, bidirectional LSTM, and best attention-based unidirectional HDL models, respectively. As per the reliability and statistical test findings, the bidirectional AtheroEdge™ 3.0<sub>att-HDL</sub> had a p-value of less than 0.001, while the unidirectional AtheroEdge™ 3.0<sub>att-HDL</sub> also complied with regulations having a p-value < 0.005.</div></div><div><h3>Conclusions</h3><div>The hypothesis was scientifically validated, assessed for reliability and stability, and deemed suitable for clinical application.</div></div>","PeriodicalId":55362,"journal":{"name":"Biomedical Signal Processing and Control","volume":"108 ","pages":"Article 107824"},"PeriodicalIF":4.9,"publicationDate":"2025-04-06","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"143783227","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"医学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2025-04-04DOI: 10.1016/j.bspc.2025.107848
Amal Kammoun , Philippe Ravier , Olivier Buttelli
In the context of low-cost and portable device for measuring pressure using insole system, selection of the relevant sensors is addressed. In a preliminary step, we compared the accuracy of Ground Reaction Force (GRF) components estimation between two pressure insoles: Fscan and Moticon. This estimation was done using Artificial Neural Network combined with Principal Component Analysis (PCA). Secondly, the focus of this study was to identify the optimal numbers and locations of the pressure sensors by a sensor ranking procedure for both insoles using PCA and three selection strategies. The ranking is determined by analyzing the loss value obtained through PCA for each pressure sensor with three selection strategies. Using data from gold standard force plates, we assessed GRF components estimation accuracies and sensor locations for both insoles during walking activities. As a first result, in our context, Moticon insole yielded superior performance for estimating GRF components compared to Fscan. Secondly, the selection procedure allowed deleting 3 among 16 sensors for Moticon (both feet) and 33/30 (left foot/right foot) among 64 sensors for Fscan. Finally, we have validated these optimal numbers by showing that the quality of GRF components estimation was minimally impacted. Remarkably, both insoles with fewer sensors led to better vertical component estimations. These results should be considered in the context of this study, which does not claim to be generalizable. As these results do not reflect a wide range of activities and subject profiles, it is therefore necessary to re-evaluate these selections with other activity conditions.
{"title":"Selection of insole pressure sensors for ground reaction force estimation through studying principal component analysis","authors":"Amal Kammoun , Philippe Ravier , Olivier Buttelli","doi":"10.1016/j.bspc.2025.107848","DOIUrl":"10.1016/j.bspc.2025.107848","url":null,"abstract":"<div><div>In the context of low-cost and portable device for measuring pressure using insole system, selection of the relevant sensors is addressed. In a preliminary step, we compared the accuracy of Ground Reaction Force (GRF) components estimation between two pressure insoles: Fscan and Moticon. This estimation was done using Artificial Neural Network combined with Principal Component Analysis (PCA). Secondly, the focus of this study was to identify the optimal numbers and locations of the pressure sensors by a sensor ranking procedure for both insoles using PCA and three selection strategies. The ranking is determined by analyzing the loss value obtained through PCA for each pressure sensor with three selection strategies. Using data from gold standard force plates, we assessed GRF components estimation accuracies and sensor locations for both insoles during walking activities. As a first result, in our context, Moticon insole yielded superior performance for estimating GRF components compared to Fscan. Secondly, the selection procedure allowed deleting 3 among 16 sensors for Moticon (both feet) and 33/30 (left foot/right foot) among 64 sensors for Fscan. Finally, we have validated these optimal numbers by showing that the quality of GRF components estimation was minimally impacted. Remarkably, both insoles with fewer sensors led to better vertical component estimations. These results should be considered in the context of this study, which does not claim to be generalizable. As these results do not reflect a wide range of activities and subject profiles, it is therefore necessary to re-evaluate these selections with other activity conditions.</div></div>","PeriodicalId":55362,"journal":{"name":"Biomedical Signal Processing and Control","volume":"107 ","pages":"Article 107848"},"PeriodicalIF":4.9,"publicationDate":"2025-04-04","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"143768532","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"医学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"OA","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2025-04-04DOI: 10.1016/j.bspc.2025.107874
Tianjiao Zhang , Yanfeng Wang , Weidi Xie , Ya Zhang
In this paper, we consider the problem of semi-automatic medical image segmentation, with the goal of segmenting the target structure in a whole 3-D volume image with only a single slice annotation to relieve the user’s annotation burden. Under such a paradigm, the segmentation of the volume is achieved by establishing the correspondence between slices and propagating the reference segmentation. We propose a more medical-suited framework denoted Slice Segmentation Propagator (SSP) that can establish reliable correspondences between slices with local attention, and maintain a running memory bank that effectively mitigates the problem of error accumulation during mask propagation. Additionally, we propose two test-time training strategies to further improve the propagation performance and generalization ability of the framework, namely, a cycle consistency mechanism to suppress error propagation, and an online adaption procedure via artificial augmentation, assisting the model to better generalize towards new structures at test time. We have conducted thorough experiments on three datasets on four anatomy structures, demonstrating promising results on both in-structure and cross-structure (test on different structures from trainset) scenarios.
{"title":"Slice Segmentation Propagator: Propagating the single slice annotation to 3D volume","authors":"Tianjiao Zhang , Yanfeng Wang , Weidi Xie , Ya Zhang","doi":"10.1016/j.bspc.2025.107874","DOIUrl":"10.1016/j.bspc.2025.107874","url":null,"abstract":"<div><div>In this paper, we consider the problem of semi-automatic medical image segmentation, with the goal of segmenting the target structure in a whole 3-D volume image with only a single slice annotation to relieve the user’s annotation burden. Under such a paradigm, the segmentation of the volume is achieved by establishing the correspondence between slices and propagating the reference segmentation. We propose a more medical-suited framework denoted Slice Segmentation Propagator (SSP) that can establish reliable correspondences between slices with local attention, and maintain a running memory bank that effectively mitigates the problem of error accumulation during mask propagation. Additionally, we propose two test-time training strategies to further improve the propagation performance and generalization ability of the framework, namely, a cycle consistency mechanism to suppress error propagation, and an online adaption procedure via artificial augmentation, assisting the model to better generalize towards new structures at test time. We have conducted thorough experiments on three datasets on four anatomy structures, demonstrating promising results on both in-structure and cross-structure (test on different structures from trainset) scenarios.</div></div>","PeriodicalId":55362,"journal":{"name":"Biomedical Signal Processing and Control","volume":"107 ","pages":"Article 107874"},"PeriodicalIF":4.9,"publicationDate":"2025-04-04","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"143768531","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"医学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2025-04-03DOI: 10.1016/j.bspc.2025.107849
Mengjie Xu , Zihao Zhao , Lanzhuju Mei , Sheng Wang , Xiaoxi Lin , Shih-Jen Chang , Qian Wang , Yajing Qiu , Dinggang Shen
Infantile hemangiomas (IH) are a common pediatric condition that, if not diagnosed and treated early, can lead to functional impairments or permanent disfigurement. However, accurate diagnosis and timely treatment recommendations often depend on the expertise of clinicians and expensive medical imaging, which presents significant challenges in resource-limited settings, especially in low- and middle-income countries. While existing computer-aided diagnosis (CAD) methods have been developed for IH, they mainly assist clinicians rather than offering direct decision-making support, which limits their impact on patient care. To address these challenges, we propose DeepIH, the first near-patient system designed for treatment recommendation of IH based on deep learning. DeepIH is methodologically innovative in two key ways: (1) it accepts camera-shot images as input, enabling patients to conveniently access treatment recommendations through accessible edge devices like smartphones or laptops; (2) it directly generates treatment recommendations, reducing reliance on clinician oversight and enabling faster, more accessible care. Through evaluation on our established dataset, DeepIH achieves an impressive 95.8% accuracy in detecting lesion regions and 84.9% top-3 accuracy in recommending treatments, which even surpasses a fine-tuned foundation model by 1.7%. These findings, for the first time, validate the viability of near-patient diagnosis for IH, highlighting its potential significance in clinical applications as it allows patients to receive treatment recommendations through everyday devices like smartphones or laptops.
{"title":"DeepIH: A deep learning-based near-patient system for treatment recommendation in infantile hemangiomas","authors":"Mengjie Xu , Zihao Zhao , Lanzhuju Mei , Sheng Wang , Xiaoxi Lin , Shih-Jen Chang , Qian Wang , Yajing Qiu , Dinggang Shen","doi":"10.1016/j.bspc.2025.107849","DOIUrl":"10.1016/j.bspc.2025.107849","url":null,"abstract":"<div><div>Infantile hemangiomas (IH) are a common pediatric condition that, if not diagnosed and treated early, can lead to functional impairments or permanent disfigurement. However, accurate diagnosis and timely treatment recommendations often depend on the expertise of clinicians and expensive medical imaging, which presents significant challenges in resource-limited settings, especially in low- and middle-income countries. While existing computer-aided diagnosis (CAD) methods have been developed for IH, they mainly assist clinicians rather than offering direct decision-making support, which limits their impact on patient care. To address these challenges, we propose DeepIH, the first near-patient system designed for treatment recommendation of IH based on deep learning. DeepIH is methodologically innovative in two key ways: (1) it accepts camera-shot images as input, enabling patients to conveniently access treatment recommendations through accessible edge devices like smartphones or laptops; (2) it directly generates treatment recommendations, reducing reliance on clinician oversight and enabling faster, more accessible care. Through evaluation on our established dataset, DeepIH achieves an impressive 95.8% accuracy in detecting lesion regions and 84.9% top-3 accuracy in recommending treatments, which even surpasses a fine-tuned foundation model by 1.7%. These findings, for the first time, validate the viability of near-patient diagnosis for IH, highlighting its potential significance in clinical applications as it allows patients to receive treatment recommendations through everyday devices like smartphones or laptops.</div></div>","PeriodicalId":55362,"journal":{"name":"Biomedical Signal Processing and Control","volume":"108 ","pages":"Article 107849"},"PeriodicalIF":4.9,"publicationDate":"2025-04-03","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"143760528","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"医学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2025-04-02DOI: 10.1016/j.bspc.2025.107837
Guangyu Yang , Dafeng Long , Kai Wang , Shuyan Xia , Juncheng Zou
Prediction of epileptic seizures is crucial for timely intervention and control. A significant challenge in this domain is the scarcity of preictal EEG data, which is important for accurate seizure prediction. To address this issue, a novel data augmentation method based on a transition network is proposed which not only enhances dataset diversity through a random walk algorithm but also preserves the spatial correlation between different EEG channels. Additionally, a noise-robust multivariate weighted fuzzy granular recurrence plot is introduced to extract nonlinear characteristics from EEG data, effectively mitigating the impact of noise on signal analysis. The multivariate weighted fuzzy granular recurrence plots are then input into the Inception V3 model for training the epilepsy prediction model. The novel method achieves state-of-the-art performance on the CHB-MIT database and American Epilepsy Society-Kaggle dataset. The key novelty of this work lies in the proposal of a transition network data augmentation method which overcomes the limitations of existing data augmentation techniques that often ignore inter-channel correlations or distort data distributions. Moreover, the introduction and development of fuzzy granular recurrence plot overcome the noise susceptibility of existing recurrence-plot-based EEG signal analysis methods and improves the extraction of detailed nonlinear features. By integrating these two novel methods into a unified framework, the performances of EEG data analysis and seizure prediction are effectively improved, offering a robust solution for clinical applications.
{"title":"Epileptic seizure prediction method based on transition network data augmentation and fuzzy granular recurrence plot","authors":"Guangyu Yang , Dafeng Long , Kai Wang , Shuyan Xia , Juncheng Zou","doi":"10.1016/j.bspc.2025.107837","DOIUrl":"10.1016/j.bspc.2025.107837","url":null,"abstract":"<div><div>Prediction of epileptic seizures is crucial for timely intervention and control. A significant challenge in this domain is the scarcity of preictal EEG data, which is important for accurate seizure prediction. To address this issue, a novel data augmentation method based on a transition network is proposed which not only enhances dataset diversity through a random walk algorithm but also preserves the spatial correlation between different EEG channels. Additionally, a noise-robust multivariate weighted fuzzy granular recurrence plot is introduced to extract nonlinear characteristics from EEG data, effectively mitigating the impact of noise on signal analysis. The multivariate weighted fuzzy granular recurrence plots are then input into the Inception V3 model for training the epilepsy prediction model. The novel method achieves state-of-the-art performance on the CHB-MIT database and American Epilepsy Society-Kaggle dataset. The key novelty of this work lies in the proposal of a transition network data augmentation method which overcomes the limitations of existing data augmentation techniques that often ignore inter-channel correlations or distort data distributions. Moreover, the introduction and development of fuzzy granular recurrence plot overcome the noise susceptibility of existing recurrence-plot-based EEG signal analysis methods and improves the extraction of detailed nonlinear features. By integrating these two novel methods into a unified framework, the performances of EEG data analysis and seizure prediction are effectively improved, offering a robust solution for clinical applications.</div></div>","PeriodicalId":55362,"journal":{"name":"Biomedical Signal Processing and Control","volume":"107 ","pages":"Article 107837"},"PeriodicalIF":4.9,"publicationDate":"2025-04-02","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"143760688","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"医学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2025-04-02DOI: 10.1016/j.bspc.2025.107875
Mira Haapatikka , Mikko Peltokangas , Saara Pietilä , Sara Protto , Velipekka Suominen , Ilkka Uurto , Damir Vakhitov , Essi Väisänen , Karem Lozano Montero , Mika-Matti Laurila , Jarmo Verho , Matti Mäntysalo , Niku Oksala , Antti Vehkaoja
Currently, most of abdominal aortic aneurysms (AAA) are detected by accident on imaging investigations of other medical conditions. The objective of this study was to investigate the classification of subjects with AAA patients and control subjects into two groups using features calculated directly from photoplethysmographic (PPG) signals measured from the index finger. PPG signals were analyzed from 48 test participants from which 25 had AAA and 23 were controls without AAA. Six pulse waveform features were computed from the PPG signals and sequential backward feature selection (SBFS) with linear discriminant analysis (LDA) and leave-one-participant-out cross validation was used to find the most relevant features. The actual classification was also done with LDA using features chosen by the SBFS. The dataset was divided to 70% training and 30% testing groups before classification. The split was stratified so that percentages of AAA subjects and controls was the same in test and train sets. Classification was repeated 500 times, and the median of the classification results was calculated. Three out of six pulse wave features were chosen for the classification. The LDA model had an area under curve (AUC) of 75%, an accuracy of 71%, a specificity of 68%, a sensitivity of 75%, score of 71%, and a positive predictive value (PPV) of 70%. Features calculated directly from PPG signals can separate individuals with AAA from controls with moderate accuracy. PPG waveform analysis could provide an easy-to-access method for AAA screening. Nonetheless, the performance should still be improved for guaranteeing clinical utility.
{"title":"Detection of abdominal aortic aneurysm using photoplethysmographic signals measured from the index finger","authors":"Mira Haapatikka , Mikko Peltokangas , Saara Pietilä , Sara Protto , Velipekka Suominen , Ilkka Uurto , Damir Vakhitov , Essi Väisänen , Karem Lozano Montero , Mika-Matti Laurila , Jarmo Verho , Matti Mäntysalo , Niku Oksala , Antti Vehkaoja","doi":"10.1016/j.bspc.2025.107875","DOIUrl":"10.1016/j.bspc.2025.107875","url":null,"abstract":"<div><div>Currently, most of abdominal aortic aneurysms (AAA) are detected by accident on imaging investigations of other medical conditions. The objective of this study was to investigate the classification of subjects with AAA patients and control subjects into two groups using features calculated directly from photoplethysmographic (PPG) signals measured from the index finger. PPG signals were analyzed from 48 test participants from which 25 had AAA and 23 were controls without AAA. Six pulse waveform features were computed from the PPG signals and sequential backward feature selection (SBFS) with linear discriminant analysis (LDA) and leave-one-participant-out cross validation was used to find the most relevant features. The actual classification was also done with LDA using features chosen by the SBFS. The dataset was divided to 70% training and 30% testing groups before classification. The split was stratified so that percentages of AAA subjects and controls was the same in test and train sets. Classification was repeated 500 times, and the median of the classification results was calculated. Three out of six pulse wave features were chosen for the classification. The LDA model had an area under curve (AUC) of 75%, an accuracy of 71%, a specificity of 68%, a sensitivity of 75%, <span><math><msub><mrow><mtext>F</mtext></mrow><mrow><mn>1</mn></mrow></msub></math></span> score of 71%, and a positive predictive value (PPV) of 70%. Features calculated directly from PPG signals can separate individuals with AAA from controls with moderate accuracy. PPG waveform analysis could provide an easy-to-access method for AAA screening. Nonetheless, the performance should still be improved for guaranteeing clinical utility.</div></div>","PeriodicalId":55362,"journal":{"name":"Biomedical Signal Processing and Control","volume":"107 ","pages":"Article 107875"},"PeriodicalIF":4.9,"publicationDate":"2025-04-02","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"143760687","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"医学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"OA","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2025-04-01DOI: 10.1016/j.bspc.2025.107818
Luca Quagliato , Sewon Kim , Olamide Robiat Hassan , Taeyong Lee
Gait analysis and heel pad mechanical properties have been largely studied by physicians and biomechanical engineers alike. However, only a few contributions deal with the intertwining relationship between these two essential aspects and no research seems to propose a modeling approach to quantitatively correlate them. To bridge this gap, indentation experiments on the heel pad and gait analysis through motion capture camera were carried out on a group composed of 40 male and female subjects in the 20′s to 50′s. To establish a robust correlation between these two sets of parameters, the Gaussian Mixture Model (GMM) features’ enhancement technique was employed and combined with the Extreme Gradient Boosting (XGB) regressor. The hyperelastic constants from models, together with the gait parameters, were employed as both features and target variables in the GMM-XGB architecture showing the ambivalence of the solution and deviations between 5% and 8% in most cases. The results show the strong reciprocal correlation between the individual’s foot plantar soft tissue’s mechanical response and the gait parameters and pave the way for further investigations in the field of biomechanics.
{"title":"Heel pad’s hyperelastic properties and gait parameters reciprocal modelling by a Gaussian Mixture Model and Extreme Gradient Boosting framework","authors":"Luca Quagliato , Sewon Kim , Olamide Robiat Hassan , Taeyong Lee","doi":"10.1016/j.bspc.2025.107818","DOIUrl":"10.1016/j.bspc.2025.107818","url":null,"abstract":"<div><div>Gait analysis and heel pad mechanical properties have been largely studied by physicians and biomechanical engineers alike. However, only a few contributions deal with the intertwining relationship between these two essential aspects and no research seems to propose a modeling approach to quantitatively correlate them. To bridge this gap, indentation experiments on the heel pad and gait analysis through motion capture camera were carried out on a group composed of 40 male and female subjects in the 20′s to 50′s. To establish a robust correlation between these two sets of parameters, the Gaussian Mixture Model (GMM) features’ enhancement technique was employed and combined with the Extreme Gradient Boosting (XGB) regressor. The hyperelastic constants from models, together with the gait parameters, were employed as both features and target variables in the GMM-XGB architecture showing the ambivalence of the solution and deviations between 5% and 8% in most cases. The results show the strong reciprocal correlation between the individual’s foot plantar soft tissue’s mechanical response and the gait parameters and pave the way for further investigations in the field of biomechanics.</div></div>","PeriodicalId":55362,"journal":{"name":"Biomedical Signal Processing and Control","volume":"107 ","pages":"Article 107818"},"PeriodicalIF":4.9,"publicationDate":"2025-04-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"143746665","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"医学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2025-04-01DOI: 10.1016/j.bspc.2025.107833
Hai Jiang , Yusuke Yamanoi , Peiji Chen , Xin Wang , Shixiong Chen , Xu Yong , Guanglin Li , Hiroshi Yokoi , Xiaobei Jing
Current pattern recognition-based myoelectric prosthetic hand control methods map electromyography (EMG) signals to specific hand postures, achieving high accuracy but often resulting in unnatural movements during transitions, reducing the hand’s anthropomorphic nature. While some studies predict single-finger joint angles from EMG signals, these approaches lack practicality since arm muscles often control multiple fingers simultaneously. This study proposed a TF2AngleNet that predicts six finger joint angles using both time domain raw signals and frequency domain features of EMG signals. A novel non-contact joint angle measurement method was used to collect EMG and joint angle data from five healthy subjects over five days. The experimental results demonstrate that TF2AngleNet achieves outstanding performance in continuous joint angle estimation, with a correlation coefficient of 94.7%, an R2 value of 89.2%, and an NRMSE of 9.5%. Notably, this represents a 12.43% improvement in NRMSE, along with average gains of 1.2% in CC and 2.42% in R2 compared to single-domain models (p-values 0.05 across all metrics). Also, hand postures were shown using a virtual hand model, providing a natural and bionic control method of myoelectric hands. Additionally, a novel conceptual framework is proposed to reduce barriers to using pattern recognition-based prosthetic hands, with this study serving as its first stage by validating the model’s performance under three experimental conditions. This research provides a promising solution for dexterous, biomimetic and practical myoelectric prosthetic hand control methods.
{"title":"TF2AngleNet: Continuous finger joint angle estimation based on multidimensional time–frequency features of sEMG signals","authors":"Hai Jiang , Yusuke Yamanoi , Peiji Chen , Xin Wang , Shixiong Chen , Xu Yong , Guanglin Li , Hiroshi Yokoi , Xiaobei Jing","doi":"10.1016/j.bspc.2025.107833","DOIUrl":"10.1016/j.bspc.2025.107833","url":null,"abstract":"<div><div>Current pattern recognition-based myoelectric prosthetic hand control methods map electromyography (EMG) signals to specific hand postures, achieving high accuracy but often resulting in unnatural movements during transitions, reducing the hand’s anthropomorphic nature. While some studies predict single-finger joint angles from EMG signals, these approaches lack practicality since arm muscles often control multiple fingers simultaneously. This study proposed a TF2AngleNet that predicts six finger joint angles using both time domain raw signals and frequency domain features of EMG signals. A novel non-contact joint angle measurement method was used to collect EMG and joint angle data from five healthy subjects over five days. The experimental results demonstrate that TF2AngleNet achieves outstanding performance in continuous joint angle estimation, with a correlation coefficient of 94.7%, an R<sup>2</sup> value of 89.2%, and an NRMSE of 9.5%. Notably, this represents a 12.43% improvement in NRMSE, along with average gains of 1.2% in CC and 2.42% in R<sup>2</sup> compared to single-domain models (p-values <span><math><mo><</mo></math></span> 0.05 across all metrics). Also, hand postures were shown using a virtual hand model, providing a natural and bionic control method of myoelectric hands. Additionally, a novel conceptual framework is proposed to reduce barriers to using pattern recognition-based prosthetic hands, with this study serving as its first stage by validating the model’s performance under three experimental conditions. This research provides a promising solution for dexterous, biomimetic and practical myoelectric prosthetic hand control methods.</div></div>","PeriodicalId":55362,"journal":{"name":"Biomedical Signal Processing and Control","volume":"107 ","pages":"Article 107833"},"PeriodicalIF":4.9,"publicationDate":"2025-04-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"143746598","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"医学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}