ACS Sens. 2025 Mar 28. doi: 10.1021/acssensors.5c00403. Online ahead of print.

ABSTRACT

In the field of flexible sensors, the development of multifunctional, highly sensitive, wide detection range, and excellent durability sensors remains a significant challenge. This paper designs and fabricates a dual-mode ion gel sensor based on the spider's sensing mechanism, integrating both wind speed and pressure detection. The wind speed sensor employs magnetic fiber flocking and inductive resonance principles, providing accurate detection within a wind speed range of 2 to 11.5 m/s, with good linear response and high sensitivity. The impedance signal exhibits a maximum variation of 6.89 times. The pressure sensor, combining microstructured ion gel and capacitive design, demonstrates high sensitivity (15.93 kPa-1) and excellent linear response within a pressure range of 0.5 Pa to 40 kPa, with strong adaptability and good stability. The sensor shows outstanding performance in human motion monitoring, accurately capturing physiological signals such as joint movements and respiratory frequency, offering robust support for motion health management. Furthermore, combined with deep learning algorithms, the sensor achieves an accuracy of 96.83% in an intelligent motion recognition system, effectively enhancing the precision of motion performance analysis. This study provides a new solution for flexible motion monitoring and health management systems, with broad application prospects.

PMID:40152352 | DOI:10.1021/acssensors.5c00403

Phys Chem Chem Phys. 2025 Mar 28. doi: 10.1039/d4cp04605f. Online ahead of print.

ABSTRACT

Due to their inherently stochastic nature, microscopic configurations and conductance values of nano-junctions fabricated using break-junction techniques vary and fluctuate in and between experiments. Unfortunately, it is extremely difficult to observe the structural evolution of nano-junctions while measuring their conductance, a fact that prevents the establishment of their structure-conductance relationship. Herein, we conduct classical molecular dynamics (MD) simulations with neural-network potentials to simulate the stretching of Au and Ag nanowires followed by training a transformer-based neural network to predict their conductance. In addition to achieving an accuracy comparable to ab initio molecular dynamics within a computational cost similar to classical force fields, our approach can acquire the conductance of a large number of junction structures efficiently. Our calculations show that the transformer-based neural network, leveraging its self-attention mechanism, exhibits remarkable stability, accuracy and scalability in the prediction of zero-bias conductance of longer, larger and even structurally different gold nanowires when trained only on smaller systems. The simulated conductance histograms of gold nanowires are highly consistent with experiments. By examining the MD trajectories of gold nanowires simulated at 150 K and 300 K, we find that the formation probability of a three-strand planar structure appearing at 300 K is much higher than that at 150 K. This may be the dominating factor for the observed blueshift of the main peak positioned between 1.5-2G0 in the conductance histogram following the temperature increase. Moreover, our transformer-based neural network pretrained on Au has an excellent transferability, which can be fine-tuned to predict accurately the conductance of Ag nanowires with much less training data. Our findings pave the way for using deep learning techniques in molecule-scale electronics and are helpful for elucidating the conducting mechanism of molecular junctions and improving their performance.

PMID:40152302 | DOI:10.1039/d4cp04605f

Sci Prog. 2025 Jan-Mar;108(1):368504251331706. doi: 10.1177/00368504251331706. Epub 2025 Mar 28.

ABSTRACT

BackgroundMonkeypox (mpox) is a zoonotic infectious disease caused by the mpox virus and characterized by painful body lesions, fever, headaches, and exhaustion. Since the report of the first human case of mpox in Africa, there have been multiple outbreaks, even in nonendemic regions of the world. The emergence and re-emergence of mpox highlight the critical need for early detection, which has spurred research into applying deep learning to improve diagnostic capabilities.ObjectiveThis research aims to develop a robust hybrid long short-term memory (LSTM)-convolutional neural network (CNN) model with a Convolutional Block Attention Module (CBAM) to provide a potential tool for the early detection of mpox.MethodsA hybrid LSTM-CNN multi-stream deep learning model with CBAM was developed and trained using the Mpox Skin Lesion Dataset Version 2.0 (MSLD v2.0). We employed LSTM layers for preliminary feature extraction, CNN layers for further feature extraction, and CBAM for feature conditioning. The model was evaluated with standard metrics, and gradient-weighted class activation maps (Grad-CAM) and local interpretable model-agnostic explanations (LIME) were used for interpretability.ResultsThe model achieved an F1-score, recall, and precision of 94%, an area under the curve of 95.04%, and an accuracy of 94%, demonstrating competitive performance compared to the state-of-the-art models. This robust performance highlights the reliability of our model. LIME and Grad-CAM offered insights into the model's decision-making process.ConclusionThe hybrid LSTM-CNN multi-stream deep learning model with CBAM successfully detects mpox, providing a promising early detection tool that can be integrated into web and mobile platforms for convenient and widespread use.

PMID:40152267 | DOI:10.1177/00368504251331706

Bioinformatics. 2025 Mar 27:btaf135. doi: 10.1093/bioinformatics/btaf135. Online ahead of print.

ABSTRACT

MOTIVATION: Advances in bacterial promoter predictors based on ML have greatly improved identification metrics. However, existing models overlooked the impact of negative datasets, previously identified in GC-content discrepancies between positive and negative datasets in single-species models. This study aims to investigate whether multiple-species models for promoter classification are inherently biased due to the selection criteria of negative datasets. We further explore whether the generation of synthetic random sequences (SRS) that mimic GC-content distribution of promoters can partly reduce this bias.

RESULTS: Multiple-species predictors exhibited GC-content bias when using CDS as negative dataset, suggested by specificity and sensibility metrics in a species-specific manner, and investigated by dimensionality reduction. We demonstrated a reduction in this bias by employing the SRS dataset, with less detection of background noise in real genomic data. In both scenarios DNABERT showed the best metrics. These findings suggest that GC-balanced datasets can enhance the generalizability of promoter predictors across Bacteria.

AVAILABILITY AND IMPLEMENTATION: The source code of the experiments is freely available at https://github.com/maigonzalezh/MultispeciesPromoterClassifier.

PMID:40152247 | DOI:10.1093/bioinformatics/btaf135

Bioinformatics. 2025 Mar 27:btaf137. doi: 10.1093/bioinformatics/btaf137. Online ahead of print.

ABSTRACT

MOTIVATION: Single-cell RNA sequencing (scRNA-seq) analysis relies heavily on effective clustering to facilitate numerous downstream applications. Although several machine learning methods have been developed to enhance single-cell clustering, most are fully unsupervised and overlook the rich repository of annotated datasets available from previous single-cell experiments. Since cells are inherently high-dimensional entities, unsupervised clustering can often result in clusters that lack biological relevance. Leveraging annotated scRNA-seq datasets as a reference can significantly enhance clustering performance, enabling the identification of biologically meaningful clusters in target datasets.

RESULTS: In this paper, we propose Single Cell MUlti-Source CLustering (scMUSCL), a novel transfer learning method designed to identify cell clusters in a target dataset by leveraging knowledge from multiple annotated reference datasets. scMUSCL employs a deep neural network to extract domain- and batch-invariant cell representations, effectively addressing discrepancies across various source datasets and between source and target datasets within the new representation space. Unlike existing methods, scMUSCL does not require prior knowledge of the number of clusters in the target dataset and eliminates the need for batch correction between source and target datasets. We conduct extensive experiments using 20 real-life datasets, demonstrating that scMUSCL consistently outperforms existing unsupervised and transfer learning-based methods. Furthermore, our experiments show that scMUSCL benefits from multiple source datasets as learning references and accurately estimates the number of clusters.

AVAILABILITY: The Python implementation of scMUSCL is available at https://github.com/arashkhoeini/scMUSCL.

SUPPLEMENTARY INFORMATION: Supplementary data are available and include additional experimental details, performance evaluations, and implementation guidelines.

PMID:40152244 | DOI:10.1093/bioinformatics/btaf137

J Chem Inf Model. 2025 Mar 28. doi: 10.1021/acs.jcim.5c00004. Online ahead of print.

ABSTRACT

With the breakthroughs in protein structure prediction technology, constructing atomic structures from cryo-electron microscopy (cryo-EM) density maps through structural fitting has become increasingly critical. However, the accuracy of the constructed models heavily relies on the precision of the structure-to-map fitting. In this study, we introduce DEMO-EMfit, a progressive method that integrates deep learning-based backbone map extraction with a global-local structural pose search to fit atomic structures into density maps. DEMO-EMfit was extensively evaluated on a benchmark data set comprising both cryo-electron tomography (cryo-ET) and cryo-EM maps of protein and nucleic acid complexes. The results demonstrate that DEMO-EMfit outperforms state-of-the-art approaches, offering an efficient and accurate tool for fitting atomic structures into density maps.

PMID:40152222 | DOI:10.1021/acs.jcim.5c00004

Technol Cancer Res Treat. 2025 Jan-Dec;24:15330338251327081. doi: 10.1177/15330338251327081. Epub 2025 Mar 28.

ABSTRACT

The aim was to evaluate a deep learning-based auto-segmentation method for liver delineation in Y-90 selective internal radiation therapy (SIRT). A deep learning (DL)-based liver segmentation model using the U-Net3D architecture was built. Auto-segmentation of the liver was tested in CT images of SIRT patients. DL auto-segmented liver contours were evaluated against physician manually-delineated contours. Dice similarity coefficient (DSC) and mean distance to agreement (MDA) were calculated. The DL-model-generated contours were compared with the contours generated using an Atlas-based method. Ratio of volume (RV, the ratio of DL-model auto-segmented liver volume to manually-delineated liver volume), and ratio of activity (RA, the ratio of Y-90 activity calculated using a DL-model auto-segmented liver volume to Y-90 activity calculated using a manually-delineated liver volume), were assessed. Compared with the contours generated with the Atlas method, the contours generated with the DL model had better agreement with the manually-delineated contours, which had larger DSCs (average: 0.94 ± 0.01 vs 0.83 ± 0.10) and smaller MDAs (average: 1.8 ± 0.4 mm vs 7.1 ± 5.1 mm). The average RV and average RA calculated using the DL-model-generated volumes are 0.99 ± 0.03 and 1.00 ± 0.00, respectively. The DL segmentation model was able to identify and segment livers in the CT images and provide reliable results. It outperformed the Atlas method. The model can be applied for SIRT procedures.

PMID:40152005 | DOI:10.1177/15330338251327081

Network. 2025 Mar 28:1-23. doi: 10.1080/0954898X.2025.2476079. Online ahead of print.

ABSTRACT

An accurate classification of brain tumour disease is an important function in diagnosing cancer disease. Several deep learning (DL) methods have been used to identify and categorize the tumour illness. Nevertheless, the better categorized result was not consistently obtained by the traditional DL procedures. Therefore, a superior answer to this problem is offered by the optimized DL approaches. Here, the brain tumour categorization (BTC) is done using the devised Hybrid Fruit Bee Optimization based Deep Convolution Neural Network (HFBO-based DCNN). Here, the noise in the image is removed through pre-processing using a Gaussian filter. Next, the feature extraction process is done using the SegNet and this helps to extract the relevant data from the input image. Then, the feature selection is done with the help of the HFBO algorithm. Additionally, the brain tumour classification is done by the Deep CNN, and the established HFBO algorithm is used to train the weight. The devised model is analysed using the testing accuracy, sensitivity, and specificity and produced the values of 0.926, 0.926, and 0.931, respectively.

PMID:40151966 | DOI:10.1080/0954898X.2025.2476079

CRISPR J. 2025 Mar 27. doi: 10.1089/crispr.2024.0058. Online ahead of print.

ABSTRACT

Design of guide RNA (gRNA) with high efficiency and specificity is vital for successful application of the CRISPR gene editing technology. Although many machine learning (ML) and deep learning (DL)-based tools have been developed to predict gRNA activities, a systematic and unbiased evaluation of their predictive performance is still needed. Here, we provide a brief overview of in silico tools for CRISPR design and assess the CRISPR datasets and statistical metrics used for evaluating model performance. We benchmark seven ML and DL-based CRISPR-Cas9 editing efficiency prediction tools across nine CRISPR datasets covering six cell types and three species. The DL models CRISPRon and DeepHF outperform the other models exhibiting greater accuracy and higher Spearman correlation coefficient across multiple datasets. We compile all CRISPR datasets and in silico prediction tools into a GuideNet resource web portal, aiming to facilitate and streamline the sharing of CRISPR datasets. Furthermore, we summarize features affecting CRISPR gene editing activity, providing important insights into model performance and the further development of more accurate CRISPR prediction models.

PMID:40151952 | DOI:10.1089/crispr.2024.0058

Front Med (Lausanne). 2025 Mar 13;12:1507226. doi: 10.3389/fmed.2025.1507226. eCollection 2025.

ABSTRACT

INTRODUCTION: Pterygium, a prevalent ocular disorder, requires accurate severity assessment to optimize treatment and alleviate patient suffering. The growing patient population and limited ophthalmologist resources necessitate efficient AI-based diagnostic solutions. This study aims to develop an automated grading system combining deep learning and image processing techniques for precise pterygium evaluation.

METHODS: The proposed system integrates two modules: 1) A semantic segmentation module utilizing an improved TransUnet architecture for pixel-level pterygium localization, trained on annotated slit-lamp microscope images from clinical datasets. 2) A severity assessment module employing enhanced curve fitting algorithms to quantify pterygium invasion depth in critical ocular regions. The framework merges deep learning with traditional computational methods for comprehensive analysis.

RESULTS: The semantic segmentation model achieved an average Dice coefficient of 0.9489 (0.9041 specifically for pterygium class) on test datasets. In clinical validation, the system attained 0.9360 grading accuracy and 0.9363 weighted F1 score. Notably, it demonstrated strong agreement with expert evaluations (Kappa coefficient: 0.8908), confirming its diagnostic reliability.

DISCUSSION: The AI-based diagnostic method proposed in this study achieves automatic grading of pterygium by integrating semantic segmentation and curve fitting technology, which is highly consistent with the clinical evaluation of doctors. The quantitative evaluation framework established in this study is expected to meet multiple clinical needs beyond basic diagnosis. The construction of the data set should continue to be optimized in future studies.

PMID:40151829 | PMC:PMC11949100 | DOI:10.3389/fmed.2025.1507226

PeerJ Comput Sci. 2025 Mar 24;11:e2644. doi: 10.7717/peerj-cs.2644. eCollection 2025.

ABSTRACT

Social media platforms such as X, Facebook, and Instagram have become essential avenues for individuals to articulate their opinions, especially during global emergencies. These platforms offer valuable insights that necessitate analysis for informed decision-making and a deeper understanding of societal trends. Sentiment analysis is crucial for assessing public sentiment toward specific issues; however, applying it to dialectal Arabic presents considerable challenges in natural language processing. The complexity arises from the language's intricate semantic and morphological structures, along with the existence of multiple dialects. This form of analysis, also referred to as sentiment classification, opinion mining, emotion mining, and review mining, is the focus of this study, which analyzes tweets from three benchmark datasets: the Arabic Sentiment Tweets Dataset (ASTD), the A Twitter-based Benchmark Arabic Sentiment Analysis Dataset (ASAD), and the Tweets Emoji Arabic Dataset (TEAD). The research involves experimentation with a variety of comparative models, including machine learning, deep learning, transformer-based models, and a transformer-based ensemble model. Feature extraction for both machine learning and deep learning approaches is performed using techniques such as AraVec, FastText, AraBERT, and Term Frequency-Inverse Document Frequency (TF-IDF). The study compares machine learning models such as support vector machine (SVM), naïve Bayes (NB), decision tree (DT), and extreme gradient boosting (XGBoost) with deep learning models such as convolutional neural networks (CNN) and bidirectional long short-term memory (BLSTM) networks. Additionally, it explores transformer-based models such as CAMeLBERT, XLM-RoBERTa, and MARBERT, along with their ensemble configurations. The findings demonstrate that the proposed transformer-based ensemble model achieved superior performance, with average accuracy, recall, precision, and F1-score of 90.4%, 88%, 87.3%, and 87.7%, respectively.

PMID:40151815 | PMC:PMC11948314 | DOI:10.7717/peerj-cs.2644

3D Print Addit Manuf. 2025 Feb 13;12(1):11-22. doi: 10.1089/3dp.2023.0285. eCollection 2025 Feb.

ABSTRACT

Vat photopolymerization is characterized by its high precision and efficiency, making it a highly promising technique in ceramic additive manufacturing. However, the process faces a significant challenge in the form of recoating defects, necessitating real-time monitoring to maintain process stability. This article presents a defect detection method that leverages multi-image fusion and deep learning for identifying recoating defects in ceramic additive manufacturing. In the image fusion process, multiple single-channel recoating images captured by monitoring camera positioned near the photopolymerization equipment are merged with curing area mask image to create a three-channel color image. The recoating images suffer from perspective distortion due to their side view. To facilitate fusion with the curing area image, image rectification technique is applied to correct the perspective distortion, transforming the side view recoating images into a top-down view. Subsequently, the fused images are processed using a channel-wise YOLO (You Only Look Once, CW-YOLO) method to extract features, enabling the distinction of various types of defects. When compared with other deep learning models, CW-YOLO achieves higher detection accuracy while maintaining a detection rate of 103.58fps, meeting the requirements for real-time detection. Furthermore, the paper introduces the F1 score as a comprehensive evaluation metric, capturing both detection accuracy and recall rate. The results show that the F1 score is enhanced by approximately 10% after image fusion, demonstrating that the proposed method can significantly improve defect detection, particularly in cases involving difficult-to-distinguish defects like material shortages and scratches.

PMID:40151680 | PMC:PMC11937757 | DOI:10.1089/3dp.2023.0285

Front Vet Sci. 2025 Mar 13;12:1404564. doi: 10.3389/fvets.2025.1404564. eCollection 2025.

ABSTRACT

Intelligent management of large-scale farms necessitates efficient monitoring of individual livestock. To address this need, a three-phase intelligent monitoring system based on deep learning was designed, integrating a multi-part detection network for flock inventory counting, a facial classification model for facial identity recognition, and a facial expression analysis network for health assessment. For multi-part detection network, The YOLOv5s path aggregation network was modified by incorporating a multi-link convolution fusion block (MCFB) to enhance fine-grained feature extraction across objects of different sizes. To improve the detection of dense small targets, a Re-Parameterizable Convolution (RepConv) structure was introduced into the YOLOv5s head. For facial identity recognition, the sixth-stage structure in GhostNet was replaced with a four-layer spatially separable self-attention mechanism (SSSA) to strengthen key feature extraction. Additionally, model compression techniques were applied to optimize the facial expression analysis network for improved efficiency. A transfer learning strategy was employed for weight pre-training, and performance was evaluated using FPS, model weight, mean average precision (mAP), and test set accuracy. Experimental results demonstrated that the enhanced multi-part identification network effectively extracted features from different regions of the sheep flock, achieving an average detection accuracy of 95.84%, with a 2.55% improvement in mAP compared to YOLOv5s. The improved facial classification network achieved a test set accuracy of 98.9%, surpassing GhostNet by 3.1%. Additionally, the facial expression analysis network attained a test set accuracy of 99.2%, representing a 3.6% increase compared to EfficientNet. The proposed system significantly enhances the accuracy and efficiency of sheep flock monitoring by integrating advanced feature extraction and model optimization techniques. The improvements in facial classification and expression analysis further enable real-time health monitoring, contributing to intelligent livestock management.

PMID:40151568 | PMC:PMC11948620 | DOI:10.3389/fvets.2025.1404564

J Med Imaging (Bellingham). 2025 Mar;12(2):024002. doi: 10.1117/1.JMI.12.2.024002. Epub 2025 Mar 26.

ABSTRACT

PURPOSE: Previous studies on echocardiogram segmentation are focused on the left ventricle in parasternal long-axis views. Deep-learning models were evaluated on the segmentation of the ventricles in parasternal short-axis echocardiograms (PSAX-echo). Segmentation of the ventricles in complementary echocardiogram views will allow the computation of important metrics with the potential to aid in diagnosing cardio-pulmonary diseases and other cardiomyopathies. Evaluating state-of-the-art models with small datasets can reveal if they improve performance on limited data.

APPROACH: PSAX-echo was performed on 33 volunteer women. An experienced cardiologist identified end-diastole and end-systole frames from 387 scans, and expert observers manually traced the contours of the cardiac structures. Traced frames were pre-processed and used to create labels to train two domain-specific (Unet-Resnet101 and Unet-ResNet50), and four general-domain [three segment anything (SAM) variants, and the Detectron2] deep-learning models. The performance of the models was evaluated using the Dice similarity coefficient (DSC), Hausdorff distance (HD), and difference in cross-sectional area (DCSA).

RESULTS: The Unet-Resnet101 model provided superior performance in the segmentation of the ventricles with 0.83, 4.93 pixels, and 106 pixel 2 on average for DSC, HD, and DCSA, respectively. A fine-tuned MedSAM model provided a performance of 0.82, 6.66 pixels, and 1252 pixel 2 , whereas the Detectron2 model provided 0.78, 2.12 pixels, and 116 pixel 2 for the same metrics, respectively.

CONCLUSIONS: Deep-learning models are suitable for the segmentation of the left and right ventricles in PSAX-echo. We demonstrated that domain-specific trained models such as Unet-ResNet provide higher accuracy for echo segmentation than general-domain segmentation models when working with small and locally acquired datasets.

PMID:40151505 | PMC:PMC11943840 | DOI:10.1117/1.JMI.12.2.024002

Front Big Data. 2025 Mar 13;8:1455442. doi: 10.3389/fdata.2025.1455442. eCollection 2025.

ABSTRACT

The exponential growth of image and video data motivates the need for practical real-time content-based searching algorithms. Features play a vital role in identifying objects within images. However, feature-based classification faces a challenge due to uneven class instance distribution. Ideally, each class should have an equal number of instances and features to ensure optimal classifier performance. However, real-world scenarios often exhibit class imbalances. Thus, this article explores the classification framework based on image features, analyzing balanced and imbalanced distributions. Through extensive experimentation, we examine the impact of class imbalance on image classification performance, primarily on large datasets. The comprehensive evaluation shows that all models perform better with balancing compared to using an imbalanced dataset, underscoring the importance of dataset balancing for model accuracy. Distributed Gaussian (D-GA) and Distributed Poisson (D-PO) are found to be the most effective techniques, especially in improving Random Forest (RF) and SVM models. The deep learning experiments also show an improvement as such.

PMID:40151465 | PMC:PMC11948280 | DOI:10.3389/fdata.2025.1455442

Ophthalmol Sci. 2024 Dec 17;5(3):100685. doi: 10.1016/j.xops.2024.100685. eCollection 2025 May-Jun.

ABSTRACT

PURPOSE: Early glaucoma detection is key to preventing vision loss, but screening often requires specialized eye examination or photography, limiting large-scale implementation. This study sought to develop artificial intelligence models that use self-reported health data from surveys to prescreen patients at high risk for glaucoma who are most in need of glaucoma screening with ophthalmic examination and imaging.

DESIGN: Cohort study.

PARTICIPANTS: Participants enrolled from May 1, 2018, to July 1, 2022, in the nationwide All of Us Research Program who were ≥18 years of age, had ≥2 eye-related diagnoses in their electronic health record (EHR), and submitted surveys with self-reported health history.

METHODS: We developed models to predict the risk of glaucoma, as determined by EHR diagnosis codes, using 3 machine learning approaches: (1) penalized logistic regression, (2) XGBoost, and (3) a fully connected neural network. Glaucoma diagnosis was identified based on International Classification of Diseases codes extracted from EHR data. An 80/20 train-test split was implemented, with cross-validation employed for hyperparameter tuning. Input features included self-reported demographics, general health, lifestyle factors, and family and personal medical history.

MAIN OUTCOME MEASURES: Models were evaluated using standard classification metrics, including area under the receiver operating characteristic curve (AUROC).

RESULTS: Among the 8205 patients, 873 (10.64%) were diagnosed with glaucoma. Across models, AUROC scores for identifying which patients had glaucoma from survey health data ranged from 0.710 to 0.890. XGBoost achieved the highest AUROC of 0.890 (95% confidence interval [CI]: 0.860-0.910). Logistic regression followed with an AUROC of 0.772 (95% CI: 0.753-0.795). Explainability studies revealed that key features included traditionally recognized risk factors for glaucoma, such as age, type 2 diabetes, and a family history of glaucoma.

CONCLUSIONS: Machine and deep learning models successfully utilized health data from self-reported surveys to predict glaucoma diagnosis without additional data from ophthalmic imaging or eye examination. These models may eventually enable prescreening for glaucoma in a wide variety of low-resource settings, after which high-risk patients can be referred for targeted screening using more specialized ophthalmic examination or imaging.

FINANCIAL DISCLOSURES: The author(s) have no proprietary or commercial interest in any materials discussed in this article.

PMID:40151359 | PMC:PMC11946806 | DOI:10.1016/j.xops.2024.100685

Comput Biol Med. 2025 Mar 26;190:110070. doi: 10.1016/j.compbiomed.2025.110070. Online ahead of print.

ABSTRACT

This study explores the development of a deep learning model using a neck-wearable piezoelectric sensor to accurately distinguish severe sleep apnea syndrome (SAS) from habitual snoring, addressing the underdiagnosis of SAS in adults. From 2018 to 2020, 60 adult habitual snorers underwent polysomnography while wearing a neck piezoelectric sensor that recorded snoring vibrations (70-250 Hz) and carotid artery pulsations (0.01-1.5 Hz). The initial dataset comprised 1167 silence, 1304 snoring, and 399 noise samples from 20 participants. Using a hybrid deep learning model comprising a one-dimensional convolutional neural network and gated-recurrent unit, the model identified snoring and apnea/hypopnea events, with sleep phases detected via pulse wave variability criteria. The model's efficacy in predicting severe SAS was assessed in the remaining 40 participants, achieving snoring detection rates of 0.88, 0.86, and 0.92, with respective loss rates of 0.39, 0.90, and 0.23. Classification accuracy for severe SAS improved from 0.85 for total sleep time to 0.90 for partial sleep time, excluding the first sleep phase, demonstrating precision of 0.84, recall of 1.00, and an F1 score of 0.91. This innovative approach of combining a hybrid deep learning model with a neck-wearable piezoelectric sensor suggests a promising route for early and precise differentiation of severe SAS from habitual snoring, aiding guiding further standard diagnostic evaluations and timely patient management. Future studies should focus on expanding the sample size, diversifying the patient population, and external validations in real-world settings to enhance the robustness and applicability of the findings.

PMID:40147187 | DOI:10.1016/j.compbiomed.2025.110070

Radiography (Lond). 2025 Mar 26;31(3):102929. doi: 10.1016/j.radi.2025.102929. Online ahead of print.

ABSTRACT

INTRODUCTION: A hybrid angio-CT system with 320-row detectors and deep learning-based reconstruction (DLR), provides additional imaging via 4D-CT angiography (CTA), potentially shortening procedure time and reducing DSA acquisitions, contrast media, and radiation dose. This study evaluates the feasibility of low-tube voltage 4D-CTA with low-contrast volume and DLR for selective renal artery embolization using a vessel phantom.

METHODS: A custom-made phantom simulating contrast-enhanced vessels filled with contrast medium was scanned. The study assessed image quality under varying image noise and vessel contrast. Quantitative analysis included peak contrast-to-noise ratio (pCNR) and image noise. Qualitative assessment was performed by seven radiologists using a 4-point scale; each radiologist independently recorded their evaluations on an assessment sheet.

RESULTS: A pCNR of approximately 15.0 was identified as the threshold for acceptable image quality. The pCNR decreased as the noise index increased (by 25-75 % when comparing a noise index of 30-70 HU).Vessels with a CT value of 500 Hounsfield units (HU) achieved sufficient image quality with a noise index of 50 HU. Dose reduction was substantial compared to traditional DSA, with effective radiation dose remaining within acceptable clinical levels.

CONCLUSION: 4D-CTA, combined with DLR, demonstrated the potential to reduce radiation and contrast agent usage while preserving diagnostic quality for renal artery angiography. Further clinical validation is required to confirm these findings in clinical settings.

IMPLICATIONS FOR PRACTICE: 4D-CTA with low-tube voltage and deep learning-based reconstruction (DLR) can reduce radiation and contrast use while maintaining image quality. This approach might improve safety, particularly in patients with renal impairment, and serve as a viable alternative to conventional DSA for selective renal artery embolization.

PMID:40147091 | DOI:10.1016/j.radi.2025.102929

Acad Radiol. 2025 Mar 26:S1076-6332(25)00263-6. doi: 10.1016/j.acra.2025.03.030. Online ahead of print.

NO ABSTRACT

PMID:40148166 | DOI:10.1016/j.acra.2025.03.030

Lancet Digit Health. 2025 Apr;7(4):e264-e274. doi: 10.1016/j.landig.2025.01.001.

ABSTRACT

BACKGROUND: Left ventricular systolic dysfunction (LVSD) is independently associated with cardiovascular events in patients with congenital heart disease. Although artificial intelligence-enhanced electrocardiogram (AI-ECG) analysis is predictive of LVSD in the general adult population, it has yet to be applied comprehensively across congenital heart disease lesions.

METHODS: We trained a convolutional neural network on paired ECG-echocardiograms (≤2 days apart) across the lifespan of a wide range of congenital heart disease lesions to detect left ventricular ejection fraction (LVEF) of 40% or less. Model performance was evaluated on single ECG-echocardiogram pairs per patient at Boston Children's Hospital (Boston, MA, USA) and externally at the Children's Hospital of Philadelphia (Philadelphia, PA, USA) using area under the receiver operating (AUROC) and precision-recall (AUPRC) curves.

FINDINGS: The training cohort comprised 124 265 ECG-echocardiogram pairs (49 158 patients; median age 10·5 years [IQR 3·5-16·8]; 3381 [2·7%] of 124 265 ECG-echocardiogram pairs with LVEF ≤40%). Test groups included internal testing (21 068 patients; median age 10·9 years [IQR 3·7-17·0]; 3381 [2·7%] of 124 265 ECG-echocardiogram pairs with LVEF ≤40%) and external validation (42 984 patients; median age 10·8 years [IQR 4·9-15·0]; 1313 [1·7%] of 76 400 ECG-echocardiogram pairs with LVEF ≤40%) cohorts. High model performance was achieved during internal testing (AUROC 0·95, AUPRC 0·33) and external validation (AUROC 0·96, AUPRC 0·25) for a wide range of congenital heart disease lesions. Patients with LVEF greater than 40% by echocardiogram who were deemed high risk by AI-ECG were more likely to have future dysfunction compared with low-risk patients (hazard ratio 12·1 [95% CI 8·4-17·3]; p<0·0001). High-risk patients by AI-ECG were at increased risk of mortality in the overall cohort and lesion-specific subgroups. Common salient features highlighted across congenital heart disaese lesions include precordial QRS complexes and T waves, with common high-risk ECG features including deep V2 S waves and lateral precordial T wave inversion. A case study on patients with ventricular pacing showed similar findings.

INTERPRETATION: Our externally validated algorithm shows promise in prediction of current and future LVSD in patients with congenital heart disease, providing a clinically impactful, inexpensive, and convenient cardiovascular health tool in this population.

FUNDING: Kostin Innovation Fund, Thrasher Research Fund Early Career Award, Boston Children's Hospital Electrophysiology Research Education Fund, National Institutes of Health, National Institute of Childhood Diseases and Human Development, and National Library of Medicine.

PMID:40148010 | DOI:10.1016/j.landig.2025.01.001