Arthritis Res Ther. 2025 Apr 11;27(1):85. doi: 10.1186/s13075-025-03508-9.

ABSTRACT

BACKGROUND: The modified Rodnan skin score (mRSS), a measure of systemic sclerosis (SSc) skin thickness, is agnostic to inflammation and vasculopathy. Previously, we demonstrated the potential of neural network-based digital pathology applied to SSc skin biopsies as a quantitative outcome. Here, we leverage deep learning and histologic analyses of clinical trial biopsies to decipher SSc skin features 'seen' by artificial intelligence (AI).

METHODS: Adults with diffuse cutaneous SSc ≤ 6 years were enrolled in an open-label trial of belumosudil [a Rho-associated coiled-coil containing protein kinase 2 (ROCK2) inhibitor]. Participants underwent serial mRSS and arm biopsies at week (W) 0, 24 and 52. Two blinded dermatopathologists scored stained sections (e.g., Masson's trichrome, hematoxylin and eosin, CD3, α-smooth muscle actin) for 16 published SSc dermal pathological parameters. We applied our deep learning model to generate QIF signatures/biopsy and obtain 'Fibrosis Scores'. Associations between Fibrosis Score and mRSS (Spearman correlation), and between Fibrosis Score and mRSS versus histologic parameters [odds ratios (OR)], were determined.

RESULTS: Only ten patients were enrolled due to early study termination, and of those, five had available biopsies due to fixation issues. Median, interquartile range (IQR) for mRSS change (0-52 W) for the ten participants was -2 (-9-7.5) and for the five with biopsies was -2.5 (-11-7.5). The correlation between Fibrosis Score and mRSS was R = 0.3; p = 0.674. Per 1-unit mRSS change (0-52 W), histologic parameters with the greatest associated changes were (OR, 95% CI, p-value): telangiectasia (2.01, [(1.31-3.07], 0.001), perivascular CD3 + (0.99, [0.97-1.02], 0.015), and % of CD8 + among CD3 + (0.95, [0.89-1.01], 0.031). Likewise, per 1-unit Fibrosis Score change, parameters with greatest changes were (OR, p-value): hyalinized collagen (1.1, [1.04 - 1.16], < 0.001), subcutaneous (SC) fat loss (1.47, [1.19-1.81], < 0.001), thickened intima (1.21, [1.06-1.38], 0.005), and eccrine entrapment (1.14, [1-1.31], 0.046).

CONCLUSIONS: Belumosudil was associated with non-clinically meaningful mRSS improvement. The histologic features that significantly correlated with Fibrosis Score changes (e.g., hyalinized collagen, SC fat loss) were distinct from those associated with mRSS changes (e.g., telangiectasia and perivascular CD3 +). These data suggest that AI applied to SSc biopsies may be useful for quantifying pathologic features of SSc beyond skin thickness.

PMID:40217251 | DOI:10.1186/s13075-025-03508-9

BMC Med Inform Decis Mak. 2025 Apr 11;25(1):161. doi: 10.1186/s12911-025-02989-7.

ABSTRACT

BACKGROUND: Thyroid nodules are frequent in clinical settings, and their diagnosis in adults is growing, with some persons experiencing symptoms. Ultrasound-guided thermal ablation can shrink nodules and alleviate discomfort. Because the degree and rate of lesion absorption vary greatly between individuals, there is no reliable model for predicting the therapeutic efficacy of thermal ablation.

METHODS: Five convolutional neural network models including VGG19, Resnet 50, EfficientNetB1, EfficientNetB0, and InceptionV3, pre-trained with ImageNet, were compared for predicting the efficacy of ultrasound-guided microwave ablation (MWA) for benign thyroid nodules using ultrasound data. The patients were randomly assigned to one of two data sets: training (70%) or validation (30%). Accuracy, sensitivity, specificity, positive predictive value, negative predictive value, and area under the curve (AUC) were all used to assess predictive performance.

RESULTS: In the validation set, fine-tuned EfficientNetB1 performed best, with an AUC of 0.85 and an ACC of 0.79.

CONCLUSIONS: The study found that our deep learning model accurately predicts nodules with VRR < 50% after a single MWA session. Indeed, when thermal therapies compete with surgery, anticipating which nodules will be poor responders provides useful information that may assist physicians and patients determine whether thermal ablation or surgery is the preferable option. This was a preliminary study of deep learning, with a gap in actual clinical applications. As a result, more in-depth study should be undertaken to develop deep-learning models that can better help clinics. Prospective studies are expected to generate high-quality evidence and improve clinical performance in subsequent research.

PMID:40217199 | DOI:10.1186/s12911-025-02989-7

Bioinformatics. 2025 Apr 11:btaf160. doi: 10.1093/bioinformatics/btaf160. Online ahead of print.

ABSTRACT

MOTIVATION: Leveraging deep learning for the representation learning of Gene Ontology (GO) and Gene Ontology Annotation (GOA) holds significant promise for enhancing downstream biological tasks such as protein-protein interaction prediction. Prior approaches have predominantly used text- and graph-based methods, embedding GO and GOA in a single geometric space (e.g., Euclidean or hyperbolic). However, since the GO graph exhibits a complex and non-monotonic hierarchy, single-space embeddings are insufficient to fully capture its structural nuances.

RESULTS: In this study, we address this limitation by exploiting geometric interaction to better reflect the intricate hierarchical structure of GO. Our proposed method, Geometry-Aware Knowledge Graph Embeddings for GO and Genes (GeOKG), leverages interactions among various geometric representations during training, thereby modeling the complex hierarchy of GO more effectively. Experiments at the GO level demonstrate the benefits of incorporating these geometric interactions, while gene-level tests reveal that GeOKG outperforms existing methods in protein-protein interaction prediction. These findings highlight the potential of using geometric interaction for embedding heterogeneous biomedical networks.

AVAILABILITY AND IMPLEMENTATION: https://github.com/ukjung21/GeOKG.

SUPPLEMENTARY INFORMATION: Supplementary data are available at Bioinformatics online.

PMID:40217132 | DOI:10.1093/bioinformatics/btaf160

Sci Rep. 2025 Apr 11;15(1):12482. doi: 10.1038/s41598-025-97565-4.

ABSTRACT

Ensuring data privacy in medical image classification is a critical challenge in healthcare, especially with the increasing reliance on AI-driven diagnostics. In fact, over 30% of healthcare organizations globally have experienced a data breach in the last year, highlighting the need for secure solutions. This study investigates the integration of transfer learning and federated learning for privacy-preserving medical image classification using GoogLeNet and VGG16 as baseline models to evaluate the generalizability of the proposed framework. Pre-trained on ImageNet and fine-tuned on three specialized medical datasets for TB chest X-rays, brain tumor MRI scans, and diabetic retinopathy images, these models achieved high classification accuracy across various aggregation methods. Additionally, the proposed dynamic aggregation method was further analyzed using modern architectures, EfficientNetV2 and ResNet-RS, to assess the scalability and robustness of the model. A key contribution is the introduction of a novel adaptive aggregation method, which dynamically alternates between Federated Averaging (FedAvg) and Federated Stochastic Gradient Descent (FedSGD), based on data divergence during communication rounds. This approach optimizes model convergence while preserving privacy in collaborative settings. The results demonstrate that transfer learning, when combined with federated learning, offers a scalable, robust, and secure solution for real-world medical diagnostics, enabling healthcare institutions to train highly accurate models without compromising sensitive patient data.

PMID:40217112 | DOI:10.1038/s41598-025-97565-4

Sci Rep. 2025 Apr 11;15(1):12383. doi: 10.1038/s41598-025-97052-w.

ABSTRACT

Hyperspectral imaging acquired from unmanned aerial vehicles (UAVs) offers detailed spectral and spatial data that holds transformative potential for precision agriculture applications, such as crop classification, health monitoring, and yield estimation. However, traditional methods struggle to effectively capture both local and global features, particularly in complex agricultural environments with diverse crop types, varying growth stages, and imbalanced data distributions. To address these challenges, we propose CMTNet, an innovative deep learning framework that integrates convolutional neural networks (CNNs) and Transformers for hyperspectral crop classification. The model combines a spectral-spatial feature extraction module to capture shallow features, a dual-branch architecture that extracts both local and global features simultaneously, and a multi-output constraint module to enhance classification accuracy through cross-constraints among multiple feature levels. Extensive experiments were conducted on three UAV-acquired datasets: WHU-Hi-LongKou, WHU-Hi-HanChuan, and WHU-Hi-HongHu. The experimental results demonstrate that CMTNet achieved overall accuracy (OA) values of 99.58%, 97.29%, and 98.31% on these three datasets, surpassing the current state-of-the-art method (CTMixer) by 0.19% (LongKou), 1.75% (HanChuan), and 2.52% (HongHu) in OA values, respectively. These findings indicate its superior potential for UAV-based agricultural monitoring in complex environments. These results advance the precision and reliability of hyperspectral crop classification, offering a valuable solution for precision agriculture challenges.

PMID:40216979 | DOI:10.1038/s41598-025-97052-w

NPJ Digit Med. 2025 Apr 11;8(1):204. doi: 10.1038/s41746-025-01606-1.

ABSTRACT

The novel coronavirus (COVID-19) pandemic has had a devastating global impact, profoundly affecting daily life, healthcare systems, and public health infrastructure. Despite the availability of treatments and vaccines, hospitalizations and deaths continue. Real-time surveillance of infection trends supports resource allocation and mitigation strategies, but reliable forecasting remains a challenge. While deep learning has advanced time-series forecasting, its effectiveness relies on large datasets, a significant obstacle given the pandemic's evolving nature. Most models use national or state-level data, limiting both dataset size and the granularity of insights. To address this, we propose the Fine-Grained Infection Forecast Network (FIGI-Net), a stacked bidirectional LSTM structure designed to leverage county-level data to produce daily forecasts up to two weeks in advance. FIGI-Net outperforms existing models, accurately predicting sudden changes such as new outbreaks or peaks, a capability many state-of-the-art models lack. This approach could enhance public health responses and outbreak preparedness.

PMID:40216974 | DOI:10.1038/s41746-025-01606-1

Sci Rep. 2025 Apr 11;15(1):12453. doi: 10.1038/s41598-025-97564-5.

ABSTRACT

Lymphedema is a chronic condition characterized by lymphatic fluid accumulation, primarily affecting the limbs. Its diagnosis is challenging due to symptom overlap with conditions like chronic venous insufficiency (CVI), deep vein thrombosis (DVT), and systemic diseases, often leading to diagnostic delays that can extend up to ten years. These delays negatively impact patient outcomes and burden healthcare systems. Conventional diagnostic methods rely heavily on clinical expertise, which may fail to distinguish subtle variations between these conditions. This study investigates the application of artificial intelligence (AI), specifically deep learning, to improve diagnostic accuracy for lower limb edema. A dataset of 1622 clinical images was used to train sixteen convolutional neural networks (CNNs) and transformer-based models, including EfficientNetV2, which achieved the highest accuracy of 78.6%. Grad-CAM analyses enhanced model interpretability, highlighting clinically relevant features such as swelling and hyperpigmentation. The AI system consistently outperformed human evaluators, whose diagnostic accuracy plateaued at 62.7%. The findings underscore the transformative potential of AI as a diagnostic tool, particularly in distinguishing conditions with overlapping clinical presentations. By integrating AI with clinical workflows, healthcare systems can reduce diagnostic delays, enhance accuracy, and alleviate the burden on medical professionals. While promising, the study acknowledges limitations, such as dataset diversity and the controlled evaluation environment, which necessitate further validation in real-world settings. This research highlights the potential of AI-driven diagnostics to revolutionize lymphedema care, bridging gaps in conventional methods and supporting healthcare professionals in delivering more precise and timely interventions. Future work should focus on external validation and hybrid systems integrating AI and clinical expertise for comprehensive diagnostic solutions.

PMID:40216943 | DOI:10.1038/s41598-025-97564-5

NPJ Digit Med. 2025 Apr 11;8(1):203. doi: 10.1038/s41746-025-01607-0.

ABSTRACT

Traditional sleep staging categorizes sleep and wakefulness into five coarse-grained classes, overlooking subtle variations within each stage. We propose a deep learning method to annotate continuous sleep depth index (SDI) with existing discrete sleep staging labels, using polysomnography from over 10,000 recordings across four large-scale cohorts. The results showcased a strong correlation between the decrease in sleep depth index and the increase in duration of arousal. Case studies indicated that SDI captured more nuanced sleep structures than conventional sleep staging. Clustering based on the digital biomarkers extracted from the SDI identified two subtypes of sleep, where participants in the disturbed subtype had a higher prevalence of several poor health conditions and were associated with a 33% increased risk of mortality and a 38% increased risk of fatal coronary heart disease. Our study underscores the utility of SDI in revealing more detailed sleep structures and yielding novel digital biomarkers for sleep medicine.

PMID:40216900 | DOI:10.1038/s41746-025-01607-0

PLoS One. 2025 Apr 11;20(4):e0319202. doi: 10.1371/journal.pone.0319202. eCollection 2025.

ABSTRACT

In industrial production, obtaining sufficient bearing fault signals is often extremely difficult, leading to a significant degradation in the performance of traditional deep learning-based fault diagnosis models. Many recent studies have shown that data augmentation using generative adversarial networks (GAN) can effectively alleviate this problem. However, the quality of generated samples is closely related to the performance of fault diagnosis models. For this reason, this paper proposes a new GAN-based small-sample bearing fault diagnosis method. Specifically, this study proposes a continuous wavelet convolution strategy (CWCL) instead of the traditional convolution operation in GAN, which can additionally capture the signal's frequency domain features. Meanwhile, this study designed a new multi-size kernel attention mechanism (MSKAM), which can extract the features of bearing vibration signals from different scales and adaptively select the features that are more important for the generation task to improve the accuracy and authenticity of the generated signals. In addition, the structural similarity index (SSIM) is adopted to quantitatively evaluate the quality of the generated signal by calculating the similarity between the generated signal and the real signal in both the time and frequency domains. Finally, we conducted extensive experiments on the CWRU and MFPT datasets and made a comprehensive comparison with existing small-sample bearing fault diagnosis methods, which verified the effectiveness of the proposed approach.

PMID:40215467 | DOI:10.1371/journal.pone.0319202

PLoS One. 2025 Apr 11;20(4):e0321116. doi: 10.1371/journal.pone.0321116. eCollection 2025.

ABSTRACT

Estimating fishing grounds is an important task in the fishing industry. This study modeled the fisher's decision-making process based on sea surface temperature patterns as a pattern recognition task. We used a deep learning-based keypoint detector to estimate fishing ground locations from these patterns. However, training the model required catch data for annotation, the amount of which was limited. To address this, we proposed a training strategy that combines weak supervision and meta-learning to estimate fishing grounds. Weak supervision involves using partially annotated or noisy data, where the labels are incomplete or imprecise. In our case, catch data cover only a subset of fishing grounds, and trajectory data, which are readily available and larger in volume than catch data, provide imprecise representations of fishing grounds. Meta-learning helps the model adapt to the noise by refining its learning rate during training. Our approach involved pre-training with trajectory data and fine-tuning with catch data, with a meta-learner further mitigating label noise during pre-training. Experimental results showed that our method improved the F1-score by 64% compared to the baseline using only catch data, demonstrating the effectiveness of pre-training and meta-learning.

PMID:40215460 | DOI:10.1371/journal.pone.0321116

PLoS One. 2025 Apr 11;20(4):e0320764. doi: 10.1371/journal.pone.0320764. eCollection 2025.

ABSTRACT

Cucumbers play a significant role as a greenhouse crop globally. In numerous countries, they are fundamental to dietary practices, contributing significantly to the nutritional patterns of various populations. Due to unfavorable environmental conditions, they are highly vulnerable to various diseases. Therefore the accurate detection of cucumber diseases is essential for maintaining crop quality and ensuring food security. Traditional methods, reliant on human inspection, are prone to errors, especially in the early stages of disease progression. Based on a VGG19 architecture, this paper uses an innovative transfer learning approach for detecting and classifying cucumber diseases, showing the applicability of artificial intelligence in this area. The model effectively distinguishes between healthy and diseased cucumber images, including Anthracnose, Bacterial Wilt, Belly Rot, Downy Mildew, Fresh Cucumber, Fresh Leaf, Pythium Fruit Rot, and Gummy Stem Blight. Using this novel approach, a balanced accuracy of 97.66% on unseen test data is achieved, compared to a balanced accuracy of 93.87% obtained with the conventional transfer learning approach, where fine-tuning is employed. This result sets a new benchmark within the dataset, highlighting the potential of deep learning techniques in agricultural disease detection. By enabling early disease diagnosis and informed agricultural management, this research contributes to enhancing crop productivity and sustainability.

PMID:40215456 | DOI:10.1371/journal.pone.0320764

Proc Natl Acad Sci U S A. 2025 Apr 15;122(15):e2423072122. doi: 10.1073/pnas.2423072122. Epub 2025 Apr 11.

ABSTRACT

Differentially private gradient descent (DP-GD) is a popular algorithm to train deep learning models with provable guarantees on the privacy of the training data. In the last decade, the problem of understanding its performance cost with respect to standard GD has received remarkable attention from the research community, which has led to upper bounds on the excess population risk [Formula: see text] in different learning settings. However, such bounds typically degrade with overparameterization, i.e., as the number of parameters [Formula: see text] gets larger than the number of training samples [Formula: see text]-a regime which is ubiquitous in current deep-learning practice. As a result, the lack of theoretical insights leaves practitioners without clear guidance, leading some to reduce the effective number of trainable parameters to improve performance, while others use larger models to achieve better results through scale. In this work, we show that in the popular random features model with quadratic loss, for any sufficiently large [Formula: see text], privacy can be obtained for free, i.e., [Formula: see text], not only when the privacy parameter [Formula: see text] has constant order but also in the strongly private setting [Formula: see text]. This challenges the common wisdom that overparameterization inherently hinders performance in private learning.

PMID:40215275 | DOI:10.1073/pnas.2423072122

IEEE Trans Image Process. 2025 Apr 11;PP. doi: 10.1109/TIP.2025.3558089. Online ahead of print.

ABSTRACT

Video Anomaly Detection (VAD) remains a fundamental yet formidable task in the video understanding community, with promising applications in areas such as information forensics and public safety protection. Due to the rarity and diversity of anomalies, existing methods only use easily collected regular events to model the inherent normality of normal spatial-temporal patterns in an unsupervised manner. Although such methods have made significant progress benefiting from the development of deep learning, they attempt to model the statistical dependency between observable videos and semantic labels, which is a crude description of normality and lacks a systematic exploration of its underlying causal relationships. Previous studies have shown that existing unsupervised VAD models are incapable of label-independent data offsets (e.g., scene changes) in real-world scenarios and may fail to respond to light anomalies due to the overgeneralization of deep neural networks. Inspired by causality learning, we argue that there exist causal factors that can adequately generalize the prototypical patterns of regular events and present significant deviations when anomalous instances occur. In this regard, we propose Causal Representation Consistency Learning (CRCL) to implicitly mine potential scene-robust causal variable in unsupervised video normality learning. Specifically, building on the structural causal models, we propose scene-debiasing learning and causality-inspired normality learning to strip away entangled scene bias in deep representations and learn causal video normality, respectively. Extensive experiments on benchmarks validate the superiority of our method over conventional deep representation learning. Moreover, ablation studies and extension validation show that the CRCL can cope with label-independent biases in multi-scene settings and maintain stable performance with only limited training data available.

PMID:40215152 | DOI:10.1109/TIP.2025.3558089

IEEE Trans Image Process. 2025 Apr 11;PP. doi: 10.1109/TIP.2025.3558420. Online ahead of print.

ABSTRACT

In this paper, we propose a new approach to train deep learning models using game theory concepts including Generative Adversarial Networks (GANs) and Adversarial Training (AT) where we deploy a double-oracle framework using best response oracles. GAN is essentially a two-player zero-sum game between the generator and the discriminator. The same concept can be applied to AT with attacker and classifier as players. Training these models is challenging as a pure Nash equilibrium may not exist and even finding the mixed Nash equilibrium is difficult as training algorithms for both GAN and AT have a large-scale strategy space. Extending our preliminary model DO-GAN, we propose the methods to apply the double oracle framework concept to Adversarial Neural Architecture Search (NAS for GAN) and Adversarial Training (NAS for AT) algorithms. We first generalize the players' strategies as the trained models of generator and discriminator from the best response oracles. We then compute the meta-strategies using a linear program. For scalability of the framework where multiple network models of best responses are stored in the memory, we prune the weakly-dominated players' strategies to keep the oracles from becoming intractable. Finally, we conduct experiments on MNIST, CIFAR-10 and TinyImageNet for DONAS-GAN. We also evaluate the robustness under FGSM and PGD attacks on CIFAR-10, SVHN and TinyImageNet for DONAS-AT. We show that all our variants have significant improvements in both subjective qualitative evaluation and quantitative metrics, compared with their respective base architectures.

PMID:40215149 | DOI:10.1109/TIP.2025.3558420

Orthod Craniofac Res. 2025 Apr 11. doi: 10.1111/ocr.12918. Online ahead of print.

ABSTRACT

Craniosynostosis is a condition characterised by the premature fusion of cranial sutures, which can lead to significant neurodevelopmental and aesthetic issues if not diagnosed and treated early. This study aimed to systematically review and conduct a meta-analysis of studies utilising machine learning (ML) models to diagnose craniosynostosis in photographs or radiographs from humans, evaluating their accuracy through sensitivity, specificity and diagnostic odds ratio. A comprehensive search was conducted on PubMed, Web of Science and Scopus until October 2024 regarding the following PECO question: 'Should ML models (E) be used to diagnose craniosynostosis in photographs or radiographs from humans (P) compared to a reference standard (C) based on their sensitivity, specificity, and diagnostic odds ratio (O)?'. Studies employing ML to diagnose craniofacial deformities on photographs and radiographs of human subjects were included. Using Quality Assessment of Diagnostic Accuracy Studies (QUADAS-2), the risk of bias was assessed. A bivariate random-effect meta-analysis was conducted to pool the diagnostic odds ratio, sensitivity and specificity of the included studies. The GRADE approach was used to evaluate the overall strength of the clinical recommendation and estimated meta-evidence. An initial search yielded 685 articles. After screening, 47 articles were selected for a full-text review. Eventually, 28 studies were selected for the systematic review, and 17 were included in the meta-analysis. The results, with an overall moderate certainty, indicated an AUC of 0.99 (95% CI: 0.98-1.00), an overall sensitivity of 97% (95% CI: 94%-98%) and an overall specificity of 97% (95% CI: 94%-99%). The estimated pooled diagnostic odds ratio was 1131 (95% CI: 290-4419). The present study showed that the ML approaches possess high efficiency and applicability in the diagnosis of craniosynostosis in photographs or radiographs from humans. These findings affirm that ML models should be considered viable diagnostic tools for craniosynostosis.

PMID:40215002 | DOI:10.1111/ocr.12918

Jpn J Radiol. 2025 Apr 11. doi: 10.1007/s11604-025-01780-y. Online ahead of print.

ABSTRACT

BACKGROUND: Esophageal squamous cell carcinoma (ESCC) poses a significant global health challenge with a particularly grim prognosis. Accurate prediction of lymph node metastasis (LNM) in ESCC is crucial for optimizing treatment strategies and improving patient outcomes. This study leverages the power of deep learning, specifically Convolutional Neural Networks (CNN) and Long Short-Term Memory (LSTM) networks, to analyze arterial phase enhanced CT images and predict LNM in ESCC patients.

METHODS: A retrospective study included 441 ESCC patients who underwent radical esophagectomy and regional lymphadenectomy. CT imaging was performed using contrast-enhanced CT scanners. Tumor region segmentation was conducted to determine the region of interest (ROI), where local tumor 3D volumes were extracted as input for the model. The novel deep learning model, LymphoReso-Net, combined CNN and LSTM networks to process and learn from medical imaging data. The model outputs a binary prediction for LNM. GRAD-CAM was integrated to enhance model interpretability. Performance was evaluated using fivefold cross-validation with metrics including accuracy, sensitivity, specificity, and AUC-ROC. The gold standard for LNM confirmation was pathologically confirmed LNM shortly after the CT.

RESULTS: LymphoReso-Net demonstrated promising performance with an average accuracy of 0.789, an AUC of 0.836, a sensitivity of 0.784, and a specificity of 0.797. GRAD-CAM provided visual explanations of the model's decision-making, aiding in identifying critical regions associated with LNM prediction.

CONCLUSION: This study introduces a novel deep learning framework, LymphoReso-Net, for predicting LNM in ESCC patients. The model's accuracy and interpretability offer valuable insights into lymphatic spread patterns, enabling more informed therapeutic decisions.

PMID:40214915 | DOI:10.1007/s11604-025-01780-y

Sleep Breath. 2025 Apr 11;29(2):155. doi: 10.1007/s11325-025-03325-z.

ABSTRACT

PURPOSE: Accurately identifying sleep states (REM, NREM, and Wake) and brief awakenings (arousals) is essential for diagnosing sleep disorders. Polysomnography (PSG) is the gold standard for such assessments but is costly and requires overnight monitoring in a lab. Home sleep testing (HST) offers a more accessible alternative, relying primarily on breathing measurements but lacks electroencephalography, limiting its ability to evaluate sleep and arousals directly. This study evaluates a deep learning algorithm which determines sleep states and arousals from breathing signals.

METHODS: A novel deep learning algorithm was developed to classify sleep states and detect arousals from respiratory inductance plethysmography signals. Sleep states were predicted for 30-s intervals (one sleep epoch), while arousal probabilities were calculated at 1-s resolution. Validation was conducted on a clinical dataset of 1,299 adults with suspected sleep disorders. Performance was assessed at the epoch level for sensitivity and specificity, with agreement analyses for arousal index (ArI) and total sleep time (TST).

RESULTS: The algorithm achieved sensitivity and specificity of 77.9% and 96.2% for Wake, 93.9% and 80.4% for NREM, 80.5% and 98.2% for REM, and 66.1% and 86.7% for arousals. Bland-Altman analysis showed ArI limits of agreement ranging from - 32 to 24 events/hour (bias: - 4.4) and TST limits from - 47 to 64 min (bias: 8.0). Intraclass correlation was 0.74 for ArI and 0.91 for TST.

CONCLUSION: The algorithm identifies sleep states and arousals from breathing signals with agreement comparable to established variability in manual scoring. These results highlight its potential to advance HST capabilities, enabling more accessible, cost-effective and reliable sleep diagnostics.

PMID:40214714 | DOI:10.1007/s11325-025-03325-z

Abdom Radiol (NY). 2025 Apr 11. doi: 10.1007/s00261-025-04929-5. Online ahead of print.

ABSTRACT

The exploration of deep learning techniques for predicting various biological characteristics of endometrial cancer (EC) is of significant importance. The objective of this study was to develop an optimized radiomics scheme combining multiparametric magnetic resonance imaging (MRI), deep learning, and machine learning to predict biological features including myometrial invasion (MI), lymph-vascular space invasion (LVSI), histologic grade (HG), and estrogen receptor (ER). This retrospective study involved 201 EC patients, who were divided into four groups according to the specific tasks. The proposed radiomics scheme extracted quantitative imaging features and multidimensional deep learning features from multiparametric MRI. Several classifiers were employed to predict biological features. Model performance and interpretability were assessed using traditional classification metrics, Gradient-weighted Class Activation Mapping (Grad-CAM), and SHapley Additive exPlanation (SHAP) techniques. In the deep MI (DMI) prediction task, the proposed protocol achieved an area under the curve (AUC) value of 0.960 (95% CI 0.9005-1.0000) in the test cohort. In the LVSI prediction task, the AUC of the proposed scheme in the test cohort was 0.924 (95% CI 0.7760-1.0000). In the HG prediction task, the AUC value of the proposed scheme in the test cohort was 0.937 (95% CI 0.8561-1.0000). In the ER prediction task, the AUC value of the proposed scheme in the test cohort was 0.929 (95% CI 0.7991-1.0000). The proposed radiomics scheme outperformed the comparative scheme and effectively extracted imaging features related to the expression of EC biological characteristics, providing potential clinical significance for accurate diagnosis and treatment decision-making.

PMID:40214699 | DOI:10.1007/s00261-025-04929-5

Minerva Anestesiol. 2025 Apr 11. doi: 10.23736/S0375-9393.25.18746-4. Online ahead of print.

ABSTRACT

The integration of artificial intelligence (AI) and all new technologies (NTs) into enhanced recovery after surgery (ERAS) protocols offers significant opportunities to address implementation challenges and improve patient care. Despite the proven benefits of ERAS, limitations such as resistance to change, resource constraints, and poor interdepartmental communication persist. AI can play a crucial role in overcoming ERAS implementation barriers by simplifying clinical plans, ensuring high compliance, and creating patient-centered approaches. Advanced techniques like machine learning and deep learning can optimize preoperative management, intraoperative phases, and postoperative recovery pathways. AI integration in ERAS protocols has the potential to revolutionize perioperative medicine by enabling personalized patient care, enhancing monitoring strategies, and improving clinical decision-making. The technology can address common postoperative challenges by developing individualized ERAS plans based on patient risk factors and optimizing perioperative processes. While challenges remain, including the need for external validation and data security, the authors suggest that the combination of AI, NTs, and ERAS protocols should become an integral part of routine clinical practice. This integration ultimately leads to improved patient outcomes and satisfaction in surgical care, transforming the perioperative medicine landscape by tailoring pathways to patients' needs.

PMID:40214219 | DOI:10.23736/S0375-9393.25.18746-4

J Med Eng Technol. 2025 Apr 11:1-8. doi: 10.1080/03091902.2025.2488827. Online ahead of print.

ABSTRACT

Medical image fusion reduces the time required for medical diagnosis by creating a composite image from a set of images belonging to different modalities. This paper introduces a deep learning framework for medical image fusion, optimising the number of convolutional layers and selecting an appropriate activation function. The conducted experiments demonstrate that employing three convolution layers with a swish activation function for the intermediate layers is sufficient to extract the salient features of the input images. The tuned features are fused using element-wise fusion rules to prevent the loss of minute details crucial for medical images. The comprehensive fused image is then reconstructed from these features using another set of three convolutional layers. Experimental results confirm that the proposed methodology outperforms other conventional medical image fusion methods in terms of various metrics and the quality of the fused image.

PMID:40214199 | DOI:10.1080/03091902.2025.2488827