Deep learning

The road to commercial success for neuromorphic technologies

Neuromorphic technologies adapt biological neural principles to synthesise high-efficiency computational devices, characterised by continuous real-time operation and sparse event-based communication. After several false starts, a confluence of advances now promises widespread commercial adoption. Gradient-based training of deep spiking neural networks is now an off-the-shelf technique for building general-purpose Neuromorphic applications, with open-source tools underwritten by theoretical results. Analog and mixed-signal Neuromorphic circuit designs are being replaced by digital equivalents in newer devices, simplifying application deployment while maintaining computational benefits. Designs for in-memory computing are also approaching commercial maturity. Solving two key problems-how to program general Neuromorphic applications; and how to deploy them at scale-clears the way to commercial success of Neuromorphic processors. Ultra-low-power Neuromorphic technology will find a home in battery-powered systems, local compute for internet-of-things devices, and consumer wearables. Inspiration from uptake of tensor processors and GPUs can help the field overcome remaining hurdles.

Multi-scale convolutional transformer network for motor imagery brain-computer interface

Brain-computer interface (BCI) systems allow users to communicate with external devices by translating neural signals into real-time commands. Convolutional neural networks (CNNs) have been effectively utilized for decoding motor imagery electroencephalography (MI-EEG) signals in BCIs. However, traditional CNN-based methods face challenges such as individual variability in EEG signals and the limited receptive fields of CNNs. This study presents the Multi-Scale Convolutional Transformer (MSCFormer) model that integrates multiple CNN branches for multi-scale feature extraction and a Transformer module to capture global dependencies, followed by a fully connected layer for classification. The multi-branch multi-scale CNN structure effectively addresses individual variability in EEG signals, enhancing the model's generalization capabilities, while the Transformer encoder strengthens global feature integration and improves decoding performance. Extensive experiments on the BCI IV-2a and IV-2b datasets show that MSCFormer achieves average accuracies of 82.95% (BCI IV-2a) and 88.00% (BCI IV-2b), with kappa values of 0.7726 and 0.7599 in five-fold cross-validation, surpassing several state-of-the-art methods. These results highlight MSCFormer's robustness and accuracy, underscoring its potential in EEG-based BCI applications. The code has been released in https://github.com/snailpt/MSCFormer .

Multi scale convolutional neural network combining BiLSTM and attention mechanism for bearing fault diagnosis under multiple working conditions

Bearing fault diagnosis is of great significance for ensuring the safety of rotating electromechanical equipment. A deep learning network framework for diagnosing bearing faults under multiple load conditions is proposed to address the problems of extracting a single feature scale from bearing vibration timing signals, inability to simultaneously utilize spatial and bidirectional time features, and difficulty in obtaining sufficient training data under multiple working conditions. The first and second convolutional layers of a convolutional neural network (CNN) are used to simultaneously extract the spatio-temporal features from the bearing vibration signal and fuse them to obtain multi-scale spatiotemporal features. Based on this, BiLSTM is further applied to extract the bi-directional temporal correlation features of the input sequence. By introducing an attention mechanism (AM) to assign greater weights to critical spatio-temporal features, a new multi-scale deep learning network which integrates CNN, BiLSTM, and AM (MSCNN-BiLSTM-AM) network is proposed to obtain key bearing state features and accurate fault diagnose results. To further improve the adaptability of the network to different load conditions, the parameters of pretrained MSCNN-BiLSTM-AM network are applied to initialize the new task model parameters. After that, the new task diagnostic network is trained and validated under new load conditions by freezing the parameters of CNN, BiLSTM and AM layer, and fine-tuning the parameters of the fully connected layer and output layer. The experiments verify the excellent performance of the proposed method, while effectively solving the challenges of model training and fault diagnosis when there are insufficient training samples under multiple working conditions.

Artificial Intelligence in Diagnosis of Heart Failure

Heart failure (HF) is a complex and varied condition that affects over 50 million people worldwide. Although there have been significant strides in understanding the underlying mechanisms of HF, several challenges persist, particularly in the accurate diagnosis of HF. These challenges include issues related to its classification, the identification of specific phenotypes, and the assessment of disease severity. Artificial intelligence (AI) algorithms have the potential to transform HF care by enhancing clinical decision-making processes, enabling the early detection of patients at risk for subclinical or worsening HF. By integrating and analyzing vast amounts of data with intricate multidimensional interactions, AI algorithms can provide critical insights that help physicians make more timely and informed decisions. In this review, we explore the challenges in current diagnosis of HF, basic AI concepts and common AI algorithms, and latest AI research in HF diagnosis.

Maize yield estimation in Northeast China's black soil region using a deep learning model with attention mechanism and remote sensing

Accurate prediction of maize yields is crucial for effective crop management. In this paper, we propose a novel deep learning framework (CNNAtBiGRU) for estimating maize yield, which is applied to typical black soil areas in Northeast China. This framework integrates a one-dimensional convolutional neural network (1D-CNN), bidirectional gated recurrent units (BiGRU), and an attention mechanism to effectively characterize and weight key segments of input data. In the predictions for the most recent year, the model demonstrated high accuracy (R² = 0.896, RMSE = 908.33 kg/ha) and exhibited strong robustness in both earlier years and during extreme climatic events. Unlike traditional yield estimation methods that primarily rely on remote sensing vegetation indices, phenological data, meteorological data, and soil characteristics, this study innovatively incorporates anthropogenic factors, such as Degree of Cultivation Mechanization (DCM), reflecting the rapid advancement of agricultural modernization. The relative importance analysis of input variables revealed that Enhanced Vegetation Index (EVI), Sun-Induced Chlorophyll Fluorescence (SIF), and DCM were the most influential factors in yield prediction. Furthermore, our framework enables maize yield prediction 1-2 months in advance by leveraging historical patterns of environmental and agricultural variables, providing valuable lead time for decision-making. This predictive capability does not rely on forecasting future weather conditions but rather captures yield-relevant signals embedded in early-season data.

AI as teacher: effectiveness of an AI-based training module to improve trainee pediatric fracture detection

OBJECTIVE

Prior work has demonstrated that AI access can help residents more accurately detect pediatric fractures. We wished to evaluate the effectiveness of an unsupervised AI-based training module as a pediatric fracture detection educational tool.

MATERIALS AND METHODS

Two hundred forty radiographic examinations from throughout the pediatric upper extremity were split into two groups of 120 examinations. A previously developed open-source deep learning fracture detection algorithm ( www.childfx.com ) was used to annotate radiographs. Four medical students and four PGY-2 radiology residents first evaluated 120 examinations for fracture without AI assistance and subsequently reviewed AI annotations on these cases via a training module. They then interpreted 120 different examinations without AI assistance. Pre- and post-intervention fracture detection accuracy was evaluated using a chi-squared test.

RESULTS

Overall resident fracture detection accuracy significantly improved from 71.3% pre-intervention to 77.5% post-intervention (p = 0.032). Medical student fracture detection accuracy was not significantly changed from 56.3% pre-intervention to 57.3% post-intervention (p = 0.794). Eighty-eight percent of responding participants (7/8) would recommend this model of learning.

CONCLUSION

We found that a tailored AI-based training module increased resident accuracy for detecting pediatric fractures by 6.2%. Medical student accuracy was not improved, likely due to their limited background familiarity with the task. AI offers a scalable method for automatically generating annotated teaching cases covering varied pathology, allowing residents to efficiently learn from simulated experience.

The Role of Artificial Intelligence in the Evaluation of Prostate Pathology

Artificial intelligence (AI) is an emerging tool in diagnostic pathology, including prostate pathology. This review summarizes the possibilities offered by AI and also discusses the challenges and risks. AI has the potential to assist in the diagnosis and grading of prostate cancer. Diagnostic safety can be enhanced by avoiding the accidental underdiagnosis of small lesions. Another possible benefit is a greater degree of standardization of grading. AI for clinical use needs to be trained on large, high-quality data sets that have been assessed by experienced pathologists. A problem with the use of AI in prostate pathology is the plethora of benign mimics of prostate cancer and morphological variants of cancer that are too unusual to allow sufficient training of AI. AI systems need to be able to account for variations in local routines for cutting, staining, and scanning of slides. We also need to be aware of the risk that users will rely too much on the output of an AI system, leading to diagnostic errors and loss of clinical competence. The reporting pathologist must ultimately be responsible for accepting or rejecting the diagnosis proposed by AI.

Prediction of Chromatographic Retention Time of a Small Molecule from SMILES Representation Using a Hybrid Transformer-LSTM Model

Accurate retention time (RT) prediction in liquid chromatography remains a significant consideration in molecular analysis. In this study, we explore the use of a transformer-based language model to predict RTs by treating simplified molecular input line entry system (SMILES) sequences as textual input, an approach that has not been previously utilized in this field. Our architecture combines a pretrained RoBERTa (robustly optimized BERT approach, a variant of BERT) with bidirectional long short-term memory (BiLSTM) networks to predict retention times in reversed-phase high-performance liquid chromatography (RP-HPLC). The METLIN small molecule retention time (SMRT) data set comprising 77,980 small molecules after preprocessing, was encoded using SMILES notation and processed through a tokenizer to enable molecular representation as sequential data. The proposed transformer-LSTM architecture incorporates layer fusion from multiple transformer layers and bidirectional sequence processing, achieving superior performance compared to existing methods with a mean absolute error (MAE) of 26.23 s, a mean absolute percentage error (MAPE) of 3.25%, and R-squared (R2) value of 0.91. The model's explainability was demonstrated through attention visualization, revealing its focus on key molecular features that can influence RT. Furthermore, we evaluated the model's transfer learning capabilities across ten data sets from the PredRet database, demonstrating robust performance across different chromatographic conditions with consistent improvement over previous approaches. Our results suggest that the hybrid model presents a valuable approach for predicting RT in liquid chromatography, with potential applications in metabolomics and small molecule analysis.

Artificial intelligence in the diagnosis and management of refractive errors

Refractive error is among the leading causes of visual impairment globally. The diagnosis and management of refractive error has traditionally relied on comprehensive eye examinations by eye care professionals, but access to these specialized services has remained limited in many areas of the world. Given this, artificial intelligence (AI) has shown immense potential in transforming the diagnosis and management of refractive error. We review AI applications across various aspects of refractive error care - from axial length prediction using fundus images to risk stratification for myopia progression. AI algorithms can be trained to analyze clinical data to detect refractive error as well as predict associated risks of myopia progression. For treatments such as implantable collamer and orthokeratology lenses, AI models facilitate vault size prediction and optimal lens fitting with high accuracy. Furthermore, AI has demonstrated promise in optimizing surgical planning and outcomes for refractive procedures. Emerging digital technologies such as telehealth, smartphone applications, and virtual reality integrated with AI present novel avenues for refractive error screening. We discuss key challenges, including limited validation datasets, lack of data standardization, image quality issues, population heterogeneity, practical deployment, and ethical considerations regarding patient privacy that need to be addressed before widespread clinical implementation.

Integrating advanced deep learning techniques for enhanced detection and classification of citrus leaf and fruit diseases

In this study, we evaluate the performance of four deep learning models, EfficientNetB0, ResNet50, DenseNet121, and InceptionV3, for the classification of citrus diseases from images. Extensive experiments were conducted on a dataset of 759 images distributed across 9 disease classes, including Black spot, Canker, Greening, Scab, Melanose, and healthy examples of fruits and leaves. Both InceptionV3 and DenseNet121 achieved a test accuracy of 99.12%, with a macro average F1-score of approximately 0.986 and a weighted average F1-score of 0.991, indicating exceptional performance in terms of precision and recall across the majority of the classes. ResNet50 and EfficientNetB0 attained test accuracies of 84.58% and 80.18%, respectively, reflecting moderate performance in comparison. These research results underscore the promise of modern convolutional neural networks for accurate and timely detection of citrus diseases, thereby providing effective tools for farmers and agricultural professionals to implement proactive disease management, reduce crop losses, and improve yield quality.

Artificial intelligence in the diagnosis of obstructive sleep apnea: a scoping review

PURPOSE

The gold standard diagnostic modality of Obstructive Sleep Apnea (OSA) is polysomnography (PSG), which is resource-intensive, requires specialized facilities, and may not be accessible to all patients. There is a growing body of research exploring the potential of artificial intelligence (AI) to offer more accessible, efficient, and cost-effective alternatives for the diagnosis of OSA.

METHODS

We conducted a scoping review of studies applying AI techniques to diagnose and assess OSA in adult populations. A comprehensive search was performed in the Web of Science database using terms related to "obstructive sleep apnea," "artificial intelligence," "machine learning," and related approaches.

RESULTS

A total of 344 articles met the inclusion criteria. The findings highlight various methodologies of disease evaluation, including binary classification distinguishing between OSA-positive and OSA-negative individuals in 118 articles, OSA event detection in 211 articles, severity evaluation in 38 articles, topographic diagnostic evaluation in 8 articles, and apnea-hypopnea index (AHI) estimation in 26 articles. 40 distinct types of data sources were identified. The three most prevalent data types were electrocardiography (ECG), used in 108 articles, photoplethysmography (PPG) in 62 articles, and respiratory effort and body movement in 44 articles. The AI techniques most frequently applied were convolutional neural networks (CNNs) in 104 articles, support vector machines (SVMs) in 91 articles, and K-Nearest Neighbors (KNN) in 57 articles. Of these studies, 229 used direct patient recruitment, and 115 utilized existing datasets.

CONCLUSION

While AI demonstrates substantial potential with high accuracy rates in certain studies, challenges remain such as model transparency, validation across diverse populations, and seamless integration into clinical practice. These challenges may stem from factors such as overfitting to specific datasets, limited generalizability, and the need for standardized protocols in clinical settings.

Continuous sleep depth index annotation with deep learning yields novel digital biomarkers for sleep health

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.

Deep learning-driven multi-class classification of brain strokes using computed tomography: A step towards enhanced diagnostic precision

OBJECTIVE

To develop and validate deep learning models leveraging CT imaging for the prediction and classification of brain stroke conditions, with the potential to enhance accuracy and support clinical decision-making.

MATERIALS AND METHODS

This retrospective, bi-center study included data from 250 patients, with a dataset of 8186 CT images collected from 2017 to 2022. Two AI models were developed using the Expanded ResNet101 deep learning framework as a two-step model. Model performance was evaluated using confusion matrices, supplemented by external validation with an independent dataset. External validation was conducted by an expert and two external members. Overall accuracy, confidence intervals, Cohen's Kappa value, and McNemar's test P-values were calculated.

RESULTS

A total of 8186 CT images were incorporated, with 6386 images used for the training and 900 datasets for testing and validation in Model 01. Further, 1619 CT images were used for training and 600 datasets for testing and validation in Model 02. The average accuracy, precision, and F1 score for both models were assessed: Model 01 achieved 99.6 %, 99.4 %, and 99.6 % respectively, whereas Model 02 achieved 99.2 %, 98.8 %, and 99.1 %. The external validation accuracies were 78.6 % (95 % CI: 0.73,0.83; P < 0.001) and 60.2 % (95 % CI: 0.48,0.70; P < 0.001) for Models 01 and 02 respectively, as evaluated by the expert.

CONCLUSION

Deep learning models demonstrated high accuracy, precision, and F1 scores in predicting outcomes for brain stroke patients. With larger cohort and diverse radiologic mimics, these models could support clinicians in prognosis and decision-making.

Fine-grained forecasting of COVID-19 trends at the county level in the United States

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.

Predicting the efficacy of microwave ablation of benign thyroid nodules from ultrasound images using deep convolutional neural networks

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.

Hierarchy of Chaotic Dynamics in Random Modular Networks

We introduce a model of randomly connected neural populations and study its dynamics by means of the dynamical mean-field theory and simulations. Our analysis uncovers a rich phase diagram, featuring high- and low-dimensional chaotic phases, separated by a crossover region characterized by low values of the maximal Lyapunov exponent and participation ratio dimension, but with high values of the Lyapunov dimension that change significantly across the region. Counterintuitively, chaos can be attenuated by either adding noise to strongly modular connectivity or by introducing modularity into random connectivity. Extending the model to include a multilevel, hierarchical connectivity reveals that a loose balance between activities across levels drives the system towards the edge of chaos.

Data-driven identification of pollution sources and water quality prediction using Apriori and LSTM models: A case study in the Hanjiang River basin

The rapid development of urbanization and industrialization has exacerbated surface water pollution, especially from point sources such as industrial discharge and urban wastewater, posing a severe challenge to global environmental health and sustainable development. This study combines the Apriori algorithm and Long Short-Term Memory (LSTM) networks to identify major pollution sources and predict dynamic changes in water quality. The study area encompasses four national monitoring hydrological stations in the core area of the South-to-North Water Diversion Project, with multi-source data collected, including water quality parameters and industry-specific discharge data. Using the Apriori algorithm, the pollutants with the highest support-chemical oxygen demand (COD), copper (Cu), suspended solids (SS), and zinc (Zn)-demonstrated a support value of 0.87, indicating that the metallurgical, electroplating, and chemical industries are the primary pollution sources. Further association rule analysis based on varying parameter thresholds revealed that when COD is present, the co-occurrence confidence for Cadmium (Cd), Cu, Lead (Pb), and SS reaches 0.9, and the combination of COD, Cu, Pb, SS, and Cyanide (CN) achieves a confidence level of 1, indicating a high degree of correlation among these pollutants. The LSTM model demonstrated high accuracy in water quality prediction, with Root Mean Square Error (RMSE) values for COD predictions at each hydrological station ranging from 0.2076 to 0.3366, and coefficients of determination (R2) all exceeding 0.9, highlighting the model's stability and predictive accuracy. This study provides a scientific basis for the sustainable management of watershed water resources and serves as a significant reference for environmental policymaking and water resource protection.

Optimized classification of dental implants using convolutional neural networks and pre-trained models with preprocessed data

OBJECTIVE

This study evaluates the performance of various classifiers and pre-trained models for dental implant state classification using preprocessed radiography images with masks.

METHODOLOGY

A dataset of 511 periapical images, including 275 for Bicon, 70 for Bego, and 166 for ITI implants, was expanded to 5110 images using data augmentation techniques such as rotation, flipping, and scaling. Preprocessing included resizing, sharpening, noise reduction, CLAHE-based contrast enhancement, implant-specific masking, and normalization. Classifiers including Convolutional Neural Networks (CNN), Convolutional Support Vector Machine (CSVM), Convolutional Decision Tree (CDT), and Convolutional Random Forest (CRF) were employed. Pre-trained models such as VGG16, ResNet50, and Xception enhanced feature extraction. Model performance was assessed using accuracy, precision, recall, F1 score, and ROC AUC, with fivefold cross-validation ensuring robustness.

RESULTS

CRF achieved the highest performance for ITI with Bego implants, with accuracy of 0.8966, precision of 0.9364, recall of 0.9253, F1 score of 0.9304, and ROC AUC of 0.9351. CNN delivered the best results for Bicon with Bego implants, achieving 0.9533 accuracy. Among pre-trained models, VGG16 with preprocessed data achieved superior results for Bicon vs. ITI classification, with 0.9865 accuracy and 0.9877 ROC AUC. Data augmentation and preprocessing significantly improved classifier performance.

CONCLUSION

Preprocessing steps, coupled with data augmentation, enhanced classification performance, ensuring robustness across models. CRF and CNN were the top-performing classifiers, with VGG16 excelling among pre-trained models. These results highlight the importance of data augmentation and preprocessing in improving dental implant classification accuracy.

Predicting PD-L1 status in NSCLC patients using deep learning radiomics based on CT images

Radiomics refers to the utilization of automated or semi-automated techniques to extract and analyze numerous quantitative features from medical images, such as computerized tomography (CT) or magnetic resonance imaging (MRI) scans. This study aims to develop a deep learning radiomics (DLR)-based approach for predicting programmed death-ligand 1 (PD-L1) expression in patients with non-small cell lung cancer (NSCLC). Data from 352 NSCLC patients with known PD-L1 expression were collected, of which 48.29% (170/352) were tested positive for PD-L1 expression. Tumor regions of interest (ROI) were semi-automatically segmented based on CT images, and DL features were extracted using Residual Network 50. The least absolute shrinkage and selection operator (LASSO) algorithm was used for feature selection and dimensionality reduction. Seven algorithms were used to build models, and the most optimal ones were identified. A combined model integrating DLR with clinical data was also developed. The predictive performance of each model was evaluated using the area under the curve (AUC) of the receiver operating characteristic (ROC) curve analysis. The DLR model, based on CT images, demonstrated an AUC of 0.85 (95% confidence interval (CI), 0.82-0.88), sensitivity of 0.80 (0.74-0.85), and specificity of 0.73 (0.70-0.77) for predicting PD-L1 status. The integrated model exhibited superior performance, with an AUC of 0.91 (0.87-0.95), sensitivity of 0.85 (0.82-0.89), and specificity of 0.75 (0.72-0.80). Our findings indicate that the DLR model holds promise as a valuable tool for predicting the PD-L1 status in patients with NSCLC, which can greatly assist in clinical decision-making and the selection of personalized treatment strategies.

Properties of neural networks identifying strongly lensed gravitational waves in time domain

Just as light, gravitational waves (GWs) can also be lensed. In the case of strong lensing, one would observe multiple copies of the same initial wave with different amplitudes, arrival times and possibly phases. As the GW detection rate increases with improving detectors, it will become more and more difficult for the analyses searching for strong-lensing signatures to keep up. Machine learning models were proposed to address this issue. In this work, we present the performance of an attention-based neural network (AttNN) trained to identify strongly lensed pairs of GWs in the time domain. We also investigate its properties and those of a convolutional neural network (CNN) to identify their advantages and limitations.This article is part of the Theo Murphy meeting issue 'Multi-messenger gravitational lensing (Part 1)'.

Evaluating the three-level approach of the U-smile method for imbalanced binary classification

Real-life binary classification problems often involve imbalanced datasets, where the majority class outnumbers the minority class. We previously developed the U-smile method, which comprises the U-smile plot and the BA, RB and I coefficients, to assess the usefulness of a new variable added to a reference prediction model and validated it under class balance. In this study, we evaluated the U-smile method under class imbalance, proposed a three-level approach of the U-smile method, and used the I coefficients as a weighting factor for point size in the U-smile plots of the BA and RB coefficients. Using real data from the Heart Disease dataset and generated random variables, we built logistic regression models to assess four new variables added to the reference model (nested setting). These models were evaluated at seven pre-defined imbalance levels of 1%, 10%, 30%, 50%, 70%, 90% and 99% of the event class. The results of the U-smile method were compared to those of certain traditional measures: Brier skill score, net reclassification index, difference in F1-score, difference in Matthews correlation coefficient, difference in the area under the receiver operating characteristic curve of the new and reference models, and the likelihood-ratio test. The reference model overfitted to the majority class at higher imbalance levels. The BA-RB-I coefficients of the U-smile method identified informative variables across the entire imbalance range. At higher imbalance levels, the U-smile method indicated both prediction improvement in the minority class (positive BA and I coefficients) and reduction in overfitting to the majority class (negative RB coefficients). The U-smile method outperformed traditional evaluation measures across most of the imbalance range. It proved highly effective in variable selection for imbalanced binary classification, making it a useful tool for real-life problems, where imbalanced datasets are prevalent.

Automatic Cry Analysis: Deep Learning for Screening of Autism Spectrum Disorder in Early Childhood

PURPOSE

The objective of this study is to identify the acoustic characteristics of cries of Typically Developing (TD) and Autism Spectrum Disorder (ASD) children via Deep Learning (DL) techniques to support clinicians in the early detection of ASD.

METHODS

We used an existing cry dataset that included 31 children with ASD and 31 TD children aged between 18 and 54 months. Statistical analysis was applied to find differences between groups for different voice acoustic features such as jitter, shimmer and harmonics-to-noise ratio (HNR). A DL model based on Recursive Convolutional Neural Networks (R-CNN) was developed to classify cries of ASD and TD children.

RESULTS

We found a statistical significant increase in jitter and shimmer for ASD cries compared to TD, as well as a decrease in HNR for ASD cries. Additionally, the DL algorithm achieved an accuracy of 90.28% in differentiating ASD cries from TD.

CONCLUSION

Empowering clinicians with automatic non-invasive Artificial Intelligence (AI) tools based on cry vocal biomarkers holds considerable promise in advancing early detection and intervention initiatives for children at risk of ASD, thereby improving their developmental trajectories.

Leveraging computer vision, large language models, and multimodal machine learning for optimal decision-making in dairy farming

This article explores various applications of artificial intelligence technologies in dairy farming, including the use of computer vision systems (CVS) for animal identification, body condition score (BCS) and body shape analysis, and potential uses of LLMs in the dairy industry. Among recent advancements in precision livestock farming (PLF) tools, CVS have gained popularity as powerful solutions for individual animal monitoring. These systems can capture phenotypes from multiple animals simultaneously using a single device in an automated and non-intrusive manner. To match animals with their corresponding predicted phenotypes, these systems require individual animal identification, which can be achieved through external identification systems or computer vision-based animal identification algorithms. Additionally, modern natural language processing techniques, such as large language models (LLMs), offer opportunities for advanced data integration, including unstructured textual data. Furthermore, we discuss the challenges associated with integrating data from different sources and modalities - such as images, text, and tabular data - into multimodal machine learning systems for phenotype prediction, which also represents a key area of artificial intelligence application. Digital technologies such as CVS and LLMs have the potential to transform dairy farming. CVS can provide individual and objective assessments of animal health, while LLMs can integrate diverse data sources for phenotype prediction. While there is much potential ahead, these technologies offer significant opportunities for advancing animal health monitoring, farm management, and individual phenotyping.

Discovering CO Adsorption and Desorption Pathways from Chemical Reaction Neural Network Modeling of Transient Kinetics Spectroscopy

We demonstrate a data-driven approach to interpreting surface reactions by combining time-resolved gas pulsing infrared spectroscopy with chemical reaction neural networks (CRNNs). Using CO adsorption and desorption on Pd(111) at 460-490 K as a model system, we show how transient kinetic data can reveal detailed reaction mechanisms. Starting with a simple one-species model, we systematically evaluate increasingly complex mechanisms involving hollow and bridge site adsorption. Despite the similar goodness of fit to the same experimental absorbance data, our models predict distinct coverage dynamics for different adsorption sites. Through analysis of spectral peak stability and predicted dynamics, we identify a mechanism in which CO primarily adsorbs on bridge sites followed by rapid conversion to hollow sites as being the most physically consistent with experimental observations. This work provides a framework for extracting mechanistic insights from limited experimental data, demonstrating how machine learning can bridge the gap between transient kinetic measurements and a molecular-level understanding of surface reactions.

Integrated deep eutectic system enrichment and AI-assisted high-throughput visual detection for Hg2+ in environmental samples

INTRODUCTION

Mercury ion (Hg2+), a prevalent heavy metal, is commonly found in environmental soils and waters. Its interaction with sulfhydryl groups in proteins and lipids can cause oxidative stress and disruption of calcium homeostasis. These lead to severe health issues, including digestive, nervous, and immune system damage. Conventional Hg2+ detection methods, such as ICP-MS and AAS, require complex procedures and bulky instruments, limiting their applicability for real-time, on-site analysis. Recently, AI-assisted detection methods have emerged as promising solutions, offering portability and rapid detection capabilities. Deep eutectic solvents (DESs), and in particularly hydrophobic DESs (HDESs), provide an environmentally friendly alternative for the enrichment and detection metal ions.

OBJECTIVES

This study aims to develop a portable, cost-effective, and environmentally friendly colorimetric sensing platform based on a silver nanoparticles hydrophobic deep eutectic system (AgNPs-HDES) for Hg2+ enrichment and detection.

METHODS

AgNPs-HDES was synthesized using silver nanoparticle-containing ethylene glycol (AgNPs-EG) as the hydrogen bond donor. Electrostatic potential maps (ESP) and density functional theory (DFT) were employed to elucidate its synthesis and enrichment mechanisms. Smartphone-based image acquisition combined with YOLOv8-based AI software enabled high-throughput colorimetric analysis for Hg2+ detection.

RESULTS

A progressive color change from brownish-yellow to colorless was observed with increasing Hg2+ concentration, thereby eliminating hydrophilic interference and improving sensitivity. The AgNPs-HDES platform demonstrated a linear detection range of 1-40 μmol·L-1 (R2 = 0.9889) and a detection limit of 0.23 μmol·L-1. Recovery rates in real samples, including lake water, soil, seawater and industrial sewage, ranged from 90.3% to 123%.

CONCLUSION

The established platform enables portable, rapid, and highly accurate Hg2+ detection across multiple environmental samples simultaneously. This AI-assisted, high-throughput detection system presents a valuable tool for environmental monitoring and pollutant tracking.

Examining the development, effectiveness, and limitations of computer-aided diagnosis systems for retained surgical items detection: a systematic review

Retained surgical items (RSIs) can lead to severe complications, and infections, with morbidity rates up to 84.32%. Computer-aided detection (CAD) systems offer potential advancement in enhancing the detection of RSIs. This systematic review aims to summarise the characteristics of CAD systems developed for the detection of RSIs, evaluate their development, effectiveness, and limitations, and propose opportunities for enhancement. The systematic review adheres to Preferred Reporting Items for Systematic Reviews and Meta-Analysis 2020 guidelines. Studies that have developed and evaluated CAD systems for identifying RSIs were eligible for inclusion. Five electronic databases were searched from inception to March 2023 and eleven studies were found eligible. The sensitivity of CAD systems ranges from 0.61 to 1 and specificity varied between 0.73 and 1. Most studies utilised synthesised RSI radiographs for developing CAD systems which raises generalisability concerns. Moreover, deep learning-based CAD systems did not incorporate explainable artificial intelligence techniques to ensure decision transparency.

A Risk Prediction Model (CMC-AKIX) for Postoperative Acute Kidney Injury Using Machine Learning: Algorithm Development and Validation

BACKGROUND

Postoperative acute kidney injury (AKI) is a significant risk associated with surgeries under general anesthesia, often leading to increased mortality and morbidity. Existing predictive models for postoperative AKI are usually limited to specific surgical areas or require external validation.

OBJECTIVE

We proposed to build a prediction model for postoperative AKI using several machine learning methods.

METHODS

We conducted a retrospective cohort analysis of noncardiac surgeries from 2009 to 2019 at seven university hospitals in South Korea. We evaluated six machine learning models: deep neural network, logistic regression, decision tree, random forest, light gradient boosting machine, and naïve Bayes for predicting postoperative AKI, defined as a significant increase in serum creatinine or the initiation of renal replacement therapy within 30 days after surgery. The performance of the models was analyzed using the area under the curve (AUC) of the receiver operating characteristic curve, accuracy, precision, sensitivity (recall), specificity, and F1-score.

RESULTS

Among the 239,267 surgeries analyzed, 7935 cases of postoperative AKI were identified. The models, using 38 preoperative predictors, showed that deep neural network (AUC=0.832), light gradient boosting machine (AUC=0.836), and logistic regression (AUC=0.825) demonstrated superior performance in predicting AKI risk. The deep neural network model was then developed into a user-friendly website for clinical use.

CONCLUSIONS

Our study introduces a robust, high-performance AKI risk prediction system that is applicable in clinical settings using preoperative data. This model's integration into a user-friendly website enhances its clinical utility, offering a significant step forward in personalized patient care and risk management.

A computed tomography-based deep learning radiomics model for predicting the gender-age-physiology stage of patients with connective tissue disease-associated interstitial lung disease

OBJECTIVES

To explore the feasibility of using a diagnostic model constructed with deep learning-radiomics (DLR) features extracted from chest computed tomography (CT) images to predict the gender-age-physiology (GAP) stage of patients with connective tissue disease-associated interstitial lung disease (CTD-ILD).

MATERIALS AND METHODS

The data of 264 CTD-ILD patients were retrospectively collected. GAP Stage I, II, III patients are 195, 56, 13 cases respectively. The latter two stages were combined into one group. The patients were randomized into a training set and a validation set. Single-input models were separately constructed using the selected radiomics and DL features, while DLR model was constructed from both sets of features. For all models, the support vector machine (SVM) and logistic regression (LR) algorithms were used for construction. The nomogram models were generated by integrating age, gender, and DLR features.

RESULTS

The DLR model outperformed the radiomics and DL models in both the training set and the validation set. The predictive performance of the DLR model based on the LR algorithm was the best among all the feature-based models (AUC = 0.923). The comprehensive models had even greater performance in predicting the GAP stage of CTD-ILD patients. The comprehensive model using the SVM algorithm had the best performance of the two models (AUC = 0.951).

CONCLUSION

The DLR model extracted from CT images can assist in the clinical prediction of the GAP stage of CTD-ILD patients. A nomogram showed even greater performance in predicting the GAP stage of CTD-ILD patients.

Coding schemes in neural networks learning classification tasks

Neural networks posses the crucial ability to generate meaningful representations of task-dependent features. Indeed, with appropriate scaling, supervised learning in neural networks can result in strong, task-dependent feature learning. However, the nature of the emergent representations is still unclear. To understand the effect of learning on representations, we investigate fully-connected, wide neural networks learning classification tasks using the Bayesian framework where learning shapes the posterior distribution of the network weights. Consistent with previous findings, our analysis of the feature learning regime (also known as 'non-lazy' regime) shows that the networks acquire strong, data-dependent features, denoted as coding schemes, where neuronal responses to each input are dominated by its class membership. Surprisingly, the nature of the coding schemes depends crucially on the neuronal nonlinearity. In linear networks, an analog coding scheme of the task emerges; in nonlinear networks, strong spontaneous symmetry breaking leads to either redundant or sparse coding schemes. Our findings highlight how network properties such as scaling of weights and neuronal nonlinearity can profoundly influence the emergent representations.

Reinforcement learning for healthcare operations management: methodological framework, recent developments, and future research directions

With the advancement in computing power and data science techniques, reinforcement learning (RL) has emerged as a powerful tool for decision-making problems in complex systems. In recent years, the research on RL for healthcare operations has grown rapidly. Especially during the COVID-19 pandemic, RL has played a critical role in optimizing decisions with greater degrees of uncertainty. RL for healthcare applications has been an exciting topic across multiple disciplines, including operations research, operations management, healthcare systems engineering, and data science. This review paper first provides a tutorial on the overall framework of RL, including its key components, training models, and approximators. Then, we present the recent advances of RL in the domain of healthcare operations management (HOM) and analyze the current trends. Our paper concludes by presenting existing challenges and future directions for RL in HOM.

Opportunities and challenges with artificial intelligence in allergy and immunology: a bibliometric study

Introduction

The fields of allergy and immunology are increasingly recognizing the transformative potential of artificial intelligence (AI). Its adoption is reshaping research directions, clinical practices, and healthcare systems. However, a systematic overview identifying current statuses, emerging trends, and future research hotspots is lacking.

Methods

This study applied bibliometric analysis methods to systematically evaluate the global research landscape of AI applications in allergy and immunology. Data from 3,883 articles published by 21,552 authors across 1,247 journals were collected and analyzed to identify leading contributors, prevalent research themes, and collaboration patterns.

Results

Analysis revealed that the USA and China are currently leading in research output and scientific impact in this domain. AI methodologies, especially machine learning (ML) and deep learning (DL), are predominantly applied in drug discovery and development, disease classification and prediction, immune response modeling, clinical decision support, diagnostics, healthcare system digitalization, and medical education. Emerging trends indicate significant movement toward personalized medical systems integration.

Discussion

The findings demonstrate the dynamic evolution of AI in allergy and immunology, highlighting the broadening scope from basic diagnostics to comprehensive personalized healthcare systems. Despite advancements, critical challenges persist, including technological limitations, ethical concerns, and regulatory frameworks that could potentially hinder further implementation and integration.

Conclusion

AI holds considerable promise for advancing allergy and immunology globally by enhancing healthcare precision, efficiency, and accessibility. Addressing existing technological, ethical, and regulatory challenges will be crucial to fully realizing its potential, ultimately improving global health outcomes and patient well-being.

Cross-institutional validation of a polar map-free 3D deep learning model for obstructive coronary artery disease prediction using myocardial perfusion imaging: insights into generalizability and bias

PURPOSE

Deep learning (DL) models for predicting obstructive coronary artery disease (CAD) using myocardial perfusion imaging (MPI) have shown potential for enhancing diagnostic accuracy. However, their ability to maintain consistent performance across institutions and demographics remains uncertain. This study aimed to investigate the generalizability and potential biases of an in-house MPI DL model between two hospital-based cohorts.

METHODS

We retrospectively included patients from two medical centers in Taiwan who underwent stress/redistribution thallium-201 MPI followed by invasive coronary angiography within 90 days as the reference standard. A polar map-free 3D DL model trained on 928 MPI images from one center to predict obstructive CAD was tested on internal (933 images) and external (3234 images from the other center) validation sets. Diagnostic performance, assessed using area under receiver operating characteristic curves (AUCs), was compared between the internal and external cohorts, demographic groups, and with the performance of stress total perfusion deficit (TPD).

RESULTS

The model showed significantly lower performance in the external cohort compared to the internal cohort in both patient-based (AUC: 0.713 vs. 0.813) and vessel-based (AUC: 0.733 vs. 0.782) analyses, but still outperformed stress TPD (all p < 0.001). The performance was lower in patients who underwent treadmill stress MPI in the internal cohort and in patients over 70 years old in the external cohort.

CONCLUSIONS

This study demonstrated adequate performance but also limitations in the generalizability of the DL-based MPI model, along with biases related to stress type and patient age. Thorough validation is essential before the clinical implementation of DL MPI models.

Non-Volatile Reconfigurable Four-Mode van der Waals Transistors and Transformable Logic Circuits

Emerging applications in data-intensive computing and circuit security demand logic circuits with high functional density, reconfigurability, and energy efficiency. Here, we demonstrate nonvolatile reconfigurable four-mode field-effect transistors (NVR4M-FETs) based on two-dimensional (2D) MoTe2 and CuInP2S6 (CIPS), offering both polarity switching and threshold voltage modulation. The device exploits the ferroelectric polarization of CIPS at the source/drain regions to achieve dynamic control over the transistor polarity, enabling transitions between n-type and p-type states through polarization-induced local electrostatic doping. Additionally, multilayer graphene floating gates are incorporated to modulate the threshold voltage, yielding four distinct nonvolatile operating modes: n-type logic, p-type logic, always-on memory, and always-off memory. Leveraging the four-mode property, the NVR4M-FET can function as a one-transistor-per-bit ternary content-addressable memory (TCAM). In addition, we demonstrate the construction of transformable logic gates with 14 distinct logic functions using two NVR4M-FETs and a reconfigurable half a dder/subtractor using three NVR4M-FETs integrated with load resistors. Furthermore, we show that a 2-input look-up table can be achieved with eight NVR4M-FETs compared to 12 transistors using reconfigurable transistors, highlighting the potential of NVR4M-FETs for high-density logic circuits. These results underscore the potential of NVR4M-FETs as essential building blocks for energy-efficient, in-memory computing, and secure hardware applications.

Artificial Intelligence-Assisted Multimode Microrobot Swarm Behaviors

Mimicking the swarm behaviors in nature, the microswarm has shown dynamic transformations and flexible assemblies in complex physiological environments, garnering increasing attention for its potential medical applications. However, because of the complexity of swarm behaviors and the corresponding influencing factors, achieving controllability, stability, and diversity of an artificial microswarm remains challenging. Here, a physically assisted artificial intelligence analysis framework was employed to predict the multimode swarm behaviors of a magnetic microswarm. By modulating 12 different parameters of a programmable magnetic field, we obtained various swarm patterns, including liquid, rod, network, ribbon, flocculence, and vortex. A physical model was developed to simulate the programmable 3D magnetic field and the corresponding collective behaviors. Explainable artificial intelligence analysis uncovered the relationship between control parameters and magnetic swarm patterns, achieving a prediction accuracy of 83.87% for pattern classification. Our stability analysis revealed that rod and vortex patterns exhibited higher stability, making them ideal for precise manipulation tasks. Leveraging this framework, we demonstrated environmentally adaptive swarm navigation through complex channels and swarm hunting of specific targets. This study could not only advance the understanding of microswarm control but also provide a strategy for targeted delivery and micromanipulation in potential clinical applications.

The role of AI for MRI-analysis in multiple sclerosis-A brief overview

Magnetic resonance imaging (MRI) has played a crucial role in the diagnosis, monitoring and treatment optimization of multiple sclerosis (MS). It is an essential component of current diagnostic criteria for its ability to non-invasively visualize both lesional and non-lesional pathology. Nevertheless, modern day usage of MRI in the clinic is limited by lengthy protocols, error-prone procedures for identifying disease markers (e.g., lesions), and the limited predictive value of existing imaging biomarkers for key disability outcomes. Recent advances in artificial intelligence (AI) have underscored the potential for AI to not only improve, but also transform how MRI is being used in MS. In this short review, we explore the role of AI in MS applications that span the entire life-cycle of an MRI image, from data collection, to lesion segmentation, detection, and volumetry, and finally to downstream clinical and scientific tasks. We conclude with a discussion on promising future directions.

Rapid COD Sensing in Complex Surface Water Using Physicochemical-Informed Spectral Transformer with UV-Vis-SWNIR Spectroscopy

Water, as a finite and vital resource, necessitates water quality monitoring to ensure its sustainable use. A key aspect of this process is the accurate measurement of critical parameters such as chemical oxygen demand (COD). However, current spectroscopic methods struggle with accurately and consistently measuring COD in large-scale, complex water environments due to an insufficient understanding of water spectra and limited generalizability. To address these limitations, we introduce the physicochemical-informed spectral Transformer (PIST) model, combined with ultraviolet-visible-shortwave-near-infrared (UV-vis-SWNIR) spectroscopy for water quality sensing. To the best of our knowledge, this is the first approach to combine Transformer with spectroscopy for water quality sensing. PIST integrates a physicochemical-informed block to incorporate existing physical and chemical information into the spectral encoding for domain adaptation, along with a feature embedding block for comprehensive spectral features extraction. We validated PIST using an actual surface water spectral data set with extensive geographic coverage including the Yangtze River and Poyang Lake. PIST demonstrated notable performance in COD sensing within complex water environments, achieving an impressive R2 value of 0.9008 and reducing root mean squared error (RMSE) by 45.20% and 29.38% compared to benchmark models such as support vector regression (SVR) and convolutional neural network (CNN). These results emphasize PIST's accuracy and generalizability, marking a significant advancement in multidisciplinary approaches that combine spectroscopy with deep learning for rapid water quality sensing.

Automated Refinement of Property-Specific Polarizable Gaussian Multipole Water Models Using Bayesian Black-Box Optimization

The critical importance of water in sustaining life highlights the need for accurate water models in computer simulations, aiming to mimic biochemical processes experimentally. The polarizable Gaussian multipole (pGM) model, recently introduced for biomolecular simulations, improves the handling of complex biomolecular interactions. As an integral part of our initial exploration, we examined a minimalist fixed geometry three-center pGM water model using ab initio quantum mechanical calculations of water oligomers. However, our final model development was based on liquid-phase water properties, leveraging automated machine learning (AutoML) techniques for optimization. This allows the development of a framework to refine both van der Waals and electrostatic parameters of the pGM model, aiming to accurately reproduce specific properties such as the oxygen-oxygen radial distribution function, density, and dipole moment, all at 298 K and 1.0 bar pressure. The efficacy of the optimized three-center pGM water model, pGM3P-25, was assessed through simulations of a water box of 512 water molecules, showcasing marked enhancements in both accuracy and practical utility. Notably, the model accurately reproduces thermodynamic properties not explicitly included in training while significantly reducing the time and human effort required for optimization. It was found that pGM3P-25 can reproduce temperature-dependent properties such as density, self-diffusion constants, heat capacity, second virial coefficient, and dielectric constant, which are important in biomolecular simulations. This study underscores the potential of AutoML-driven frameworks to streamline parameter refinement for molecular dynamics simulations, paving the way for broader applications in computational chemistry and beyond.

Onet: Twin U-Net Architecture for Unsupervised Binary Semantic Segmentation in Radar and Remote Sensing Images

Segmenting objects from cluttered backgrounds in single-channel images, such as marine radar echoes, medical images, and remote sensing images, poses significant challenges due to limited texture, color information, and diverse target types. This paper proposes a novel solution: the Onet, an O-shaped assembly of twin U-Net deep neural networks, designed for unsupervised binary semantic segmentation. The Onet, trained with an intensity-complementary image pair and without the need for annotated labels, maximizes the Jensen-Shannon divergence (JSD) between the densely localized features and the class probability maps. By leveraging the symmetry of U-Net, Onet subtly strengthens the dependence between dense local features, global features, and class probability maps during the training process. The design of the complementary input pair aligns with the theoretical requirement that optimizing JSD needs the class probability of negative samples to accurately estimate the marginal distribution. Compared to the current leading unsupervised segmentation methods, the Onet demonstrates superior performance in target segmentation in marine radar frames and cloud segmentation in remote sensing images. Notably, we found that Onet's foreground prediction significantly enhances the signal-to-noise ratio (SNR) of targets amidst marine radar clutter. Onet's source code is publicly accessible at https://github.com/joeyee/Onet.

Evaluating the validity and consistency of artificial intelligence chatbots in responding to patients' frequently asked questions in prosthodontics

STATEMENT OF PROBLEM

Healthcare-related information provided by artificial intelligence (AI) chatbots may pose challenges such as inaccuracies, lack of empathy, biases, over-reliance, limited scope, and ethical concerns.

PURPOSE

The purpose of this study was to evaluate and compare the validity and consistency of responses to prosthodontics-related frequently asked questions (FAQ) generated by 4 different chatbot systems.

MATERIAL AND METHODS

Four prosthodontics domains were evaluated: implant, fixed prosthodontics, complete denture (CD), and removable partial denture (RPD). Within each domain, 10 questions were prepared by full-time prosthodontic faculty members, and 10 questions were generated by GPT-3.5, representing its top frequently asked questions in each domain. The validity and consistency of responses provided by 4 chatbots: GPT-3.5, GPT-4, Gemini, and Bing were evaluated. The chi-squared test with the Yates correction was used to compare the validity of responses between different chatbots (α=.05). The Cronbach alpha was calculated for 3 sets of responses collected in the morning, afternoon, and evening to evaluate the consistency of the responses.

RESULTS

According to the low threshold validity test, the chatbots' answers to ChatGPT's implant-related, ChatGPT's RPD-related, and prosthodontists' CD-related FAQs were statistically different (P<.001, P<.001, and P=.004, respectively), with Bing being the lowest. At the high threshold validity test, the chatbots' answers to ChatGPT's implant-related and RPD-related FAQs and ChatGPT's and prosthodontists' fixed prosthetics-related and CD-related FAQs were statistically different (P<.001, P<.001, P=.004, P=.002, and P=.003, respectively), with Bing being the lowest. Overall, all 4 chatbots demonstrated lower validity at the high threshold than the low threshold. Bing, Gemini, and ChatGPT-4 chatbots displayed an acceptable level of consistency, while ChatGPT-3.5 did not.

CONCLUSIONS

Currently, AI chatbots show limitations in delivering answers to patients' prosthodontic-related FAQs with high validity and consistency.

Perturbation of mRNA splicing in liver cancer: insights, opportunities and challenges

Perturbation of mRNA splicing is commonly observed in human cancers and plays a role in various aspects of cancer hallmarks. Understanding the mechanisms and functions of alternative splicing (AS) not only enables us to explore the complex regulatory network involved in tumour initiation and progression but also reveals potential for RNA-based cancer treatment strategies. This review provides a comprehensive summary of the significance of AS in liver cancer, covering the regulatory mechanisms, cancer-related AS events, abnormal splicing regulators, as well as the interplay between AS and post-transcriptional and post-translational regulations. We present the current bioinformatic approaches and databases to detect and analyse AS in cancer, and discuss the implications and perspectives of AS in the treatment of liver cancer.

Analyzing the performance of biomedical time-series segmentation with electrophysiology data

Accurate segmentation of biomedical time-series, such as intracardiac electrograms, is vital for understanding physiological states and supporting clinical interventions. Traditional rule-based and feature engineering approaches often struggle with complex clinical patterns and noise. Recent deep learning advancements offer solutions, showing various benefits and drawbacks in segmentation tasks. This study evaluates five segmentation algorithms, from traditional rule-based methods to advanced deep learning models, using a unique clinical dataset of intracardiac signals from 100 patients. We compared a rule-based method, a support vector machine (SVM), fully convolutional semantic neural network (UNet), region proposal network (Faster R-CNN), and recurrent neural network for electrocardiographic signals (DENS-ECG). Notably, Faster R-CNN has never been applied to 1D signals segmentation before. Each model underwent Bayesian optimization to minimize hyperparameter bias. Results indicated that deep learning models outperformed traditional methods, with UNet achieving the highest segmentation score of 88.9 % (root mean square errors for onset and offset of 8.43 ms and 7.49 ms), closely followed by DENS-ECG at 87.8 %. Faster R-CNN and SVM showed moderate performance, while the rule-based method had the lowest accuracy (77.7 %). UNet and DENS-ECG excelled in capturing detailed features and handling noise, highlighting their potential for clinical application. Despite greater computational demands, their superior performance and diagnostic potential support further exploration in biomedical time-series analysis.

Optimization on multifractal loss landscapes explains a diverse range of geometrical and dynamical properties of deep learning

Gradient descent and its variants are foundational in solving optimization problems across many disciplines. In deep learning, these optimizers demonstrate a remarkable ability to dynamically navigate complex loss landscapes, ultimately converging to solutions that generalize well. To elucidate the mechanism underlying this ability, we introduce a theoretical framework that models the complexities of loss landscapes as multifractal. Our model unifies and explains a broad range of realistic geometrical signatures of loss landscapes, including clustered degenerate minima, multiscale structure, and rich optimization dynamics in deep neural networks, such as the edge of stability, non-stationary anomalous diffusion, and the extended edge of chaos without requiring fine-tuning parameters. We further develop a fractional diffusion theory to illustrate how these optimization dynamics, coupled with multifractal structure, effectively guide optimizers toward smooth solution spaces housing flatter minima, thus enhancing generalization. Our findings suggest that the complexities of loss landscapes do not hinder optimization; rather, they facilitate the process. This perspective not only has important implications for understanding deep learning but also extends potential applicability to other disciplines where optimization unfolds on complex landscapes.

ECG Sensor Design Assessment with Variational Autoencoder-Based Digital Watermarking

Designing an ECG sensor circuit requires a comprehensive approach to detect, amplify, filter, and condition the weak electrical signals produced by the heart. To evaluate sensor performance under realistic conditions, diverse ECG signals with embedded watermarks are generated, enabling an assessment of how effectively the sensor and its signal-conditioning circuitry handle these modified signals. A Variational Autoencoder (VAE) framework is employed to generate the watermarked ECG signals, addressing critical concerns in the digital era, such as data security, authenticity, and copyright protection. Three watermarking strategies are examined in this study: embedding watermarks in the mean (μ) of the VAE's latent space, embedding them through the latent variable (z), and using post-reconstruction watermarking in the frequency domain. Experimental results demonstrate that watermarking applied through the mean (μ) and in the frequency domain achieves a low Mean Squared Error (MSE) while maintaining stable signal fidelity across varying watermark strengths (α), latent space dimensions, and noise levels. These findings indicate that the mean (μ) and frequency domain methods offer robust performance and are minimally affected by changes in these parameters, making them particularly suitable for preserving ECG signal quality. By contrasting these methods, this study provides insights into selecting the most appropriate watermarking technique for ECG sensor applications. Incorporating watermarking into sensor design not only strengthens data security and authenticity but also supports reliable signal acquisition in modern healthcare environments. Overall, the results underscore the effectiveness of combining VAEs with watermarking strategies to produce high-fidelity, resilient ECG signals for both sensor performance evaluation and the protection of digital content.

Adaptive Whole-Brain Dynamics Predictive Method: Relevancy to Mental Disorders

The Hopf whole-brain model, based on structural connectivity, overcomes limitations of traditional structural or functional connectivity-focused methods by incorporating heterogeneity parameters, quantifying dynamic brain characteristics in healthy and diseased states. Traditional parameter fitting techniques lack precision, restricting broader use. To address this, we validated parameter fitting methods using simulated networks and synthetic models, introducing improvements such as individual-specific initialization and optimized gradient descent, which reduced individual data loss. We also developed an approximate loss function and gradient adjustment mechanism, enhancing parameter fitting accuracy and stability. Applying this refined method to datasets for major depressive disorder (MDD) and autism spectrum disorder (ASD), we identified differences in brain regions between patients and healthy controls, explaining related anomalies. This rigorous validation is crucial for clinical application, paving the way for precise neuropathological identification and novel treatments in neuropsychiatric research, demonstrating substantial potential in clinical neurology.

Microplastics migration mechanisms in high-erosion watersheds under climate warming

Understanding Microplastics (MPs) migration in small watersheds is crucial for pollution management, but progress has been hindered by limited long-term data and modeling approaches. This study investigated three watersheds on the Qinghai-Xizang Plateau, each with distinct land uses (grassland, cropland, urban). Using 15 years of sediment data, a novel MPs migration model was developed with machine learning (RF, SHAP, DNN), achieving exceptionally high accuracy in source tracing (R² = 0.93) and pathway analysis (R² = 0.97). The results revealed that under conditions of sediment thickness < 6.5 cm (Scenario 1), MPs primarily migrated from cropland to sediment driven by southerly winds and surface runoff, with an MPs migration flux (nMPs) of 2.09 × 10⁴ items/m² and an MPs migration content (ρMPs) of 372.99 items/kg. For sediment thicknesses between 6.5 and 10 cm (Scenario 2), contributions from both cropland and grassland led to a 127.6 % increase in nMPs. When sediment thickness exceeds 10 cm (Scenario 3), grassland contributions become more significant, leading to a 284.52 % increase in nMPs and a 21.31 % reduction in ρMPs. Between 2000 and 2020, climate warming significantly intensified extreme precipitation (p < 0.05), shifting MPs migration patterns toward Scenario 3. Future projections (2030-2100) under a high-emission scenario indicated MPs migration and contents would increase by 111.64 % and 4.29 items/kg per decade, respectively. Under a low-emission scenario, migration would decrease by 1.48 % per decade, while MPs content would slightly increase by 1.05 items/kg per decade. This study provides a robust modeling framework for understanding MPs migration and supporting sustainable pollution management.

Retinal OCT image segmentation with deep learning: A review of advances, datasets, and evaluation metrics

Optical coherence tomography (OCT) is a widely used imaging technology in ophthalmic clinical practice, providing non-invasive access to high-resolution retinal images. Segmentation of anatomical structures and pathological lesions in retinal OCT images, directly impacts clinical decisions. While commercial OCT devices segment multiple retinal layers in healthy eyes, their performance degrades severely under pathological conditions. In recent years, the rapid advancements in deep learning have significantly driven research in OCT image segmentation. This review provides a comprehensive overview of the latest developments in deep learning-based segmentation methods for retinal OCT images. Additionally, it summarizes the medical significance, publicly available datasets, and commonly used evaluation metrics in this field. The review also discusses the current challenges faced by the research community and highlights potential future directions.

Predicting Atlantic and Benguela Niño events with deep learning

Atlantic and Benguela Niño events substantially affect the tropical Atlantic region, with far-reaching consequences on local marine ecosystems, African climates, and El Niño Southern Oscillation. While accurate forecasts of these events are invaluable, state-of-the-art dynamic forecasting systems have shown limited predictive capabilities. Thus, the extent to which the tropical Atlantic variability is predictable remains an open question. This study explores the potential of deep learning in this context. Using a simple convolutional neural network architecture, we show that Atlantic/Benguela Niños can be predicted up to 3 to 4 months ahead. Our model excels in forecasting peak-season events with remarkable accuracy extending lead time to 5 months. Detailed analysis reveals our model's ability to exploit known physical precursors, such as long-wave ocean dynamics, for accurate predictions of these events. This study challenges the perception that the tropical Atlantic is unpredictable and highlights deep learning's potential to advance our understanding and forecasting of critical climate events.

Deep cross entropy fusion for pulmonary nodule classification based on ultrasound Imagery

Introduction

Accurate differentiation of benign and malignant pulmonary nodules in ultrasound remains a clinical challenge due to insufficient diagnostic precision. We propose the Deep Cross-Entropy Fusion (DCEF) model to enhance classification accuracy.

Methods

A retrospective dataset of 135 patients (27 benign, 68 malignant training; 11 benign, 29 malignant testing) was analyzed. Manually annotated ultrasound ROIs were preprocessed and input into DCEF, which integrates ResNet, DenseNet, VGG, and InceptionV3 via entropy-based fusion. Performance was evaluated using AUC, accuracy, sensitivity, specificity, precision, and F1-score.

Results

DCEF achieved an AUC of 0.873 (training) and 0.792 (testing), outperforming traditional methods. Test metrics included 71.5% accuracy, 70.69% sensitivity, 70.58% specificity, 72.55% precision, and 71.13% F1-score, demonstrating robust diagnostic capability.

Discussion

DCEF's multi-architecture fusion enhances diagnostic reliability for ultrasound-based nodule assessment. While promising, validation in larger multi-center cohorts is needed to address single-center data limitations. Future work will explore next-generation architectures and multi-modal integration.

MSFSegNet: A multi-scale feature fusion model for instance segmentation in adult liver ultrasound images

BACKGROUND AND OBJECTIVE

Liver diseases often remain undetected until advanced stages due to the lack of early symptoms. Two-dimensional ultrasonography is a key diagnostic tool, but manual segmentation of liver and its accessory structures (LAS) is time-consuming and prone to human error. To address this, we propose MSFSegNet, a novel instance segmentation model designed for adult liver ultrasound images.

METHODS

MSFSegNet integrates a multi-scale feature fusion network (CCMC), an adaptive downsampling method (ODConv), and the Convolutional Block Attention Module (CBAM) to enhance segmentation accuracy, particularly for small anatomical structures.

RESULTS

MSFSegNet achieves superior performance with Precision, Recall, and mAP@0.5 of 94.4 %, 91.8 %, and 95.7 % in position evaluation, and 93.9 %, 91.3 %, and 94.8 % in segmentation tasks, outperforming existing methods by a significant margin.

CONCLUSIONS

The proposed model demonstrates significant potential for computer-aided diagnosis in liver ultrasound imaging, offering a robust solution for accurate segmentation of LAS. Future work will focus on optimizing computational efficiency and expanding the model's applicability to pathological cases.