Training Dataset Research Articles

Self-supervised learning improves robustness of deep learning lung tumor segmentation models to CT imaging differences.

Self-supervised learning (SSL) is an approach to extract useful feature representations from unlabeled data, and enable fine-tuning on downstream tasks with limited labeled examples. Self-pretraining is a SSL approach that uses curated downstream task dataset for both pretraining and fine-tuning. Availability of large, diverse, and uncurated public medical image sets presents the opportunity to potentially create foundation models by applying SSL in the "wild" that are robust to imaging variations. However, the benefit of wild- versus self-pretraining has not been studied for medical imageanalysis. Compare robustness of wild versus self-pretrained models created using convolutional neural network (CNN) and transformer (vision transformer [ViT] and hierarchical shifted window [Swin]) models for non-small cell lung cancer (NSCLC) segmentation from 3D computed tomography (CT) scans. CNN, ViT, and Swin models were wild-pretrained using unlabeled 10,412 3D CTs sourced from the cancer imaging archive and internal datasets. Self-pretraining was applied to same networks using a curated public downstream task dataset (n = 377) of patients with NSCLC. Pretext tasks introduced in self-distilled masked image transformer were used for both pretraining approaches. All models were fine-tuned to segment NSCLC (n = 377 training dataset) and tested on two separate datasets containing early (public n = 156) and advanced stage (internal n = 196) NSCLC. Models were evaluated in terms of: (a) accuracy, (b) robustness to image differences from contrast, slice thickness, and reconstruction kernels, and (c) impact of pretext tasks for pretraining. Feature reuse was evaluated using centered kernel alignment. Wild-pretrained Swin models resulted in higher feature reuse at earlier level layers and increased feature differentiation close to output. Wild-pretrained Swin outperformed self-pretrained models for analyzed imaging acquisitions. Neither ViT nor CNN showed a clear benefit of wild-pretraining compared to self-pretraining. Masked image prediction pretext task that forces networks to learn the local structure resulted in higher accuracy compared to contrastive task that models global image information. Wild-pretrained Swin networks were more robust to analyzed CT imaging differences for lung tumor segmentation than self-pretrained methods. ViT and CNN models did not show a clear benefit for wild-pretraining overself-pretraining.

Neural operators for hydrodynamic modeling of underground gas storage facilities

Objectives. Much of the research in deep learning has focused on studying mappings between finite-dimensional spaces. While hydrodynamic processes of gas filtration in underground storage facilities can be described by partial differential equations (PDE), the requirement to study the mappings between functional spaces of infinite dimension distinguishes this problem from those solved using traditional mapping approaches. One of the most promising approaches involves the construction of neural operators, i.e., a generalization of neural networks to approximate mappings between functional spaces. The purpose of the work is to develop a neural operator to speed up calculations involved in hydrodynamic modeling of underground gas storages (UGS) to an acceptable degree of accuracy.Methods. In this work, a modified Fourier neural operator was built and trained for hydrodynamic modeling of gas filtration processes in underground gas storages.Results. The described method is shown to be capable of successful application to problems of three-dimensional gas filtration in a Cartesian coordinate system at objects with many wells. Despite the use of the fast Fourier transform algorithm in the architecture, the developed model is also effective for modeling objects having a nonuniform grid and complex geometry. As demonstrated not only on the test set, but also on artificially generated scenarios with significant changes made to the structure of the modeled object, the neural operator does not require a large training dataset size to achieve high accuracy of approximation of PDE solutions. A trained neural operator can simulate a given scenario in a fraction of a second, which is at least 106 times faster than a traditional numerical simulator.Conclusions. The constructed and trained neural operator demonstrated efficient hydrodynamic modeling of underground gas storages. The resulting algorithm reproduces adequate solutions even in the case of significant changes in the modeled object that had not occurred during the training process. The model can be recommended for use in planning and decision-making purposes regarding various aspects of UGS operation, such as optimal control of gas wells, pressure control, and management of gas reserves.

Comparing human-labeled and AI-labeled speech datasets for TTS

As the output quality of neural networks in the fields of automatic speech recognition (ASR) and text-to-speech (TTS) continues to improve, new opportunities are becoming available to train models in a weakly supervised fashion, thus minimizing the manual effort required to annotate new audio data for supervised training. While weak supervision has recently shown very promising results in the domain of ASR, speech synthesis has not yet been thoroughly investigated regarding this technique despite requiring the equivalent training dataset structure of aligned audio-transcript pairs. In this work, we compare the performance of TTS models trained using a well-curated and manually labeled training dataset to others trained on the same audio data with text labels generated using both grapheme- and phoneme-based ASR models. Phoneme-based approaches seem especially promising, since even for wrongly predicted phonemes, the resulting word is more likely to sound similar to the originally spoken word than for grapheme-based predictions. For evaluation and ranking, we generate synthesized audio outputs from all previously trained models using input texts sourced from a selection of speech recognition datasets covering a wide range of application domains. These synthesized outputs are subsequently fed into multiple state-of-the-art ASR models with their output text predictions being compared to the initial TTS model input texts. This comparison enables an objective assessment of the intelligibility of the audio outputs from all TTS models, by utilizing metrics like word error rate and character error rate. Our results not only show that models trained on data generated with weak supervision achieve comparable quality to models trained on manually labeled datasets, but can outperform the latter, even for small, well-curated speech datasets. These findings suggest that the future creation of labeled datasets for supervised training of TTS models may not require any manual annotation but can be fully automated.

Identification of potential biomarkers from amino acid transporter in the activation of hepatic stellate cells via bioinformatics

BackgroundThe etiopathogenesis of hepatic stellate cells (HSC) activation has yet to be completely comprehended, and there has been broad concern about the interplay between amino acid transporter and cell proliferation. This study proposed exploring the molecular mechanism from amino acid transport-related genes in HSC activation by bioinformatic methods, seeking to identify the potentially crucial biomarkers.MethodsGSE68000, the mRNA expression profile dataset of activated HSC, was applied as the training dataset, and GSE67664 as the validation dataset. Differently expressed amino acid transport-related genes (DEAATGs), GO, DO, and KEGG analyses were utilized. We applied the protein-protein interaction analysis and machine learning of LASSO and random forests to identify the target genes. Moreover, single-gene GESA was executed to investigate the potential functions of target genes via the KEGG pathway terms. Then, a ceRNA network and a drug-gene interaction network were constructed. Ultimately, correlation analysis was explored between target genes and collagen alpha I (COL1A), alpha-smooth muscle actin (α-SMA), and immune checkpoints.ResultsWe identified 15 DEAATGs, whose enrichment analyses indicated that they were primarily enriched in the transport and metabolic process of amino acids. Moreover, two target genes (SLC7A5 and SLC1A5) were recognized from the PPI network and machine learning, confirmed through the validation dataset. Then single-gene GESA analysis revealed that SLC7A5 and SLC1A5 had a significant positive correlation to ECM−receptor interaction, cell cycle, and TGF−β signaling pathway and negative association with retinol metabolism conversely. Furthermore, the mRNA expression of target genes was closely correlated with the COL1A and α-SMA, as well as immune checkpoints. Additionally, 12 potential therapeutic drugs were in the drug-gene interaction network, and the ceRNA network was constructed and visualized.ConclusionSLC7A5 and SLC1A5, with their relevant molecules, could be potentially vital biomarkers for the activation of HSC.

Application of artificial intelligence-based detection of furcation involvement in mandibular first molar using cone beam tomography images- a preliminary study

BackgroundRadiographs play a key role in diagnosis of periodontal diseases. Deep learning models have been explored for image analysis in periodontal diseases. However, there is lacuna of research in the deep learning model-based detection of furcation involvements [FI]. The objective of this study was to determine the accuracy of deep learning model in the detection of FI in axial CBCT images.MethodologyWe obtained initial dataset 285 axial CBCT images among which 143 were normal (without FI) and 142 were abnormal (with FI). Data augmentation technique was used to create 600(300 normal and 300 abnormal) images by using 200 images from the training dataset. Remaining 85(43 normal and 42 abnormal) images were kept for testing of model. ResNet101V2 with transfer learning was used employed for the analysis of images.ResultsTraining accuracy of model is 98%, valid accuracy is 97% and test accuracy is 91%. The precision and F1 score were 0.98 and 0.98 respectively. The Area under curve (AUC) was reported at 0.98. The test loss was reported at 0.2170.ConclusionThe deep learning model (ResNet101V2) can accurately detect the FI in axial CBCT images. However, since our study was preliminary in nature and carried out with relatively smaller dataset, a study with larger dataset will further confirm the accuracy of deep learning models.

Integrated explainable machine learning and multi-omics analysis for survival prediction in cancer with immunotherapy response.

To demonstrate the efficacy of machine learning models in predicting mortality in melanoma cancer, we developed an interpretability model for better understanding the survival prediction of cancer. To this end, the optimal features were identified, ten different machine learning models were utilized to predict mortality across various datasets. Then we have utilized the important features identified by those machines learning methods to construct a new model named NKECLR to forecast mortality of patient with cancer. To explicitly clarify the model's decision-making process and uncover novel findings, an interpretable technique incorporating machine learning and SHapley Additive exPlanations (SHAP), as well as LIME, has been employed, and four genes EPGN, PHF11, RBM34, and ZFP36 were identified from those machine learning(ML). The experimental analysis conducted on training and validation datasets demonstrated that the proposed model has a good performance com- pared to existing methods with AUC value 81.8%, and 79.3%, respectively. Moreover, when combined our NKECLR with PD-L1, PD-1, and CTLA-4 the AUC value was 83%0. Finally, these findings have been applied to comprehend the response of drugs and immunotherapy. Our research introduced an innovative predictive NKECLR model utilizing natural killer(NK) cell marker genes for cohorts with melanoma cancer. The NKECLR model can effectively predict the survival of melanoma cancer cohorts and treatment results, revealing distinct immune cell infiltration in the high-risk group.

NeCA: 3D Coronary Artery Tree Reconstruction from Two 2D Projections via Neural Implicit Representation

Cardiovascular diseases (CVDs) are the most common health threats worldwide. 2D X-ray invasive coronary angiography (ICA) remains the most widely adopted imaging modality for CVD assessment during real-time cardiac interventions. However, it is often difficult for the cardiologists to interpret the 3D geometry of coronary vessels based on 2D planes. Moreover, due to the radiation limit, often only two angiographic projections are acquired, providing limited information of the vessel geometry and necessitating 3D coronary tree reconstruction based only on two ICA projections. In this paper, we propose a self-supervised deep learning method called NeCA, which is based on neural implicit representation using the multiresolution hash encoder and differentiable cone-beam forward projector layer, in order to achieve 3D coronary artery tree reconstruction from two 2D projections. We validate our method using six different metrics on a dataset generated from coronary computed tomography angiography of right coronary artery and left anterior descending artery. The evaluation results demonstrate that our NeCA method, without requiring 3D ground truth for supervision or large datasets for training, achieves promising performance in both vessel topology and branch-connectivity preservation compared to the supervised deep learning model.

Reducing Uncertainty of Groundwater Redox Condition Predictions at National Scale, for Decision Making and Policy.

Understanding hydrogeochemical heterogeneity, associated with natural nitrate attenuation, is an integral part of implementing integrated land and water management on a regional or national scale. Redox conditions are a key indicator of naturally occurring denitrification in the groundwater environment, and often used to inform spatial planning and targeted regulation. This work describes the development of a statistical redox condition model for the groundwater environment at a national scale, using spatially variable physiochemical descriptors as predictors. The proposed approach builds on previous work, by complementing the available data with expert knowledge, in the form of synthetic data. Special care is given so that the synthetic data do not overfit and create further imbalances to the training dataset. The predictor dataset is further complemented by the results of a data driven model of the water table developed for this study, which is used both as a predictive parameter and a reference level for groundwater redox condition predictions at different depths. The developed model predicted the redox class for 84% of the samples in the out-of-bag datasets. We also propose an alternative approach for the communication of prediction uncertainty. We use the concept of a discriminate function to identify model classifications that may be ambiguous. Our results show a marked reduction in prediction uncertainty at shallow depths, with uncertainty in reduced environments decreasing from 76 to 12%, and overall uncertainty reduced by approximately 20%, though improvements at greater depths are less pronounced. We conclude that this approach can highlight robust model predictions that are defendable for decision making and can identify areas where monitoring or sampling efforts can be focused for improved outcomes.

Low dose threshold for measuring cardiac functional metrics using four-dimensional CT with deep learning.

Four-dimensional CT is increasingly used for functional cardiac imaging, including prognosis for conditions such as heart failure and post myocardial infarction. However, radiation dose from an acquisition spanning the full cardiac cycle remains a concern. This work investigates the possibility of dose reduction in 4DCT using deep learning (DL)-based segmentation techniques as an objective observer. A 3D residual U-Net was developed for segmentation of left ventricle (LV) myocardium and blood pool. Two networks were trained: Standard DL (trained with only standard-dose [SD] data) and Noise-Robust DL (additionally trained with low-dose data). The primary goal of the proposed DL methods is to serve as an unbiased and consistent observer for functional analysis performance. Functional cardiac metrics including ejection fraction (EF), global longitudinal strain (GLS), circumferential strain (CS), and wall thickness (WT), were measured for an external test set of 250 Cardiac CT volumes reconstructed at five different dose levels. Functional metrics obtained from DL segmentations of standard dose images matched well with those from expert manual analysis. Utilizing Standard-DL, absolute difference between DL-derived metrics obtained with standard dose data and 100mA (corresponding to ∼76±13% dose reduction) data was less than 0.8±1.0% for EF, GLS, and CS, and 5.6±6.7% for Average WT. Performance variation of Noise-Robust DL remained acceptable at even 50mA. We demonstrate that on average radiation dose can be reduced by a factor of 5 while introducing minimal changes to global functional metrics (especially EF, GLS, and CS). The robustness to reduced image quality can be further boosted by using emulated low-dose data in the DL training set.

Description and Modeling of Relevant Demographic and Laboratory Variables in a Large Oncology Cohort to Generate Virtual Populations

Background/Objectives: Pathophysiological variability in patients with cancer is associated with differences in responses to pharmacotherapy. In this work, we aimed to describe the demographic characteristics and hematological, biochemical, and coagulation variables in a large oncology cohort and to develop, optimize, and provide open access to modeling equations for the estimation of variables potentially relevant in pharmacokinetic modeling. Methods: Using data from 1793 patients with cancer, divided into training (n = 1259) and validation (n = 534) datasets, a modeling network was developed and used to simulate virtual oncology populations. All analyses were conducted in RStudio 4.3.2 Build 494. Results: The simulation network based on sex, age, biogeographic origin/ethnicity, and tumor type (fixed or primary factors) was successfully validated, able to predict age, height, weight, alpha-1-acid glycoprotein, albumin, hemoglobin, C-reactive protein and lactate dehydrogenase serum levels, platelet–lymphocyte and neutrophil–lymphocyte ratios, and hematocrit. This network was then successfully extrapolated to simulate the laboratory variables of eight oncology populations (n = 1200); only East Asians, Sub-Saharan Africans, Europeans, only males, females, patients with an ECOG performance status equal to 2, and only patients with pancreas cancer or ovarian cancer. Conclusions: this network constitutes a valuable tool to predict relevant characteristics/variables of patients with cancer, which may be useful in the evaluation and prediction of pharmacokinetics in virtual oncology populations, as well as for model-based optimization of oncology treatments.

Deepfake Detection Model Combining Texture Differences and Frequency Domain Information

In recent years, public security incidents caused by deepfake technology have occurred frequently around the world, which makes an efficient and accurate deepfake detection model crucial. The existing advanced methods use the manipulation features in the image to realize the binary classification of real and fake images by training complex neural network models. However, these models rely on a single manipulation feature, and the detection accuracy of these methods will be greatly reduced when the forgery technology or image quality of the training data set and the validation data set are different. Inspired by the existing work, we propose a two-stream collaborative learning framework that combines spatial texture differences and frequency information. The average difference convolution (ADC) is designed to extract the spatial texture difference information of the image, and the gray image frequency-aware decomposition (GFAD) is used to extract the artifact information of the image in the frequency domain. At the same time, the ViT idea is combined with cross attention mechanism for feature fusion to comprehensively mine forged features in forged images. Experimental results show that the proposed model has good detection effects on three benchmark data sets. In terms of cross-dataset evaluation, the AUC on Celeb-DF dataset reaches 82.86%, which is better than the existing advanced methods.

Selecting the Right Metric: A Detailed Study on Image Segmentation Evaluation

<p dir="ltr"><span>Image segmentation is critical role in various computer vision tasks, such as object recognition, medical imaging, autonomous driving, and document analysis. To evaluate the performance of segmentation algorithms, various metrics have been developed, each providing insights into different aspects of segmentation accuracy, object completeness, and boundary precision. This paper presents a comprehensive review of the key evaluation metrics used in image segmentation, including popular metrics such as Intersection over Union, Dice Coefficient, F1-score, and boundary-based metrics like average and maximum boundary distance. We explore the taxonomy of evaluation techniques, distinguishing between supervised and unsupervised methods, and discuss the strengths, limitations, and applications of each metric. Furthermore, we address the challenges associated with image segmentation, including handling noise, occlusions, illumination variation, and the scarcity of annotated datasets for training. By offering a detailed analysis of evaluation metrics, this paper aims to guide researchers and practitioners in selecting appropriate metrics for specific tasks and improving algorithm performance through informed comparisons and parameter tuning.</span></p><div><span><br /></span></div>

Can diffusion models capture extreme event statistics?

For many important problems it is essential to be able to accurately quantify the statistics of extremes for specific quantities of interest, such as extreme atmospheric weather events or ocean-related quantities. While there are many classical approaches to perform such modeling tasks, recent interest has been increasing in the usage of generative models trained on available data. Despite the sporadic success of such methods, it is not clear for what systems or datasets a system-agnostic generative AI tool is capable of generating previously ‘unseen’ extreme events in a manner that accurately extrapolates the tails for the observable of interest. Here, we propose an apriori criterion, which based on the geometry of the training dataset, it can predict whether a generative AI tool will be able to extrapolate the tails, i.e. generate previously unseen extreme events. The idea is to quantify whether existing extreme events lie in the interior of the dataset or its boundary. In the former case it is shown that generative AI tools can work in an ‘interpolation’ mode and generate new extreme events. On the other hand, if the topology of the dataset is such that extremes live in the boundary of the domain then the generative AI algorithm needs to operate in an extrapolation mode which does not lead to accurate results. We illustrate our findings on a specific class of Diffusion Models (DMs) called Denoising Diffusion Probabilistic Models (DDPMs) and we test on three datasets, a simple on-hyperball dataset following a Weibull distribution for the radii of the data points of dimensionality 2⋅103, a dataset sampled from the so-called Majda–McLaughlin–Tabak Wave Model (MMT), of dimensionality 8.1⋅103 and a dataset consisting of Lagrangian turbulence trajectories, of dimensionality 2⋅103.

Particulate matter estimation using satellite datasets: a machine learning approach.

In the present work, it is the first time an interpretable machine learning model has been developed for the estimation of Particulate Matter 10µm (PM10) concentrations over India using Aerosol Optical Depth (AOD) from two different satellites, i.e. INSAT-3D and Moderate Resolution Imaging Spectroradiometer (MODIS) for the period of 7years (2014 to 2020). Ground datasets of AOD are taken from the Aerosol Robotic Network (AERONET) for the validation of satellite-retrieved AOD. The observation of particulate matter (PM) data is acquired from the Central Pollution Control Board (CPCB) station across India. Analysis has been performed on a monthly basis for the given time period. The result shows that AOD products of MODIS exhibit good correlation with AERONET AOD whereas INSAT-3D AOD is not well correlated with AERONET AOD. However, after applying an error envelope and threshold-based filtering technique, we have found that INSAT-3D shows significant correlation with ground-level AOD with approximate correlation of 0.66 for Jaipur and 0.57 for Kanpur exhibiting almost similar performance as MODIS-derived AOD. Satellite AOD data together with ground PM concentration data is used to train the machine learning model (random forest) for the estimation of the PM distribution across India for the year 2020. An encouraging correlation of R-squared (R2) value 0.78 has been observed between the estimated and observed PM10 concentrations. The model demonstrates effective training, mitigating huge overestimation and underestimation. However, despite closely tracking the trends of estimated PM10 with observed PM10, few instances of overestimation persist. This suggests the need for an expanded training dataset to further refine and enhance the model's accuracy. Finally, the machine learning model used for PM10 estimation is found to be optimal for a calibrated satellite AOD product.

A feasibility study using quantitative and interpretable histological analyses of celiac disease for automated cell type and tissue area classification

Histological assessment is essential for the diagnosis and management of celiac disease. Current scoring systems, including modified Marsh (Marsh–Oberhuber) score, lack inter-pathologist agreement. To address this unmet need, we aimed to develop a fully automated, quantitative approach for histology characterisation of celiac disease. Convolutional neural network models were trained using pathologist annotations of hematoxylin and eosin-stained biopsies of celiac disease mucosa and normal duodenum to identify cells, tissue and artifact regions. Biopsies of duodenal mucosa of varying celiac disease severity, and normal duodenum were collected from a large central laboratory. Celiac disease slides (N = 318) were split into training (n = 230; 72.3%), validation (n = 60; 18.9%) and test (n = 28; 8.8%) datasets. Normal duodenum slides (N = 58) were similarly divided into training (n = 40; 69.0%), validation (n = 12; 20.7%) and test (n = 6; 10.3%) datasets. Human interpretable features were extracted and the strength of their correlation with Marsh scores were calculated using Spearman rank correlations. Our model identified cells, tissue regions and artifacts, including distinguishing intraepithelial lymphocytes and differentiating villous epithelium from crypt epithelium. Proportional area measurements representing villous atrophy negatively correlated with Marsh scores (r = − 0.79), while measurements indicative of crypt hyperplasia positively correlated (r = 0.71). Furthermore, features distinguishing celiac disease from normal duodenum were identified. Our novel model provides an explainable and fully automated approach for histology characterisation of celiac disease that correlates with modified Marsh scores, potentially facilitating diagnosis, prognosis, clinical trials and treatment response monitoring.

Comparison of Multiple Regression and Model Averaging Model-Building Approach for Missing Data with Multiple Imputation

Model construction is of significant importance for the extraction of information from datasets and the prediction of responses based on predictor variables. The objective of this study is to compare the Multiple Regression (MR) and model averaging approaches in the context of missing data and to validate the effectiveness of the Multiple Imputation (MI) method used to address missing data issues. A comparison was performed between the results obtained from the multiple-imputed data and those derived from the Complete Case (CC) data, using a diabetes dataset from Hospital Besar Alor Setar. Prior to the application of MI and model building, k-fold cross-validation was employed to partition the dataset, resulting in 90% of the data lacking complete covariates for training and 10% of the data comprising complete covariates for testing. Subsequently, MI was applied to the 90% training dataset. Model M115, derived from the multiple-imputed data, was identified as the optimal model for MR. In the model averaging approach, two models were identified as optimal: Model 1 (without interaction variables) and Model 2 (with interaction variables). The first one, exhibited the lowest values of Mean Square Error (MSE), Root Mean Square Error (RMSE), and Mean Absolute Error (MAE). These results indicate that model averaging, specifically Model 1, is the superior model-building approach for this study, demonstrating improved performance compared to MR and validating the effectiveness of the MI method.

Deep-learning surrogate models for the stability of a wide rectangular tunnel

The stability of a wide rectangular tunnel in spatially variable soils is investigated using deep-learning surrogate models combined with a mixed formulation of limit analysis. The mixed formulation is employed to perform stochastic simulations of tunnels in spatially random soils, generating the data to train the surrogate models. Two types of deep-learning models are tested: Fully Connected Neural Networks (FCNs) and Convolutional Neural Networks (CNNs). To enhance model performance, the study introduces two novel encoding techniques to enhance the identification of tunnel regions in the input data. The first, Mask-FCN, directly incorporates spatial information into the model by adding a binary mask layer to distinguish tunnel regions from the surrounding soil. The second, Grad-FCN, modifies training gradients to indirectly highlight these differences. Additionally, a Mask-CNN model is developed, leveraging spatial filtering to better capture local patterns and correlations. Our results show that the surrogate models, trained on data from stochastic reliability analyses, demonstrate robustness and accuracy across various evaluation metrics. The encoding neural networks outperform the baseline model in accuracy and overall regression performance on training datasets. Importantly, both encoding networks achieve acceptable accuracy on unseen datasets with new tunnel geometric features, a task the baseline model fails to accomplish. The FCN-based encoding neural network achieves prediction capabilities comparable to ordinary CNNs, while the CNN-based encoding neural network delivers high prediction accuracy for unknown datasets with new tunnel geometric characteristics. The study concludes that the generated surrogate models offer a computationally efficient alternative for predicting tunnel stability in spatially variable soils.

Predictors of Death or Severe Impairment in Neonates With Hypoxic-Ischemic Encephalopathy

Outcomes after hypoxic-ischemic encephalopathy (HIE) are variable. Predicting death or severe neurodevelopmental impairment (NDI) in affected neonates is crucial for guiding management and parent communication. To predict death or severe NDI in neonates who receive hypothermia for HIE. This prognostic study included participants enrolled in a large US clinical trial conducted in US neonatal intensive care units who were born between January 2017 and October 2019 and followed up to age 2 years. Eligible participants were neonates with moderate-severe HIE born at 36 weeks or more gestation and with 2-year outcome data. Data were analyzed June 2023. External validation was performed with a UK cohort. Clinical, electroencephalography (EEG), and magnetic resonance imaging (MRI) variables were curated and examined at 24 hours and following cooling. Death or severe NDI at age 2 years. Severe NDI was defined as Bayley Scales of Infant Toddler Development cognitive score below 70, Gross Motor Function Classification System score of 3 or higher, or quadriparesis. Model performance metrics were derived from training, internal, and external validation datasets. Among 424 neonates (mean [SD] gestational age, 39.1 [1.4] weeks; 192 female [45.3%]; 28 Asian [6.6%], 50 Black [11.8%], 311 White [73.3%]), 105 (24.7%) had severe encephalopathy at enrollment. Overall, 59 (13.9%) died and 46 (10.8%) had severe NDI. In the 24-hour model, the combined presence of 3 clinical characteristics-(1) severely abnormal EEG, (2) pH level of 7.11 or below, and (3) 5-minute Apgar score of 0-had a specificity of 99.6% (95% CI, 97.5%-100%) and a positive predictive value (PPV) of 95.2% (95% CI, 73.2%-99.3%). Validation model metrics were 97.9% (95% CI, 92.7%-99.8%) for internal specificity, with a PPV of 77.8% (95% CI, 43.4%-94.1%), and 97.6% (95% CI, 95.1%-99.0%) for external specificity, with a PPV of 46.2% (95% CI, 23.3%-70.8%). In the postcooling model, specificity for T1, T2, or diffusion-weighted imaging (DWI) abnormality in at least 2 of 3 deep gray regions (ie, thalamus, caudate, putamen and/or globus pallidus) plus a severely abnormal EEG within the first 24 hours was 99.1% (95% CI, 96.8%-99.9%), with a PPV of 91.7% (95% CI, 72.8%-97.8%). Internal specificity in this model was 98.9% (95% CI, 94.1%-100%), with a PPV of 92.9% (95% CI, 64.2%-99.0%); external specificity was 98.6% (95% CI, 96.5%-99.6%), with a PPV of 83.3% (95% CI, 64.1%-93.4%). In this prognostic study of neonates with moderate or severe HIE who were treated with therapeutic hypothermia, simple models using readily available clinical, EEG, and MRI results during the hospital admission had high specificity and PPV for death or severe NDI at age 2 years.

An Ensemble Kernelized-based Approach for Precise Emotion Recognition in Depressed People

As the COVID-19 pandemic created serious challenges for mental health worldwide, with a noticeable increase in depression cases, it has become important to quickly and accurately assess emotional states. Facial expression recognition technology is a key tool for this task. To address this need, this study proposes a new approach to emotion recognition using the Ensemble Kernelized Learning System (EKLS). Nonverbal cues, such as facial expressions, are crucial in showing emotional states. This study uses the Extended Cohn-Kanade (CK+) dataset, which was enhanced with images and videos from the COVID-19 era related to depression. Each of these images and videos is manually labeled with the corresponding emotions, creating a strong dataset for training and testing the proposed model. Facial feature detection techniques were used along with key facial measurements to aid in emotion recognition. EKLS is a flexible machine-learning framework that combines different techniques, including Support Vector Machines (SVMs), Self-Organizing Maps (SOMs), kernel methods, Random Forest (RF), and Gradient Boosting (GB). The ensemble model was thoroughly trained and fine-tuned to ensure high accuracy and consistency. EKLS is a powerful tool for real-time emotion recognition in both images and videos, achieving an impressive accuracy of 99.82%. This study offers a practical and effective approach to emotion recognition and makes a significant contribution to the field.

Compensated Neural Network Training Algorithm with Minimized Training Dataset for Modeling the Switching Transients of SiC MOSFETs

Accurate modeling of the switching transients of SiC MOSFETs is essential for overvoltage evaluation, EMI prediction, and other critical applications. Due to the fast switching speed, the switching transients of SiC MOSFETs are highly sensitive to parasitic parameters and nonlinear components, making precise modeling challenging. This paper proposes a hybrid model for SiC MOSFET, in which the analytical model is treated as the basis to provide the fundamental waveforms (knowledge-driven), while the neural network (NN) is utilized to fit the high-order and nonlinear features (data-driven). An NN training method with augmented data is proposed to minimize the training datasets. Verification results show that, even though the NN is trained with the data from a single operating condition, the model can accurately predict switching transients of other operating conditions. The proposed methodology has the potential to co-work with the “black-box” or “grey-box” models to enhance the model accuracy.