Depression assessment using integrated multi-featured EEG bands deep neural network models: Leveraging ensemble learning techniques

Mental Status Assessment (MSA) holds significant importance in psychiatry. In recent years, several studies have leveraged Electroencephalogram (EEG) technology to gauge an individual's mental state or level of depression. This study introduces a novel multi-tier ensemble learning approach to integrate multiple EEG bands for conducting mental state or depression assessments. Initially, the EEG signal is divided into eight sub-bands, and then a Long Short-Term Memory (LSTM)-based Deep Neural Network (DNN) model is trained for each band. Subsequently, the integration of multi-band EEG frequency models and the evaluation of mental state or depression level are facilitated through a two-tier ensemble learning approach based on Multiple Linear Regression (MLR). The authors conducted numerous experiments to validate the performance of the proposed method under different evaluation metrics. For clarity and conciseness, the research employs the simplest commercialized one-channel EEG sensor, positioned at FP1, to collect data from 57 subjects (49 depressed and 18 healthy subjects). The obtained results, including an accuracy of 0.897, F1-score of 0.921, precision of 0.935, negative predictive value of 0.829, recall of 0.908, specificity of 0.875, and AUC of 0.8917, provide evidence of the superior performance of the proposed method compared to other ensemble learning techniques. This method not only proves effective but also holds the potential to significantly enhance the accuracy of depression assessment.

A lightweight and gradient-stable neural layer

To enhance resource efficiency and model deployability of neural networks, we propose a neural-layer architecture based on Householder weighting and absolute-value activating, called Householder-absolute neural layer or simply Han-layer. Compared to a fully connected layer with d-neurons and d outputs, a Han-layer reduces the number of parameters and the corresponding computational complexity from O(d2) to O(d). The Han-layer structure guarantees that the Jacobian of the layer function is always orthogonal, thus ensuring gradient stability (i.e., free of gradient vanishing or exploding issues) for any Han-layer sub-networks. Extensive numerical experiments show that one can strategically use Han-layers to replace fully connected (FC) layers, reducing the number of model parameters while maintaining or even improving the generalization performance. We will also showcase the capabilities of the Han-layer architecture on a few small stylized models, and discuss its current limitations.

An artificial intelligence-based model for predicting reproductive toxicity of bisphenol analogues mixtures to the rotifer Brachionus calyciflorus

The joint toxicity effects of mixtures, particularly reproductive toxicity, one of the main causes of aquatic ecosystem degradation, are often overlooked as it is impractical to test all mixtures. This study developed and evaluated the following models to predict the concentration response curve concerning the joint reproductive toxicity of mixtures of three bisphenol analogues (BPA, BPF, BPAF) on the rotifer Brachionus calyciflorus: concentration addition (CA), independent action (IA), and two deep neural network (DNN) models. One applied mixture molecular descriptors as input variables (DNN-QSAR), while the other applied the ratios of chemicals in the mixtures (DNN-Ratio). Descriptors related to molecular mass were found to be of greater importance and exhibited a proportional relationship with toxic effects. The results indicate that the range of correlation coefficients (R2) between predicted and measured values for various mixture rays by CA and IA models is 0.372 to 0.974 and - 0.970 to 0.586, respectively. The R2 values for DNN-Ratio and DNN-QSAR were 0.841 to 0.984 and 0.834 to 0.991, respectively, demonstrating that models developed by DNN significantly outperform traditional models in predicting the joint toxicity of mixtures. Furthermore, DNN-QSAR not only predicts mixture toxicity but also provides accurate toxicity predictions for BPA, BPF, and BPAF, with R2 values of 0.990, 0.616, and 0.887, respectively, while DNN-Ratio yields values of 0.920, 0.355, and - 0.495. The study also found that the joint effects of mixtures are primarily influenced by the total concentration of the mixtures, and an increase in total concentration shifts the joint effects towards addition. This study introduces a novel approach to predict joint toxicity and analyze the influencing factors of joint effects, providing a more comprehensive assessment of the ecological risk posed by mixtures.

Prediction of systemic lupus erythematosus-related genes based on graph attention network and deep neural network

Systemic lupus erythematosus (SLE) is an autoimmune disorder intricately linked to genetic factors, with numerous approaches having identified genes linked to its development, diagnosis and prognosis. Despite genome-wide association analysis and gene knockout experiments confirming some genes associated with SLE, there are still numerous potential genes yet to be discovered. The search for relevant genes through biological experiments entails significant financial and human resources. With the advancement of computational technologies like deep learning, we aim to identify SLE-related genes through deep learning methods, thereby narrowing down the scope for biological experimentation. This study introduces SLEDL, a deep learning-based approach that leverages DNN and graph neural networks to effectively identify SLE-related genes by capturing relevant features in the gene interaction network. The above steps transform the identification of SLE related genes into a binary classification problem, ultimately solved through a fully connected layer. The results demonstrate the superiority of SLEDL, achieving higher AUC (0.7274) and AUPR (0.7599), further validated through case studies.

Functional loops: Monitoring functional organization of deep neural networks using algebraic topology

Various topological methods have emerged in recent years to investigate the inner workings of deep neural networks (DNNs) based on the structural and weight information. However, their effectiveness is restricted due to the stratified structure and volatile weight information. In this study, we explore the relationship between functional organizations and network performance using algebraic topology. Our results indicate that functional loops reveal functional interaction patterns of multiple neurons in DNNs. We also propose functional persistence as a measure of functional complexity and develop an early stopping criterion that achieves competitive results without requiring a validation set.

Low dimensional approximation and generalization of multivariate functions on smooth manifolds using deep ReLU neural networks

The expressive power of deep neural networks is manifested by their remarkable ability to approximate multivariate functions in a way that appears to overcome the curse of dimensionality. This ability is exemplified by their success in solving high-dimensional problems where traditional numerical solvers fail due to their limitations in accurately representing high-dimensional structures. To provide a theoretical framework for explaining this phenomenon, we analyze the approximation of Hölder functions defined on a d-dimensional smooth manifold M embedded in RD, with d≪D, using deep neural networks. We prove that the uniform convergence estimates of the approximation and generalization errors by deep neural networks with ReLU activation functions do not depend on the ambient dimension D of the function but only on its lower manifold dimension d, in a precise sense. Our result improves existing results from the literature where approximation and generalization errors were shown to depend weakly on D.

Spatiotemporal variation reconstruction of total phosphorus in the Great Lakes since 2002 using remote sensing and deep neural network

Total phosphorus (TP) is non-optically active, thus TP concentration (CTP) estimation using remote sensing still exists grand challenge. This study developed a deep neural network model (DNN) for CTP estimation with synchronous in-situ measurements and MODIS-derived remote sensing reflectance (Rrs) (N = 3916). Using DNN, the annual and intra-annual CTP spatial distributions of the Great Lakes since 2002 were reconstructed. Then, the reconstructions were correlated to nine potential factors, e.g., Chlorophyll-a, snowmelt, and cropland, to explain seasonal and long-term CTP variations. The results showed that DNN reliably estimated CTP from MODIS Rrs, with R2, mean absolute error (MAE), root mean squared error (RMSE), mean absolute percentage error (MAPE), and root mean squared logarithmic error (RMSLE) of 0.83, 1.05 μg/L, 2.95 μg/L, 9.92%, and 0.13 on the test set. The near-surface CTP in the Great Lakes decreased significantly (p < 0.05) during 2002 - 2022, primarily attributed to cropland reduction, coupled with improvements in basin natural ecosystems. The sensitivity analysis verified the model robustness when confronted with input feature changes < 35%. This result along with the marginal difference between CTP derived from two sensors (R2 = 0.76, MAE = 2.12 μg/L, RMSE = 2.51 μg/L, MAPE = 11.52%, RMSLE = 0.24) suggested the model transferability from MODIS to VIIRS. This transformation facilitated optimal usage of MODIS-related archive and enhanced the continuity of CTP estimation at moderate resolution. This study presents a practical method for spatiotemporal reconstruction of CTP using remote sensing, and contributes to better understandings of driving factors behind CTP variations in the Great Lakes.

Image analysis and teaching strategy optimization of folk dance training based on the deep neural network

To improve the recognition effect of the folk dance image recognition model and put forward new suggestions for teachers' teaching strategies, this study introduces a Deep Neural Network (DNN) to optimize the folk dance training image recognition model. Moreover, a corresponding teaching strategy optimization scheme is proposed according to the experimental results. Firstly, the image preprocessing and feature extraction of DNN are optimized. Secondly, classification and target detection models are established to analyze the folk dance training images, and the C-dance dataset is used for experiments. Finally, the results are compared with those of the Naive Bayes classifier, K-nearest neighbor, decision tree classifier, support vector machine, and logistic regression models. The results of this study provide new suggestions for teaching strategies. The research results indicate that the optimized classification model shows a significant improvement in classification accuracy across various aspects such as action complexity, dance types, movement speed, dance styles, body dynamics, and rhythm. The accuracy, precision, recall, and F1 scores have increased by approximately 14.7, 11.8, 13.2, and 17.4%, respectively. In the study of factors such as different training images, changes in perspective, lighting conditions, and noise interference, the optimized model demonstrates a substantial enhancement in recognition accuracy and robustness. These findings suggest that, compared to traditional models, the optimized model performs better in identifying various dances and movements, enhancing the accuracy and stability of classification. Based on the experimental results, strategies for optimizing the real-time feedback and assessment mechanism in folk dance teaching, as well as the design of personalized learning paths, are proposed. Therefore, this study holds the potential to be applied in the field of folk dance, promoting the development and innovation of folk dance education.

Enhancing cervical cancer detection and robust classification through a fusion of deep learning models

Cervical cancer, the second most prevalent cancer affecting women, arises from abnormal cell growth in the cervix, a crucial anatomical structure within the uterus. The significance of early detection cannot be overstated, prompting the use of various screening methods such as Pap smears, colposcopy, and Human Papillomavirus (HPV) testing to identify potential risks and initiate timely intervention. These screening procedures encompass visual inspections, Pap smears, colposcopies, biopsies, and HPV-DNA testing, each demanding the specialized knowledge and skills of experienced physicians and pathologists due to the inherently subjective nature of cancer diagnosis. In response to the imperative for efficient and intelligent screening, this article introduces a groundbreaking methodology that leverages pre-trained deep neural network models, including Alexnet, Resnet-101, Resnet-152, and InceptionV3, for feature extraction. The fine-tuning of these models is accompanied by the integration of diverse machine learning algorithms, with ResNet152 showcasing exceptional performance, achieving an impressive accuracy rate of 98.08%. It is noteworthy that the SIPaKMeD dataset, publicly accessible and utilized in this study, contributes to the transparency and reproducibility of our findings. The proposed hybrid methodology combines aspects of DL and ML for cervical cancer classification. Most intricate and complicated features from images can be extracted through DL. Further various ML algorithms can be implemented on extracted features. This innovative approach not only holds promise for significantly improving cervical cancer detection but also underscores the transformative potential of intelligent automation within the realm of medical diagnostics, paving the way for more accurate and timely interventions.

GEnDDn: An lncRNA-Disease Association Identification Framework Based on Dual-Net Neural Architecture and Deep Neural Network

Accumulating studies have demonstrated close relationships between long non-coding RNAs (lncRNAs) and diseases. Identification of new lncRNA-disease associations (LDAs) enables us to better understand disease mechanisms and further provides promising insights into cancer targeted therapy and anti-cancer drug design. Here, we present an LDA prediction framework called GEnDDn based on deep learning. GEnDDn mainly comprises two steps: First, features of both lncRNAs and diseases are extracted by combining similarity computation, non-negative matrix factorization, and graph attention auto-encoder, respectively. And each lncRNA-disease pair (LDP) is depicted as a vector based on concatenation operation on the extracted features. Subsequently, unknown LDPs are classified by aggregating dual-net neural architecture and deep neural network. Using six different evaluation metrics, we found that GEnDDn surpassed four competing LDA identification methods (SDLDA, LDNFSGB, IPCARF, LDASR) on the lncRNADisease and MNDR databases under fivefold cross-validation experiments on lncRNAs, diseases, LDPs, and independent lncRNAs and independent diseases, respectively. Ablation experiments further validated the powerful LDA prediction performance of GEnDDn. Furthermore, we utilized GEnDDn to find underlying lncRNAs for lung cancer and breast cancer. The results elucidated that there may be dense linkages between IFNG-AS1 and lung cancer as well as between HIF1A-AS1 and breast cancer. The results require further biomedical experimental verification. GEnDDn is publicly available at https://github.com/plhhnu/GEnDDn.

TIRPnet: Risk prediction of traditional Chinese medicine ingredients based on a deep neural network

ETHNOPHARMACOLOGICAL RELEVANCE

Traditional Chinese medicine (TCM) has a history of over 3000 years of medical practice. Due to the complex ingredients and unclear pharmacological mechanism of TCM, it is very difficult to predict its risks. With the increase in the number and severity of spontaneous reports of adverse drug reactions (ADRs) of TCM, its safety has received widespread attention.

AIM OF THE STUDY

In this study, we proposed a framework based on deep learning to predict the probability of adverse reactions caused by TCM ingredients and validated the model using real-world data.

MATERIALS AND METHODS

The spontaneous reporting data from Jiangsu Province of China was selected as the research data, which included 72,561 ADR reports of TCMs. All the ingredients of these TCMs were collected from the medical website and correlated with the corresponding ADRs. Then, a risk prediction model was constructed based on a deep neural network (DNN), named TIRPnet. Based on one-hot encoded data, our model achieved the optimal performance by fine-tuning some hyperparameters. The ten most commonly used TCM ingredients and their ADRs were collected as the test set to evaluate their performance as objective criteria.

RESULTS

TIRPnet was constructed as a 7-layer DNN. The experimental results showed that TIRPnet performs excellently in all indicators, with a sensitivity of 0.950, specificity of 0.995, accuracy of 0.994, precision of 0.708, and F1 of 0.811.

CONCLUSIONS

The proposed TIRPnet owns the ability to predict the ADRs of a single TCM ingredient by learning a large number of TCM-related spontaneous reports, which can help doctors design safe prescriptions and provide technical support for the pharmacovigilance of TCM.

Structural- and DTI- MRI enable automated prediction of IDH Mutation Status in CNS WHO Grade 2-4 glioma patients: a deep Radiomics Approach

BACKGROUND

The role of isocitrate dehydrogenase (IDH) mutation status for glioma stratification and prognosis is established. While structural magnetic resonance image (MRI) is a promising biomarker, it may not be sufficient for non-invasive characterisation of IDH mutation status. We investigated the diagnostic value of combined diffusion tensor imaging (DTI) and structural MRI enhanced by a deep radiomics approach based on convolutional neural networks (CNNs) and support vector machine (SVM), to determine the IDH mutation status in Central Nervous System World Health Organization (CNS WHO) grade 2-4 gliomas.

METHODS

This retrospective study analyzed the DTI-derived fractional anisotropy (FA) and mean diffusivity (MD) images and structural images including fluid attenuated inversion recovery (FLAIR), non-enhanced T1-, and T2-weighted images of 206 treatment-naïve gliomas, including 146 IDH mutant and 60 IDH-wildtype ones. The lesions were manually segmented by experienced neuroradiologists and the masks were applied to the FA and MD maps. Deep radiomics features were extracted from each subject by applying a pre-trained CNN and statistical description. An SVM classifier was applied to predict IDH status using imaging features in combination with demographic data.

RESULTS

We comparatively assessed the CNN-SVM classifier performance in predicting IDH mutation status using standalone and combined structural and DTI-based imaging features. Combined imaging features surpassed stand-alone modalities for the prediction of IDH mutation status [area under the curve (AUC) = 0.846; sensitivity = 0.925; and specificity = 0.567]. Importantly, optimal model performance was noted following the addition of demographic data (patients' age) to structural and DTI imaging features [area under the curve (AUC) = 0.847; sensitivity = 0.911; and specificity = 0.617].

CONCLUSIONS

Imaging features derived from DTI-based FA and MD maps combined with structural MRI, have superior diagnostic value to that provided by standalone structural or DTI sequences. In combination with demographic information, this CNN-SVM model offers a further enhanced non-invasive prediction of IDH mutation status in gliomas.

A deep neural network predictor to predict the sensitivity of neoadjuvant chemoradiotherapy in locally advanced rectal cancer

Neoadjuvant chemoradiotherapy (NCRT) is widely used for locally advanced rectal cancer (LARC). This study aimed to conduct an effective model to predict NCRT sensitivity and provide guidance for clinical treatment. Biomarkers for NCRT sensitivity were identified by applying transcriptome profiles using logistic regression and subsequently screened out by Spearman correlation analysis and four machine learning algorithms. A deep neural network (DNN) predictor was constructed by using in-house dataset and validated in two independent datasets. Additionally, a web-based program was developed. Wnt/β-catenin signaling and linoleic acid metabolism (LA) pathways were associated with NCRT sensitivity and prognosis in LARC, antagonistically. A DNN predictor with an 18-gene signature was conducted within in-house datasets. In two validation cohorts, area under ROC curve (AUC) achieved 0.706 and 0.897. The DNN subtypes were significantly associated with NCRT sensitivity, survival status et al. Moreover, NK and cytotoxic T cells were observed contribution to NCRT sensitivity while regulatory T, myeloid-derived suppressor cells and dysfunction of CD4 T effector memory cells could impede NCRT response. A DNN predictor could predict NCRT sensitivity in LARC and stratify LARC patients with different clinical and immunity characteristic.

High cell density cultivation of Corynebacterium glutamicum by deep learning-assisted medium design and the subsequent feeding strategy

To improve the cell productivity of Corynebacterium glutamicum, its initial specific growth rate was improved by medium improvement using deep neural network (DNN)-assisted design with Bayesian optimization (BO) and a genetic algorithm (GA). To obtain training data for the DNN, experimental design with an orthogonal array was set up using a chemically defined basal medium (GC XII). Based on the cultivation results for the training data, specific growth rates were observed between 0.04 and 0.3/h. The resulting DNN model estimated the test data with high accuracy (R2test ≥ 0.98). According to the validation cultivation, specific growth rates in the optimal media components estimated by DNN-BO and DNN-GA increased from 0.242 to 0.355/h. Using the optimal media (UCB_3), the specific growth rate, along with other parameters, was evaluated in batch culture. The specific growth rate reached 0.371/h from 3 to 12 h, and the dry cell weight was 28.0 g/L at 22.5 h. From the cultivation, the cell yields against glucose, ammonium ion, phosphate ion, sulfate ion, potassium ion, and magnesium ion were calculated. The cell yield calculation was used to estimate the required amounts of each component, and magnesium was found to limit the cell growth. However, in the follow-up fed-batch cultivation, glucose and magnesium addition was required to achieve the high initial specific growth rate, while appropriate feeding of glucose and magnesium during cultivation resulted in maintaining the high specific growth rate, and obtaining a cell yield of 80 g/Lini.

Diagnostic performance of deep-learning-based virtual chromoendoscopy in gastric neoplasms

BACKGROUNDS

Cycle-consistent generative adversarial network (CycleGAN) is a deep neural network model that performs image-to-image translations. We generated virtual indigo carmine (IC) chromoendoscopy images of gastric neoplasms using CycleGAN and compared their diagnostic performance with that of white light endoscopy (WLE).

METHODS

WLE and IC images of 176 patients with gastric neoplasms who underwent endoscopic resection were obtained. We used 1,633 images (911 WLE and 722 IC) of 146 cases in the training dataset to develop virtual IC images using CycleGAN. The remaining 30 WLE images were translated into 30 virtual IC images using the trained CycleGAN and used for validation. The lesion borders were evaluated by 118 endoscopists from 22 institutions using the 60 paired virtual IC and WLE images. The lesion area concordance rate and successful whole-lesion diagnosis were compared.

RESULTS

The lesion area concordance rate based on the pathological diagnosis in virtual IC was lower than in WLE (44.1% vs. 48.5%, p < 0.01). The successful whole-lesion diagnosis was higher in the virtual IC than in WLE images; however, the difference was insignificant (28.2% vs. 26.4%, p = 0.11). Conversely, subgroup analyses revealed a significantly higher diagnosis in virtual IC than in WLE for depressed morphology (41.9% vs. 36.9%, p = 0.02), differentiated histology (27.6% vs. 24.8%, p = 0.02), smaller lesion size (42.3% vs. 38.3%, p = 0.01), and assessed by expert endoscopists (27.3% vs. 23.6%, p = 0.03).

CONCLUSIONS

The diagnostic ability of virtual IC was higher for some lesions, but not completely superior to that of WLE. Adjustments are required to improve the imaging system's performance.

DRSPRING: Graph convolutional network (GCN)-Based drug synergy prediction utilizing drug-induced gene expression profile

Great efforts have been made over the years to identify novel drug pairs with synergistic effects. Although numerous computational approaches have been proposed to analyze diverse types of biological big data, the pharmacogenomic profiles, presumably the most direct proxy of drug effects, have been rarely used due to the data sparsity problem. In this study, we developed a composite deep-learning-based model that predicts the drug synergy effect utilizing pharmacogenomic profiles as well as molecular properties. Graph convolutional network (GCN) was used to represent and integrate the chemical structure, genetic interactions, drug-target information, and gene expression profiles of cell lines. Insufficient amount of pharmacogenomic data, i.e., drug-induced expression profiles from the LINCS project, was resolved by augmenting the data with the predicted profiles. Our method learned and predicted the Loewe synergy score in the DrugComb database and achieved a better or comparable performance compared to other published methods in a benchmark test. We also investigated contribution of various input features, which highlighted the value of basal gene expression and pharmacogenomic profiles of each cell line. Importantly, DRSPRING (DRug Synergy PRediction by INtegrated GCN) can be applied to any drug pairs and any cell lines, greatly expanding its applicability compared to previous methods.

ECG-surv: A deep learning-based model to predict time to 1-year mortality from 12-lead electrocardiogram

BACKGROUND

Electrocardiogram (ECG) abnormalities have demonstrated potential as prognostic indicators of patient survival. However, the traditional statistical approach is constrained by structured data input, limiting its ability to fully leverage the predictive value of ECG data in prognostic modeling.

METHODS

This study aims to introduce and evaluate a deep-learning model to simultaneously handle censored data and unstructured ECG data for survival analysis. We herein introduce a novel deep neural network called ECG-surv, which includes a feature extraction neural network and a time-to-event analysis neural network. The proposed model is specifically designed to predict the time to 1-year mortality by extracting and analyzing unique features from 12-lead ECG data. ECG-surv was evaluated using both an independent test set and an external set, which were collected using different ECG devices.

RESULTS

The performance of ECG-surv surpassed that of the Cox proportional model, which included demographics and ECG waveform parameters, in predicting 1-year all-cause mortality, with a significantly higher concordance index (C-index) in ECG-surv than in the Cox model using both the independent test set (0.860 [95% CI: 0.859- 0.861] vs. 0.796 [95% CI: 0.791- 0.800]) and the external test set (0.813 [95% CI: 0.807- 0.814] vs. 0.764 [95% CI: 0.755- 0.770]). ECG-surv also demonstrated exceptional predictive ability for cardiovascular death (C-index of 0.891 [95% CI: 0.890- 0.893]), outperforming the Framingham risk Cox model (C-index of 0.734 [95% CI: 0.715-0.752]).

CONCLUSION

ECG-surv effectively utilized unstructured ECG data in a survival analysis. It outperformed traditional statistical approaches in predicting 1-year all-cause mortality and cardiovascular death, which makes it a valuable tool for predicting patient survival.

Non-invasive screening and subtyping for breast cancer by serum SERS combined with LGB-DNN algorithms

Early detection of breast cancer and its molecular subtyping is crucial for guiding clinical treatment and improving survival rate. Current diagnostic methods for breast cancer are invasive, time consuming and complicated. In this work, an optical detection method integrating surface-enhanced Raman spectroscopy (SERS) technology with feature selection and deep learning algorithm was developed for identifying serum components and building diagnostic model, with the aim of efficient and accurate noninvasive screening of breast cancer. First, the high quality of serum SERS spectra from breast cancer (BC), breast benign disease (BBD) patients and healthy controls (HC) were obtained. Chi-square tests were conducted to exclude confounding factors, enhancing the reliability of the study. Then, LightGBM (LGB) algorithm was used as the base model to retain useful features to significantly improve classification performance. The DNN algorithm was trained through backpropagation, adjusting the weights and biases between neurons to improve the network's predictive ability. In comparison to traditional machine learning algorithms, this method provided more accurate information for breast cancer classification, with classification accuracies of 91.38 % for BC and BBD, and 96.40 % for BC, BBD, and HC. Furthermore, the accuracies of 90.11 % for HR+/HR- and 88.89 % for HER2+/HER2- can be reached when evaluating BC patients' molecular subtypes. These results demonstrate that serum SERS combined with powerful LGB-DNN algorithm would provide a supplementary method for clinical breast cancer screening.

Carbon sequestration costs and spatial spillover effects in China's collective forests

BACKGROUND

Global climate change is one of the major challenges facing the world today, and forests play a crucial role as significant carbon sinks and providers of ecosystem services in mitigating climate change and protecting the environment. China, as one of the largest developing countries globally, owns 60% of its forest resources collectively. Evaluating the carbon sequestration cost of collective forests not only helps assess the contribution of China's forest resources to global climate change mitigation but also provides important evidence for formulating relevant policies and measures.

RESULTS

Over the past 30 years, the carbon sequestration cost of collective forests in China has shown an overall upward trend. Except for coastal provinces, southern collective forest areas, as well as some southwestern and northeastern regions, have the advantage of lower carbon sequestration costs. Furthermore, LSTM network predictions indicate that the carbon sequestration cost of collective forests in China will continue to rise. By 2030, the average carbon sequestration cost of collective forests is projected to reach 125 CNY per ton(= 16.06 Euros/t). Additionally, there is spatial correlation in the carbon sequestration cost of collective forests. Timber production, labor costs, and labor prices have negative spatial spillover effects on carbon sequestration costs, while land opportunity costs, forest accumulation, and rural resident consumption have positive spatial spillover effects.

CONCLUSION

The results of this study indicate regional disparities in the spatial distribution of carbon sequestration costs of collective forests, with an undeniable upward trend in future cost growth. It is essential to focus on areas with lower carbon sequestration costs and formulate targeted carbon sink economic policies and management measures to maximize the carbon sequestration potential of collective forests and promote the sustainable development of forestry.

Multi-scale full spike pattern for semantic segmentation

Spiking neural networks (SNNs), as the brain-inspired neural networks, encode information in spatio-temporal dynamics. They have the potential to serve as low-power alternatives to artificial neural networks (ANNs) due to their sparse and event-driven nature. However, existing SNN-based models for pixel-level semantic segmentation tasks suffer from poor performance and high memory overhead, failing to fully exploit the computational effectiveness and efficiency of SNNs. To address these challenges, we propose the multi-scale and full spike segmentation network (MFS-Seg), which is based on the deep direct trained SNN and represents the first attempt to train a deep SNN with surrogate gradients for semantic segmentation. Specifically, we design an efficient fully-spike residual block (EFS-Res) to alleviate representation issues caused by spiking noise on different channels. EFS-Res utilizes depthwise separable convolution to improve the distributions of spiking feature maps. The visualization shows that our model can effectively extract the edge features of segmented objects. Furthermore, it can significantly reduce the memory overhead and energy consumption of the network. In addition, we theoretically analyze and prove that EFS-Res can avoid the degradation problem based on block dynamical isometry theory. Experimental results on the Camvid dataset, the DDD17 dataset, and the DSEC-Semantic dataset show that our model achieves comparable performance to the mainstream UNet network with up to 31× fewer parameters, while significantly reducing power consumption by over 13×. Overall, our MFS-Seg model demonstrates promising results in terms of performance, memory efficiency, and energy consumption, showcasing the potential of deep SNNs for semantic segmentation tasks. Our code is available in https://github.com/BICLab/MFS-Seg.

Quantitative detection of multicomponent SF6 decomposition products based on Fourier transform infrared spectroscopy combined with SCARS-DNN

Accurate and efficient quantitative analysis of the decomposition products of the insulating medium SF6 in gas-insulated switchgear (GIS) is important for an effective assessment of its internal insulation status. In this work, a quantitative calibration model of Fourier Transform Infrared Spectroscopy (FTIR) combined with SCARS-DNN (Stability Competitive Adaptive Reweighted Sampling-Deep Neural Network) is proposed for the rapid non-destructive detection of SF6 decomposition products. First, the interference of the background gas SF6 on the absorption spectra of the decomposition products is eliminated according to the Lambert-Beer law, while baseline correction and Savitzky-Golay (S-G) smoothing are used to remove baseline drift and noise. Subsequently, a Monte Carlo cross-validation method is used to detect and eliminate the anomalous samples. Then feature selection is performed using uninformative variable elimination (UVE) and stability competitive adaptive reweighted sampling (SCARS), and finally quantitative calibration models of FULL-DNN (full spectral band), UVE-DNN, and SCARS-DNN are developed. For the quantitative detection of SF6 decomposition products, the SCARS-DNN model had the best prediction performance with a maximum reduction of 96.18% in the root mean square error (RMSE) and 96.11% in the mean absolute percentage error (MAPE). Results reveal that the relative errors are basically kept below 1.36% when predicting the three decomposition products, even in the presence of a high level of SF6 interference. Therefore, the SCARS-DNN model is suitable for high-precision quantitative detection of SF6 decomposition gas.

Investigation of multidirectional toxicity induced by high-dose molybdenum exposure with Allium test

In this study, the multifaceted toxicity induced by high doses of the essential trace element molybdenum in Allium cepa L. was investigated. Germination, root elongation, weight gain, mitotic index (MI), micronucleus (MN), chromosomal abnormalities (CAs), Comet assay, malondialdehyde (MDA), proline, superoxide dismutase (SOD), catalase (CAT) and anatomical parameters were used as biomarkers of toxicity. In addition, detailed correlation and PCA analyzes were performed for all parameters discussed. On the other hand, this study focused on the development of a two hidden layer deep neural network (DNN) using Matlab. Four experimental groups were designed: control group bulbs were germinated in tap water and application group bulbs were germinated with 1000, 2000 and 4000 mg/L doses of molybdenum for 72 h. After germination, root tips were collected and prepared for analysis. As a result, molybdenum exposure caused a dose-dependent decrease (p < 0.05) in the investigated physiological parameter values, and an increase (p < 0.05) in the cytogenetic (except MI) and biochemical parameter values. Molybdenum exposure induced different types of CAs and various anatomical damages in root meristem cells. Comet assay results showed that the severity of DNA damage increased depending on the increasing molybdenum dose. Detailed correlation and PCA analysis results determined significant positive and negative interactions between the investigated parameters and confirmed the relationships of these parameters with molybdenum doses. It has been found that the DNN model is in close agreement with the actual data showing the accuracy of the predictions. MAE, MAPE, RMSE and R2 were used to evaluate the effectiveness of the DNN model. Collective analysis of these metrics showed that the DNN model performed well. As a result, it has been determined once again that high doses of molybdenum cause multiple toxicity in A. cepa and the Allium test is a reliable universal test for determining this toxicity. Therefore, periodic measurement of molybdenum levels in agricultural soils should be the first priority in preventing molybdenum toxicity.

Deep neural network-estimated age using optical coherence tomography predicts mortality

The concept of biological age has emerged as a measurement that reflects physiological and functional decline with ageing. Here we aimed to develop a deep neural network (DNN) model that predicts biological age from optical coherence tomography (OCT). A total of 84,753 high-quality OCT images from 53,159 individuals in the UK Biobank were included, among which 12,631 3D-OCT images from 8,541 participants without any reported medical conditions at baseline were used to develop an age prediction model. For the remaining 44,618 participants, OCT age gap, the difference between the OCT-predicted age and chronological age, was calculated for each participant. Cox regression models assessed the association between OCT age gap and mortality. The DNN model predicted age with a mean absolute error of 3.27 years and showed a strong correlation of 0.85 with chronological age. After a median follow-up of 11.0 years (IQR 10.9-11.1 years), 2,429 deaths (5.44%) were recorded. For each 5-year increase in OCT age gap, there was an 8% increased mortality risk (hazard ratio [HR] = 1.08, CI:1.02-1.13, P = 0.004). Compared with an OCT age gap within ± 4 years, OCT age gap less than minus 4 years was associated with a 16% decreased mortality risk (HR = 0.84, CI: 0.75-0.94, P = 0.002) and OCT age gap more than 4 years showed an 18% increased risk of death incidence (HR = 1.18, CI: 1.02-1.37, P = 0.026). OCT imaging could serve as an ageing biomarker to predict biological age with high accuracy and the OCT age gap, defined as the difference between the OCT-predicted age and chronological age, can be used as a marker of the risk of mortality.

Invariant feature based label correction for DNN when Learning with Noisy Labels

Learning with Noisy Labels (LNL) methods have been widely studied in recent years, which aims to improve the performance of Deep Neural Networks (DNNs) when the training dataset contains incorrectly annotated labels. Popular existing LNL methods rely on semantic features extracted by the DNN to detect and mitigate label noise. However, these extracted features are often spurious and contain unstable correlations with the label across different environments (domains), which can occasionally lead to incorrect prediction and compromise the efficacy of LNL methods. To mitigate this insufficiency, we propose Invariant Feature based Label Correction (IFLC), which reduces spurious features and accurately utilizes the learned invariant features that contain stable correlation to correct label noise. To the best of our knowledge, this is the first attempt to mitigate the issue of spurious features for LNL methods. IFLC consists of two critical processes: The Label Disturbing (LD) process and the Representation Decorrelation (RD) process. The LD process aims to encourage DNN to attain stable performance across different environments, thus reducing the captured spurious features. The RD process strengthens independence between each dimension of the representation vector, thus enabling accurate utilization of the learned invariant features for label correction. We then utilize robust linear regression for the feature representation to conduct label correction. We evaluated the effectiveness of our proposed method and compared it with state-of-the-art (sota) LNL methods on four benchmark datasets, CIFAR-10, CIFAR-100, Animal-10N, and Clothing1M. The experimental results show that our proposed method achieved comparable or even better performance than the existing sota methods. The source codes are available at https://github.com/yangbo1973/IFLC.

Multi-modal deep learning networks for RGB-D pavement waste detection and recognition

To create a clean living environment, governments around the world have hired a large number of workers to clean up waste on pavements, which is inefficient for waste management. To better alleviate this problem, relevant scholars have proposed several deep learning methods based on RGB images to achieve waste detection and recognition. Considering the limitations of color images, we propose an efficient multi-modal learning solution for pavement waste detection and recognition. Specifically, we construct a high-quality outdoor pavement waste dataset called OPWaste, which is more in line with real needs. Compared to other waste datasets, OPWaste dataset not only has the advantages of rich background and high diversity, but also provides color and depth images. Meanwhile, we explore six different multi-modal fusion methods and propose a novel multi-modal multi-scale network (MM-Net) for RGB-D waste detection and recognition. MM-Net introduces a novel multi-scale refinement module (MRM) and multi-scale interaction module (MIM). MRM can effectively refine critical features using attention mechanisms. MIM can gradually realize information interaction between hierarchical features. In addition, we select several representative methods and perform comparative experiments. Experimental results show that MM-Net based on the image addition fusion method outperforms other deep learning models and reaches 97.3% and 84.4% on mAP0.5 and AR metrics. In fact, multi-modal learning plays an important role in intelligent waste recycling. As a promising auxiliary tool, our solution can be applied to intelligent cleaning robots for automatic outdoor waste management.

Short-term wind power forecasting through stacked and bi directional LSTM techniques

Background

Computational intelligence (CI) based prediction models increase the efficient and effective utilization of resources for wind prediction. However, the traditional recurrent neural networks (RNN) are difficult to train on data having long-term temporal dependencies, thus susceptible to an inherent problem of vanishing gradient. This work proposed a method based on an advanced version of RNN known as long short-term memory (LSTM) architecture, which updates recurrent weights to overcome the vanishing gradient problem. This, in turn, improves training performance.

Methods

The RNN model is developed based on stack LSTM and bidirectional LSTM. The parameters like mean absolute error (MAE), standard deviation error (SDE), and root mean squared error (RMSE) are utilized as performance measures for comparison with recent state-of-the-art techniques.

Results

Results showed that the proposed technique outperformed the existing techniques in terms of RMSE and MAE against all the used wind farm datasets. Whereas, a reduction in SDE is observed for larger wind farm datasets. The proposed RNN approach performed better than the existing models despite fewer parameters. In addition, the approach requires minimum processing power to achieve compatible results.

Levenberg-Marquardt deep neural watermarking for 3D mesh using nearest centroid salient point learning

Watermarking is one of the crucial techniques in the domain of information security, preventing the exploitation of 3D Mesh models in the era of Internet. In 3D Mesh watermark embedding, moderately perturbing the vertices is commonly required to retain them in certain pre-arranged relationship with their neighboring vertices. This paper proposes a novel watermarking authentication method, called Nearest Centroid Discrete Gaussian and Levenberg-Marquardt (NCDG-LV), for distortion detection and recovery using salient point detection. In this method, the salient points are selected using the Nearest Centroid and Discrete Gaussian Geometric (NC-DGG) salient point detection model. Map segmentation is applied to the 3D Mesh model to segment into distinct sub regions according to the selected salient points. Finally, the watermark is embedded by employing the Multi-function Barycenter into each spatially selected and segmented region. In the extraction process, the embedded 3D Mesh image is extracted from each re-segmented region by means of Levenberg-Marquardt Deep Neural Network Watermark Extraction. In the authentication stage, watermark bits are extracted by analyzing the geometry via Levenberg-Marquardt back-propagation. Based on a performance evaluation, the proposed method exhibits high imperceptibility and tolerance against attacks, such as smoothing, cropping, translation, and rotation. The experimental results further demonstrate that the proposed method is superior in terms of salient point detection time, distortion rate, true positive rate, peak signal to noise ratio, bit error rate, and root mean square error compared to the state-of-the-art methods.

Using machine learning models to predict falls in hospitalised adults

BACKGROUND

Identifying patients at high risk of falling is crucial in implementing effective fall prevention programs. While the integration of information systems is becoming more widespread in the healthcare industry, it poses a significant challenge in analysing vast amounts of data to identify factors that could enhance patient safety.

OBJECTIVE

To determine fall-associated factors and develop high-performance prediction tools for at-risk patients in acute and sub-acute care services in Australia.

METHODS

A retrospective study of 672,400 patients admitted to acute and sub-acute care services within a large metropolitan tertiary health service in Victoria, Australia, between January 1, 2019, and December 31, 2021. Data were obtained from four sources: the Department of Health Victorian Admitted Episodes Dataset, RiskManTM, electronic health records, and the health workforce dataset. Machine learning techniques, including Random Forest and Deep Neural Network models, were used to analyse the data, predict patient falls, and identify the most important risk factors for falls in this population. Model performance was evaluated using accuracy, F1-score, precision, recall, specificity, Matthew's correlation coefficient, and the area under the receiver operating characteristic curve (AUC).

RESULTS

The deep neural network and random forest models were highly accurate in predicting hospital patient falls. The deep neural network model achieved an accuracy of 0.988 and a specificity of 0.999, while the RF achieved an accuracy of 0.989 and a specificity of 1.000. The top 20 variables impacting falls were compared across both models, and 12 common factors were identified. These factors can be broadly classified into three categories: patient-related factors, staffing-related factors, and admission-related factors. Although not all factors are modifiable, they must be considered when planning fall prevention interventions.

CONCLUSION

The study demonstrated machine learning's potential to predict falls and identify key risk factors. Further validation across diverse populations and settings is essential for broader applicability.

scFSNN: a feature selection method based on neural network for single-cell RNA-seq data

While single-cell RNA sequencing (scRNA-seq) allows researchers to analyze gene expression in individual cells, its unique characteristics like over-dispersion, zero-inflation, high gene-gene correlation, and large data volume with many features pose challenges for most existing feature selection methods. In this paper, we present a feature selection method based on neural network (scFSNN) to solve classification problem for the scRNA-seq data. scFSNN is an embedded method that can automatically select features (genes) during model training, control the false discovery rate of selected features and adaptively determine the number of features to be eliminated. Extensive simulation and real data studies demonstrate its excellent feature selection ability and predictive performance.

A comparative study of a traditional localization algorithm and a deep learning model for radioactive particle tracking application

Radioactive particle tracking is a nuclear technique that tracks a sealed radioactive particle inside a volume through a mathematical location algorithm, which is widely applied in many fields such as chemical and civil engineering in hydrodynamics flows. It is possible to reconstruct the trajectory of the radioactive particle using a traditional mathematical algorithm or artificial intelligence methods. In this paper, the traditional algorithm is based on solving a minimization problem between the simulated events and a calibration dataset, and it was written using C++ language. The artificial intelligence method is represented by a deep neural network, in which hyperparameters were defined using a Python optimization library called Optuna. This paper aims to compare the potentiality of both methods to evaluate the accuracy of the radioactive particle tracking technique. This study proposes a simplified model of a concrete mixer, six NaI(Tl) detectors, and a137Cs sealed radioactive particle. The simulated measurement geometry and the dataset (3615 patterns) were developed using the MCNPX code, which is a mathematical code based on the Monte Carlo Method. The results show a mean absolute percentage error (MAPE) of 20.81%, 10.33%, and 16.84% for x, y and z coordinates, respectively, for the traditional algorithm. For the deep neural network, MAPE is 6.87%, 2.70%, and 22.79% respectively for x, y and z coordinates. In addition, an investigation is carried out to analyze whether the size of the calibration dataset influences the performance of both methods.

Recent developments in machine learning modeling methods for hypertension treatment

Hypertension is the leading cause of cardiovascular complications. This review focuses on the advancements in medical artificial intelligence (AI) models aimed at individualized treatment for hypertension, with particular emphasis on the approach to time-series big data on blood pressure and the development of interpretable medical AI models. The digitalization of daily blood pressure records and the downsizing of measurement devices enable the accumulation and utilization of time-series data. As mainstream blood pressure data shift from snapshots to time series, the clinical significance of blood pressure variability will be clarified. The time-series blood pressure prediction model demonstrated the capability to forecast blood pressure variabilities with a reasonable degree of accuracy for up to four weeks in advance. In recent years, various explainable AI techniques have been proposed for different purposes of model interpretation. It is essential to select the appropriate technique based on the clinical aspects; for example, actionable path-planning techniques can present individualized intervention plans to efficiently improve outcomes such as hypertension. Despite considerable progress in this field, challenges remain, such as the need for the prospective validation of AI-driven interventions and the development of comprehensive systems that integrate multiple AI methods. Future research should focus on addressing these challenges and refining the AI models to ensure their practical applicability in real-world clinical settings. Furthermore, the implementation of interdisciplinary collaborations among AI experts, clinicians, and healthcare providers are crucial to further optimizing and validate AI-driven solutions for hypertension management.

Understanding the performance of RHO type zeolite membrane for CH4/N2 separation based on molecular dynamics and deep neural network methods

This study shows a molecular dynamics (MD) simulation study on the performance of the RHO zeolite membrane for separating nitrogen from methane/nitrogen gas mixtures. The contamination of natural gas, predominantly composed of methane, with nitrogen diminishes its value. Zeolite membranes offer promising prospects for gas separation due to their stability, rigid pore structure, and molecular sieving properties. The study investigates the impact of pressure difference (up to 30 MPa), feed composition, and membrane thickness on the separation rate at a system temperature of 298 K. Results demonstrate that the RHO zeolite membrane exhibits high permeability and selectivity for N2 separation, surpassing the upper limit defined by Robson with a maximum permeability of 2.14 × 105 GPU (Gas Permeation Units). Exceptional selectivity of N2 over CH4 molecules is observed. Additionally, altering the feed composition and membrane thickness positively influences the membrane's separation performance, thereby enhancing its efficiency. The findings contribute to the advancement of separation technologies, providing valuable insights into the potential application of zeolite membranes for efficient N2 separation from CH4/N2 gas mixtures in natural gas processing. Furthermore, the study explores the use of Deep Neural Network (DNN) models to predict the membrane's performance under diverse operating conditions. The DNN models, trained using simulation data from MD simulations, exhibit high accuracy with a coefficient of determination (R2) exceeding 0.9, ensuring reliable predictions. The integration of DNN models facilitates the optimization of zeolite membrane-based gas separation systems, improving their design and operation.

Interpreting drug synergy in breast cancer with deep learning using target-protein inhibition profiles

BACKGROUND

Breast cancer is the most common malignancy among women worldwide. Despite advances in treating breast cancer over the past decades, drug resistance and adverse effects remain challenging. Recent therapeutic progress has shifted toward using drug combinations for better treatment efficiency. However, with a growing number of potential small-molecule cancer inhibitors, in silico strategies to predict pharmacological synergy before experimental trials are required to compensate for time and cost restrictions. Many deep learning models have been previously proposed to predict the synergistic effects of drug combinations with high performance. However, these models heavily relied on a large number of drug chemical structural fingerprints as their main features, which made model interpretation a challenge.

RESULTS

This study developed a deep neural network model that predicts synergy between small-molecule pairs based on their inhibitory activities against 13 selected key proteins. The synergy prediction model achieved a Pearson correlation coefficient between model predictions and experimental data of 0.63 across five breast cancer cell lines. BT-549 and MCF-7 achieved the highest correlation of 0.67 when considering individual cell lines. Despite achieving a moderate correlation compared to previous deep learning models, our model offers a distinctive advantage in terms of interpretability. Using the inhibitory activities against key protein targets as the main features allowed a straightforward interpretation of the model since the individual features had direct biological meaning. By tracing the synergistic interactions of compounds through their target proteins, we gained insights into the patterns our model recognized as indicative of synergistic effects.

CONCLUSIONS

The framework employed in the present study lays the groundwork for future advancements, especially in model interpretation. By combining deep learning techniques and target-specific models, this study shed light on potential patterns of target-protein inhibition profiles that could be exploited in breast cancer treatment.

Highly accurate and effective deep neural networks in pathological diagnosis of prostate cancer

PURPOSE

To established an AI system to make the pathological diagnosis of prostate cancer.

METHODS

Prostate histopathological whole mount (WM) sections from patients underwent robot-assisted laparoscopic prostatectomy were prepared. All the prostate WM pathological sections were converted to digital image data and marked with different colors on the basis of the ISUP Gleason grade group. The image was then fed into a segmentation algorithm. We chose modified U-Net as our fundamental network architecture.

RESULTS

172 patients were involved in this study. 896 pieces of prostate WM pathological sections from 160 patients, in which 826 pieces of WM sections from 148 patients were assigned to the training set randomly. After image segmentation there were totally 2,138,895 patches, of which 1,646,535 patches were valid for training. The other WM section was arranged for testing. Based on the whole image testing, AI and pathologists presented the same answers among 21 of 22 pieces of sections. To evaluate the diagnostic results at the pixel level, we anticipated correct cancer or non-cancer diagnose from this AI system. The area under the ROC curve as 96.8%. The value of pixel accuracy of three methods (binary analysis, clinically oriented analysis and analysis for different ISUP Gleason grade) were 96.93%, 95.43% and 93.88%, respectively. The value of frequency weighted IoU were 94.32%, 92.13% and 90.21%, respectively.

CONCLUSIONS

This AI system is able to assist pathologists to make a final diagnosis, indicating the great potential and a wide-range of applications of AI in the medical field.

Geospatial prediction of total soil carbon in European agricultural land based on deep learning

Accurate geospatial prediction of soil parameters provides a basis for large-scale digital soil mapping, making efficient use of the expensive and time-consuming process of field soil sampling. To date, few studies have used deep learning for geospatial prediction of soil parameters, but there is evidence that it may provide higher accuracy compared to machine learning methods. To address this research gap, this study proposed a deep neural network (DNN) for geospatial prediction of total soil carbon (TC) in European agricultural land and compared it with the eight most commonly used machine learning methods based on studies indexed in the Web of Science Core Collection. A total of 6209 preprocessed soil samples from the Geochemical mapping of agricultural and grazing land soil (GEMAS) dataset in heterogeneous agricultural areas covering 4,899,602 km2 in Europe were used. Prediction was performed based on 96 environmental covariates from climate and remote sensing sources, with extensive comprehensive hyperparameter tuning for all evaluated methods. DNN outperformed all evaluated machine learning methods (R2 = 0.663, RMSE = 9.595, MAE = 5.565), followed by Quantile Random Forest (QRF) (R2 = 0.635, RMSE = 25.993, MAE = 22.081). The ability of DNN to accurately predict small TC values and thus produce relatively low absolute residuals was a major reason for the higher prediction accuracy compared to machine learning methods. Climate parameters were the main factors in the achieved prediction accuracy, with 23 of the 25 environmental covariates with the highest variable importance being climate or land surface temperature parameters. These results demonstrate the superiority of DNN over machine learning methods for TC prediction, while highlighting the need for more recent soil sampling to assess the impact of climate change on TC content in European agricultural land.

RUBICON: a framework for designing efficient deep learning-based genomic basecallers

Nanopore sequencing generates noisy electrical signals that need to be converted into a standard string of DNA nucleotide bases using a computational step called basecalling. The performance of basecalling has critical implications for all later steps in genome analysis. Therefore, there is a need to reduce the computation and memory cost of basecalling while maintaining accuracy. We present RUBICON, a framework to develop efficient hardware-optimized basecallers. We demonstrate the effectiveness of RUBICON by developing RUBICALL, the first hardware-optimized mixed-precision basecaller that performs efficient basecalling, outperforming the state-of-the-art basecallers. We believe RUBICON offers a promising path to develop future hardware-optimized basecallers.

DeepMPTB: a vaginal microbiome-based deep neural network as artificial intelligence strategy for efficient preterm birth prediction

In recent decades, preterm birth (PTB) has become a significant research focus in the healthcare field, as it is a leading cause of neonatal mortality worldwide. Using five independent study cohorts including 1290 vaginal samples from 561 pregnant women who delivered at term (n = 1029) or prematurely (n = 261), we analysed vaginal metagenomics data for precise microbiome structure characterization. Then, a deep neural network (DNN) was trained to predict term birth (TB) and PTB with an accuracy of 84.10% and an area under the receiver operating characteristic curve (AUROC) of 0.875 ± 0.11. During a benchmarking process, we demonstrated that our DL model outperformed seven currently used machine learning algorithms. Finally, our results indicate that overall diversity of the vaginal microbiota should be taken in account to predict PTB and not specific species. This artificial-intelligence based strategy should be highly helpful for clinicians in predicting preterm birth risk, allowing personalized assistance to address various health issues. DeepMPTB is open source and free for academic use. It is licensed under a GNU Affero General Public License 3.0 and is available at https://deepmptb.streamlit.app/ . Source code is available at https://github.com/oschakoory/DeepMPTB and can be easily installed using Docker ( https://www.docker.com/ ).

Explainable deep-neural-network supported scheme for tuberculosis detection from chest radiographs

Chest radiographs are examined in typical clinical settings by competent physicians for tuberculosis diagnosis. However, this procedure is time consuming and subjective. Due to the growing usage of machine learning techniques in applied sciences, researchers have begun applying comparable concepts to medical diagnostics, such as tuberculosis screening. In the period of extremely deep neural nets which comprised of hundreds of convolution layers for feature extraction, we create a shallow-CNN for screening of TB condition from Chest X-rays so that the model is able to offer appropriate interpretation for right diagnosis. The suggested model consists of four convolution-maxpooling layers with various hyperparameters that were optimized for optimal performance using a Bayesian optimization technique. The model was reported with a peak classification accuracy, F1-score, sensitivity and specificity of 0.95. In addition, the receiver operating characteristic (ROC) curve for the proposed shallow-CNN showed a peak area under the curve value of 0.976. Moreover, we have employed class activation maps (CAM) and Local Interpretable Model-agnostic Explanations (LIME), explainer systems for assessing the transparency and explainability of the model in comparison to a state-of-the-art pre-trained neural net such as the DenseNet.

Differential effects of intrinsic properties of natural scenes and interference mechanisms on recognition processes in long-term visual memory

Humans display remarkable long-term visual memory (LTVM) processes. Even though images may be intrinsically memorable, the fidelity of their visual representations, and consequently the likelihood of successfully retrieving them, hinges on their similarity when concurrently held in LTVM. In this debate, it is still unclear whether intrinsic features of images (perceptual and semantic) may be mediated by mechanisms of interference generated at encoding, or during retrieval, and how these factors impinge on recognition processes. In the current study, participants (32) studied a stream of 120 natural scenes from 8 semantic categories, which varied in frequencies (4, 8, 16 or 32 exemplars per category) to generate different levels of category interference, in preparation for a recognition test. Then they were asked to indicate which of two images, presented side by side (i.e. two-alternative forced-choice), they remembered. The two images belonged to the same semantic category but varied in their perceptual similarity (similar or dissimilar). Participants also expressed their confidence (sure/not sure) about their recognition response, enabling us to tap into their metacognitive efficacy (meta-d'). Additionally, we extracted the activation of perceptual and semantic features in images (i.e. their informational richness) through deep neural network modelling and examined their impact on recognition processes. Corroborating previous literature, we found that category interference and perceptual similarity negatively impact recognition processes, as well as response times and metacognitive efficacy. Moreover, images semantically rich were less likely remembered, an effect that trumped a positive memorability boost coming from perceptual information. Critically, we did not observe any significant interaction between intrinsic features of images and interference generated either at encoding or during retrieval. All in all, our study calls for a more integrative understanding of the representational dynamics during encoding and recognition enabling us to form, maintain and access visual information.

Ultrahigh-porosity Ranunculus-like MgO adsorbent coupled with predictive deep belief networks: A transformative method for phosphorus treatment

Phosphorus is a nonrenewable material with a finite supply on Earth; however, due to the rapid growth of the manufacturing industry, phosphorus contamination has become a global concern. Therefore, this study highlights the remarkable potential of ranunculus-like MgO (MO4-MO6) as superior adsorbents for phosphate removal and recovery. Furthermore, MO6 stands out with an impressive adsorption capacity of 596.88 mg/g and a high efficacy across a wide pH range (2-10) under varying coexisting ion concentrations. MO6 outperforms the top current adsorbents for phosphate removal. The process follows Pseudo-second-order and Langmuir models, indicating chemical interactions between the phosphate species and homogeneous MO6 monolayer. MO6 maintains 80 % removal and 96 % recovery after five cycles and adheres to the WHO and EUWFD regulations for residual elements in water. FT-IR and XPS analyses further reveal the underlying mechanisms, including ion exchange, electrostatic, and acid-base interactions. Ten machine learning (ML) models were applied to simultaneously predict multi-criteria (sorption capacity, removal efficiency, final pH, and Mg leakage) affected by 15 diverse environmental conditions. Traditional ML models and deep neural networks have poor accuracy, particularly for removal efficiency. However, a breakthrough was achieved by the developed deep belief network (DBN) with unparalleled performance (MAE = 1.3289, RMSE = 5.2552, R2 = 0.9926) across all output features, surpassing all current studies using thousands of data points for only one output factor. These captivating MO6 and DBN models also have immense potential for effectively applying in the real water test with error < 5 %, opening immense horizons for transformative methods, particularly in phosphate removal and recovery.

A computational investigation of DMSO/water separation through functionalized GO multilayer nanosheet membrane using molecular dynamics simulation and deep neural network model for membrane performance prediction

In this molecular dynamics (MD) simulation study, the separation of dimethyl sulfoxide (DMSO) from water was investigated using multilayer functionalized graphene oxide (GO) membranes. The GO nanosheets were modified with chemical groups (-F, -H) to alter their properties. The study analyzed the influence of pressure and functional groups on the separation rate. Additionally, a deep neural network (DNN) model was developed to predict membrane behavior under different conditions in water treatment processes. Results revealed that the fluorine-functionalized membrane exhibited higher permeation compared to the hydrogen-functionalized one, with potential of mean force (PMF) analysis indicating higher energy barriers for water molecules passing through the hydrogen-functionalized membrane. The study used density profile, water density map analysis, and radial distribution function (RDF) analysis to understand water and DMSO molecule interactions. The diffusion coefficient of water molecules was also calculated, showing higher diffusion in the fluorine-functionalized system. Overall, the findings suggest that functionalized GO membranes are effective for DMSO-water separation, with the fluorine-functionalized membrane showing superior performance. The DNN model accurately predicts membrane behavior, contributing to the optimization of membrane separation systems.

Deep Learning Algorithm Based on Molecular Fingerprint for Prediction of Drug-Induced Liver Injury

Drug-induced liver injury (DILI) is one the rare adverse drug reaction (ADR) and multifactorial endpoints. Current preclinical animal models struggle to anticipate it, and in silico methods have emerged as a way with significant potential for doing so. In this study, a high-quality dataset of 1573 compounds was assembled. The 48 classification models, which depended on six different molecular fingerprints, were built via deep neural network (DNN) and seven machine learning algorithms. Comparing the results of the DNN and machine learning models, the optional performing model was found as the one developed based on the DNN with ECFP_6 as input, which achieved the area under the receiver operating characteristic curve (AUC) of 0.713, balanced accuracy (BA) of 0.680, and F1 of 0.753. In addition, we used the SHapley Additive exPlanations (SHAP) algorithm to interpret the models, identified the crucial structural fragments related to DILI risk, and selected the top ten substructures with the highest contribution rankings to serve as warning indicators for subsequent drug hepatotoxicity screening studies. The study demonstrates that the DNN models developed based on molecular fingerprints can be a trustworthy and efficient tool for determining the risk of DILI during the pre-development of novel medications.

[Development and evaluation of ultrasound navigation for free-hand biopsies of small masses in the head and neck area]

BACKGROUND

Ultrasound is an important imaging method in the head and neck area. It is readily available, dynamic, inexpensive, and does not involve radiation exposure. Interventions in the complex head and neck anatomy require good orientation, which is supported by navigation systems.

OBJECTIVE

This work aimed to develop a new ultrasound-controlled navigation system for taking biopsies of small target structures in the head and neck region.

METHODS

A neck phantom with sonographically detectable masses (size: 8-10 mm) was constructed. These were automatically segmented using a ResNet-50-based deep neural network. The ultrasound scanner was equipped with an individually manufactured tracking tool.

RESULTS

The positions of the ultrasound device, the masses, and a puncture needle were recorded in the world coordinate system. In 8 out of 10 cases, an 8‑mm mass was hit. In a special evaluation phantom, the average deviation was calculated to be 2.5 mm. The tracked biopsy needle is aligned and navigated to the masses by auditory feedback.

CONCLUSION

Outstanding advantages compared to conventional navigation systems include renunciation of preoperative tomographic imaging, automatic three-dimensional real-time registration that considers intraoperative tissue displacements, maintenance of the surgeon's optical axis at the surgical site without having to look at a navigation monitor, and working freely with both hands without holding the ultrasound scanner during biopsy taking. The described functional model can also be used in open head and neck surgery.

Rapid estimation of electroporation-dependent tissue properties in canine lung tumors using a deep neural network

The efficiency of electroporation treatments depends on the application of a critical electric field over the targeted tissue volume. Both the electric field and temperature distribution strongly depend on the tissue-specific electrical properties, which both differ between patients in healthy and malignant tissues and change in an electric field-dependent manner from the electroporation process itself. Therefore, tissue property estimations are paramount for treatment planning with electroporation therapies. Ex vivo methods to find electrical tissue properties often misrepresent the targeted tissue, especially when translating results to tumors. A voltage ramp is an in situ method that applies a series of increasing electric potentials across treatment electrodes and measures the resulting current. Here, we develop a robust deep neural network, trained on finite element model simulations, to directly predict tissue properties from a measured voltage ramp. There was minimal test error (R2>0.94;p<0.0001) in three important electric tissue properties. Further, our model was validated to correctly predict the complete dynamic conductivity curve in a previously characterized ex vivo liver model (R2>0.93;p<0.0001) within 100 s from probe insertion, showing great utility for a clinical application. Lastly, we characterize the first reported electrical tissue properties of lung tumors from five canine patients (R2>0.99;p<0.0001). We believe this platform can be incorporated prior to treatment to quickly ascertain patient-specific tissue properties required for electroporation treatment planning models or real-time treatment prediction algorithms. Further, this method can be used over traditional ex vivo methods for in situ tissue characterization with clinically relevant geometries.

NOx emissions prediction in diesel engines: a deep neural network approach

The reduction of various nitrogen oxide (NOx) emissions from diesel engines is an important environmental issue due to their negative impact on air quality and public health. Selective catalytic reduction (SCR) has emerged as an effective technology to mitigate NOx emissions, but predicting the performance of SCR systems remains a challenge due to the complex chemistry involved. In this study, we propose using DNN models to predict NOx emission reductions in SCR systems. Four types of datasets were created; each consisted of five variables as inputs. We evaluated the models using experimental data collected from a diesel engine equipped with an SCR system. Our results indicated that the deep neural network (DNN) model produces precise estimates for exhaust gas temperature, NOx concentration, and De-NOx efficiency. Moreover, inclusion of additional input features, such as engine speed and temperature, improved the prediction accuracy of the DNN model. The mean absolute error (MAE) values for these parameters were 3.1 °C, 3.04 ppm, and 3.65%, respectively. Furthermore, the R-squared coefficient of determination values for the estimates were 0.912, 0.983, and 0.905, respectively. Overall, this study demonstrates the potential of using DNNs to accurately predict NOx emissions from diesel engines and provides insights into the impact of input features on the performance of the model.

TSANet: multibranch attention deep neural network for classifying tactile selective attention in brain-computer interfaces

Brain-computer interfaces (BCIs) enable communication between the brain and a computer and electroencephalography (EEG) has been widely used to implement BCIs because of its high temporal resolution and noninvasiveness. Recently, a tactile-based EEG task was introduced to overcome the current limitations of visual-based tasks, such as visual fatigue from sustained attention. However, the classification performance of tactile-based BCIs as control signals is unsatisfactory. Therefore, a novel classification approach is required for this purpose. Here, we propose TSANet, a deep neural network, that uses multibranch convolutional neural networks and a feature-attention mechanism to classify tactile selective attention (TSA) in a tactile-based BCI system. We tested TSANet under three evaluation conditions, namely, within-subject, leave-one-out, and cross-subject. We found that TSANet achieved the highest classification performance compared with conventional deep neural network models under all evaluation conditions. Additionally, we show that TSANet extracts reasonable features for TSA by investigating the weights of spatial filters. Our results demonstrate that TSANet has the potential to be used as an efficient end-to-end learning approach in tactile-based BCIs.

Supplementary Information

The online version contains supplementary material available at 10.1007/s13534-023-00309-4.

SetQuence & SetOmic: Deep set transformers for whole genome and exome tumour analysis

In oncology, Deep Learning has shown great potential to personalise tasks such as tumour type classification, based on per-patient omics data-sets. Being high dimensional, incorporation of such data in one model is a challenge, often leading to one-dimensional studies and, therefore, information loss. Instead, we first propose relying on non-fixed sets of whole genome or whole exome variant-associated sequences, which can be used for supervised learning of oncology-relevant tasks by our Set Transformer based Deep Neural Network, SetQuence. We optimise this architecture to improve its efficiency. This allows for exploration of not just coding but also non-coding variants, from large datasets. Second, we extend the model to incorporate these representations together with multiple other sources of omics data in a flexible way with SetOmic. Evaluation, using these representations, shows improved robustness and reduced information loss compared to previous approaches, while still being computationally tractable. By means of Explainable Artificial Intelligence methods, our models are able to recapitulate the biological contribution of highly attributed features in the tumours studied. This validation opens the door to novel directions in multi-faceted genome and exome wide biomarker discovery and personalised treatment among other presently clinically relevant tasks.

Artificial Intelligence-Based Counting Algorithm Enables Accurate and Detailed Analysis of the Broad Spectrum of Spot Morphologies Observed in Antigen-Specific B-Cell ELISPOT and FluoroSpot Assays

Antigen-specific B-cell ELISPOT and multicolor FluoroSpot assays, in which the membrane-bound antigen itself serves as the capture reagent for the antibodies that B cells secrete, inherently result in a broad range of spot sizes and intensities. The diversity of secretory footprint morphologies reflects the polyclonal nature of the antigen-specific B cell repertoire, with individual antibody-secreting B cells in the test sample differing in their affinity for the antigen, fine epitope specificity, and activation/secretion kinetics. To account for these heterogeneous spot morphologies, and to eliminate the need for setting up subjective counting parameters well-by-well, CTL introduces here its cutting-edge deep learning-based IntelliCount™ algorithm within the ImmunoSpot® Studio Software Suite, which integrates CTL's proprietary deep neural network. Here, we report detailed analyses of spots with a broad range of morphologies that were challenging to analyze using standard parameter-based counting approaches. IntelliCount™, especially in conjunction with high dynamic range (HDR) imaging, permits the extraction of accurate, high-content information of such spots, as required for assessing the affinity distribution of an antigen-specific memory B-cell repertoire ex vivo. IntelliCount™ also extends the range in which the number of antibody-secreting B cells plated and spots detected follow a linear function; that is, in which the frequencies of antigen-specific B cells can be accurately established. Introducing high-content analysis of secretory footprints in B-cell ELISPOT/FluoroSpot assays, therefore, fundamentally enhances the depth in which an antigen-specific B-cell repertoire can be studied using freshly isolated or cryopreserved primary cell material, such as peripheral blood mononuclear cells.

Machine learning and deep neural network-based learning in osteoarthritis knee

Osteoarthritis (OA) of the knee joint is considered the commonest musculoskeletal condition leading to marked disability for patients residing in various regions around the globe. Application of machine learning (ML) in doing research regarding OA has brought about various clinical advances viz, OA being diagnosed at preliminary stages, prediction of chances of development of OA among the population, discovering various phenotypes of OA, calculating the severity in OA structure and also discovering people with slow and fast progression of disease pathology, etc. Various publications are available regarding machine learning methods for the early detection of osteoarthritis. The key features are detected by morphology, molecular architecture, and electrical and mechanical functions. In addition, this particular technique was utilized to assess non-interfering, non-ionizing, and in-vivo techniques using magnetic resonance imaging. ML is being utilized in OA, chiefly with the formulation of large cohorts viz, the OA Initiative, a cohort observational study, the Multi-centre Osteoarthritis Study, an observational, prospective longitudinal study and the Cohort Hip & Cohort Knee, an observational cohort prospective study of both hip and knee OA. Though ML has various contributions and enhancing applications, it remains an imminent field with high potential, also with its limitations. Many more studies are to be carried out to find more about the link between machine learning and knee osteoarthritis, which would help in the improvement of making decisions clinically, and expedite the necessary interventions.

Shortening image registration time using a deep neural network for patient positional verification in radiotherapy

We sought to accelerate 2D/3D image registration computation time using image synthesis with a deep neural network (DNN) to generate digitally reconstructed radiographic (DRR) images from X-ray flat panel detector (FPD) images. And we explored the feasibility of using our DNN in the patient setup verification application. Images of the prostate and of the head and neck (H&N) regions were acquired by two oblique X-ray fluoroscopic units and the treatment planning CT. DNN was designed to generate DRR images from the FPD image data. We evaluated the quality of the synthesized DRR images to compare the ground-truth DRR images using the peak signal-to-noise ratio (PSNR) and structural similarity index measure (SSIM). Image registration accuracy and computation time were evaluated by comparing the 2D-3D image registration algorithm using DRR and FPD image data with DRR and synthesized DRR images. Mean PSNR values were 23.4 ± 3.7 dB and 24.1 ± 3.9 dB for the pelvic and H&N regions, respectively. Mean SSIM values for both cases were also similar (= 0.90). Image registration accuracy was degraded by a mean of 0.43 mm and 0.30°, it was clinically acceptable. Computation time was accelerated by a factor of 0.69. Our DNN successfully generated DRR images from FPD image data, and improved 2D-3D image registration computation time up to 37% in average.