Aesthetic Plast Surg. 2025 May 27. doi: 10.1007/s00266-025-04938-1. Online ahead of print.

NO ABSTRACT

PMID:40425882 | DOI:10.1007/s00266-025-04938-1

Nat Med. 2025 May 27. doi: 10.1038/s41591-025-03715-6. Online ahead of print.

ABSTRACT

There is limited evidence supporting the feasibility of using omics and functional technologies to inform treatment decisions. Here we present results from a cohort of 116 melanoma patients in the prospective, multicentric observational Tumor Profiler (TuPro) precision oncology project. Nine independent technologies, mostly at single-cell level, were used to analyze 126 patient samples, generating up to 500 Gb of data per sample (40,000 potential markers) within 4 weeks. Among established and experimental markers, the molecular tumor board selected 54 to inform its treatment recommendations. In 75% of cases, TuPro-based data were judged to be useful in informing recommendations. Patients received either standard of care (SOC) treatments or highly individualized, polybiomarker-driven treatments (beyond SOC). The objective response rate in difficult-to-treat palliative, beyond SOC patients (n = 37) was 38%, with a disease control rate of 54%. Progression-free survival of patients with TuPro-informed therapy decisions was 6.04 months, (95% confidence interval, 3.75-12.06) and 5.35 months (95% confidence interval, 2.89-12.06) in ≥third therapy lines. The proof-of-concept TuPro project demonstrated the feasibility and relevance of omics-based tumor profiling to support data-guided clinical decision-making. ClinicalTrials.gov identifier: NCT06463509 .

PMID:40425842 | DOI:10.1038/s41591-025-03715-6

Sci Rep. 2025 May 27;15(1):18556. doi: 10.1038/s41598-025-03010-x.

ABSTRACT

Alzheimer's disease (AD) is the most prevalent type of dementia. It is linked with a gradual decline in various brain functions, such as memory. Many research efforts are now directed toward non-invasive procedures for early diagnosis because early detection greatly benefits the patient care and treatment outcome. Additional to an accurate diagnosis and reduction of the rate of misdiagnosis; Computer-Aided Design (CAD) systems are built to give definitive diagnosis. This paper presents a novel CAD system to determine stages of AD. Initially, deep learning techniques are utilized to extract features from the AD brain MRIs. Then, the extracted features are reduced using a proposed feature reduction technique named Deep Downsized Kernel Partial Least Squares (DDKPLS). The proposed approach selects a reduced number of samples from the initial information matrix. The samples chosen give rise to a new data matrix further processed by KPLS to deal with the high dimensionality. The reduced feature space is finally classified using ELM. The implementation is named DDKPLS-ELM. Reference tests have been performed on the Kaggle MRI dataset, which exhibit the efficacy of the DDKPLS-based classifier; it achieves accuracy up to 95.4% and an F1 score of 95.1%.

PMID:40425715 | DOI:10.1038/s41598-025-03010-x

Sci Rep. 2025 May 27;15(1):18552. doi: 10.1038/s41598-025-02739-9.

ABSTRACT

Markerless motion capture (ML) systems, which utilize deep learning algorithms, have significantly expanded the applications of biomechanical analysis. Jump tests are now essential tools for athlete monitoring and injury prevention. However, the validity of kinematic and kinetic parameters derived from ML for lower limb joints requires further validation in populations engaged in high-intensity jumping sports. The purpose of this study was to compare lower limb kinematic and kinetic estimates between marker-based (MB) and ML motion capture systems during jumps. Fourteen male Division I movement collegiate athletes performed a minimum of three squat jumps (SJ), drop jumps (DJ), and countermovement jumps (CMJ) in a fixed sequence. The movements were synchronized using ten infrared cameras, six high-resolution cameras, and two force measurement platforms, all controlled by Vicon Nexus software. Motion data were collected, and the angles, moments, and power at the hip, knee, and ankle joints were calculated using Theia3D software. These results were then compared with those obtained from the Vicon system. Comparative analyses included Pearson correlation coefficients (r), root mean square differences (RMSD), extreme error values, and statistical parametric mapping (SPM).SPM analysis of the three movements in the sagittal plane revealed significant differences in hip joint angles, with joint angle RMSD ≤ 5.6°, hip joint moments RMSD ≤ 0.26 N·M/kg, and power RMSD ≤ 2.12 W/kg showing considerable variation, though not reaching statistical significance. ML systems demonstrate high measurement accuracy in estimating knee and ankle kinematics and kinetics in the sagittal plane during these conventional jump tests; however, the accuracy of hip joint kinematic measurements in the sagittal plane requires further validation.

PMID:40425708 | DOI:10.1038/s41598-025-02739-9

PLoS One. 2025 May 27;20(5):e0324293. doi: 10.1371/journal.pone.0324293. eCollection 2025.

ABSTRACT

Cotton is a major cash crop, and increasing its production is extremely important worldwide, especially in agriculture-led economies. The crop is susceptible to various diseases, leading to decreased yields. In recent years, advancements in deep learning methods have enabled researchers to develop automated methods for detecting diseases in cotton crops. Such automation not only assists farmers in mitigating the effects of the disease but also conserves resources in terms of labor and fertilizer costs. However, accurate classification of multiple diseases simultaneously in cotton remains challenging due to multiple factors, including class imbalance, variation in disease symptoms, and the need for real-time detection, as most existing datasets are acquired under controlled conditions. This research proposes a novel method for addressing these challenges and accurately classifying seven classes, including six diseases and a healthy class. We address the class imbalance issue through synthetic data generation using conventional methods like scaling, rotating, transforming, shearing, and zooming and propose a customized StyleGAN for synthetic data generation. After preprocessing, we combine features extracted from MobileNet and VGG16 to create a comprehensive feature vector, passed to three classifiers: Long Short Term Memory Units, Support Vector Machines, and Random Forest. We propose a StackNet-based ensemble classifier that takes the output probabilities of these three classifiers and predicts the class label among six diseases-Bacterial blight, Curl virus, Fusarium wilt, Alternaria, Cercospora, Greymildew-and a healthy class. We trained and tested our method on publicly available datasets, achieving an average accuracy of 97%. Our robust method outperforms state-of-the-art techniques to identify the six diseases and the healthy class.

PMID:40424461 | DOI:10.1371/journal.pone.0324293

Bioinformatics. 2025 May 27:btaf319. doi: 10.1093/bioinformatics/btaf319. Online ahead of print.

ABSTRACT

MOTIVATION: Drug side effects refer to harmful or adverse reactions that occur during drug use, unrelated to the therapeutic purpose. A core issue in drug side effect prediction is determining the frequency of these drug side effects in the population, which can guide patient medication use and drug development. Many computational methods have been developed to predict the frequency of drug side effects as an alternative to clinical trials. However, existing methods typically build regression models on five frequency classes of drug side effects and tend to overfit the training set, leading to boundary handling issues and the risk of overfitting.

RESULTS: To address this problem, we develop a multi-source similarity fusion-based model, named MSSF, for predicting five frequency classes of drug side effects. Compared to existing methods, our model utilizes the multi-source feature fusion module and the self-attention mechanism to explore the relationships between drugs and side effects deeply and employs Bayesian variational inference to more accurately predict the frequency classes of drug side effects. The experimental results indicate that MSSF consistently achieves superior performance compared to existing models across multiple evaluation settings, including cross-validation, cold-start experiments, and independent testing. The visual analysis and case studies further demonstrate MSSF's reliable feature extraction capability and promise in predicting the frequency classes of drug side effects.

AVAILABILITY: The source code of MSSF is available on GitHub (https://github.com/dingxlcse/MSSF.git) and archived on Zenodo (DOI: 10.5281/zenodo.15462041).

SUPPLEMENTARY INFORMATION: Additional files are available at Bioinformatics online.

PMID:40424358 | DOI:10.1093/bioinformatics/btaf319

PLoS One. 2025 May 27;20(5):e0322225. doi: 10.1371/journal.pone.0322225. eCollection 2025.

ABSTRACT

With the digital transformation of the financial industry, credit score prediction, as a key component of risk management, faces increasingly complex challenges. Traditional credit scoring methods often have difficulty in fully capturing the characteristics of large-scale, high-dimensional financial data, resulting in limited prediction performance. To address these issues, this paper proposes a credit score prediction model that combines CNNs and BiGRUs, and uses the GWO algorithm for hyperparameter tuning. CNN performs well in feature extraction and can effectively capture patterns in customer historical behaviors, while BiGRU is good at handling time dependencies, which further improves the prediction accuracy of the model. The GWO algorithm is introduced to further improve the overall performance of the model by optimizing key parameters. Experimental results show that the CNN-BiGRU-GWO model proposed in this paper performs well on multiple public credit score datasets, significantly improving the accuracy and efficiency of prediction. On the LendingClub loan dataset, the MAE of this model is 15.63, MAPE is 4.65%, RMSE is 3.34, and MSE is 12.01, which are 64.5%, 68.0%, 21.4%, and 52.5% lower than the traditional method plawiak of 44.07, 14.51%, 4.25, and 25.29, respectively. In addition, compared with traditional methods, this model also shows stronger advantages in adaptability and generalization ability. By integrating advanced technologies, this model not only provides an innovative technical solution for credit score prediction, but also provides valuable insights into the application of deep learning in the financial field, making up for the shortcomings of existing methods and demonstrating its potential for wide application in financial risk management.

PMID:40424348 | DOI:10.1371/journal.pone.0322225

PLoS One. 2025 May 27;20(5):e0322555. doi: 10.1371/journal.pone.0322555. eCollection 2025.

ABSTRACT

Human Action Recognition (HAR) has grown significantly because of its many uses, including real-time surveillance and human-computer interaction. Various variations in routine human actions make the recognition process of action more difficult. In this paper, we proposed a novel deep learning architecture known as Inverted Bottleneck Residual with Self-Attention (InBRwSA). The proposed architecture is based on two different modules. In the first module, 6-parallel inverted bottleneck residual blocks are designed, and each block is connected with a skip connection. These blocks aim to learn complex human actions in many convolutional layers. After that, the second module is designed based on the self-attention mechanism. The learned weights of the first module are passed to self-attention, extract the most essential features, and can easily discriminate complex human actions. The proposed architecture is trained on the selected datasets, whereas the hyperparameters are chosen using the particle swarm optimization (PSO) algorithm. The trained model is employed in the testing phase for the feature extraction from the self-attention layer and passed to the shallow wide neural network classifier for the final classification. The HMDB51 and UCF 101 are frequently used as action recognition standard datasets. These datasets are chosen to allow for meaningful comparison with earlier research. UCF101 dataset has a wide range of activity classes, and HMDB51 has varied real-world behaviors. These features test the generalizability and flexibility of the presented model. Moreover, these datasets define the evaluation scope within a particular domain and guarantee relevance to real-world circumstances. The proposed technique is tested on both datasets, and accuracies of 78.80% and 91.80% were achieved, respectively. The ablation study demonstrated that a margin of error value of 70.1338 ± 3.053 (±4.35%) and 82.7813 ± 2.852 (±3.45%) for the confidence level 95%,1.960σx̄ is obtained for HMDB51 and UCF datasets respectively. The training time for the highest accuracy for HDMB51 and UCF101 is 134.09 and 252.10 seconds, respectively. The proposed architecture is compared with several pre-trained deep models and state-of-the-art (SOTA) existing techniques. Based on the results, the proposed architecture outperformed existing techniques.

PMID:40424287 | DOI:10.1371/journal.pone.0322555

PLoS One. 2025 May 27;20(5):e0324916. doi: 10.1371/journal.pone.0324916. eCollection 2025.

ABSTRACT

Traditional communication signal detection heavily relies on manually designed features, making it difficult to fully characterize the essential characteristics of the signal, resulting in limited detection accuracy. Based on this, the study innovatively combines Multiple Input Multiple Output (MIMO) with orthogonal frequency division multiplexing technology to construct a data-driven detection system. The system adopts a Multi-DNN method with a dual-DNN cascade structure and mixed activation function design to optimize the channel estimation and signal detection coordination process of the MIMO part. At the same time, a DCNet decoder based on a convolutional neural network batch normalization mechanism is designed to suppress inter-subcarrier interference in OFDM systems effectively. The results showed that on the simulation training set, the accuracy of the research model was 93.8%, the symbol error rate was 17.6%, the throughput was 81.3%, and the modulation error rate was 0.004%. On the simulation test set, its accuracy, symbol error rate, throughput, and modulation error rate were 90.7%, 18.1%, 81.2%, and 0.006%. In both 2.4 GHz and 5 GHz WiFi signals, the signal detection accuracy of the research model reached 91.5% and 91.6%, with false detection rates of 1.9% and 1.5%, and missed detection rates of 1.6% and 4.2%. In resource consumption assessment, the detection speed of this model reached 120 signals/s, with an average latency of 50 ms. The model loading time was only 2.4 s, and the CPU usage was as low as 25%, with moderate memory usage. Overall, the research model has achieved significant results in improving detection accuracy, optimizing real-time performance, and reducing resource consumption. It has broad application prospects in the field of communication signal detection.

PMID:40424260 | DOI:10.1371/journal.pone.0324916

PLoS One. 2025 May 27;20(5):e0321270. doi: 10.1371/journal.pone.0321270. eCollection 2025.

ABSTRACT

Deep learning techniques are widely used in the field of medicine and image classification. In past studies, SwimTransformer and RepVGG are very efficient and classical deep learning models. Multi-scale feature fusion and attention mechanisms are effective means to enhance the performance of deep learning models. In this paper, we introduce a novel Swim-Rep fusion network, along with a new multi-scale feature fusion module called multi-scale strip pooling fusion module(MPF) and a new attention module called Faster Recurrent Criss Cross Polarized Attention (FRCPA), both of which excel at extracting multi-dimensional cross-attention and fine-grained features. Our fully supervised model achieved an impressive accuracy of 99.82% on the MIT-BIH database, outperforming the ViT model classifier by 0.12%. Additionally, our semi-supervised model demonstrated strong performance, achieving 98.4% accuracy on the validation set. Experimental results on the remote sensing image classification dataset RSSCN7 demonstrate that our new base model achieves a classification accuracy of 92.5%, which is 8.57% better than the classification performance of swim-transformer-base and 12.9% better than that of RepVGG-base, and increasing the depth of the module yields superior performance.

PMID:40424251 | DOI:10.1371/journal.pone.0321270

PLoS One. 2025 May 27;20(5):e0324595. doi: 10.1371/journal.pone.0324595. eCollection 2025.

ABSTRACT

A zero-day vulnerability is a critical security weakness of software or hardware that has not yet been found and, for that reason, neither the vendor nor the users are informed about it. These vulnerabilities may be taken advantage of by malicious people to execute cyber-attacks leading to severe effects on organizations and individuals. Given that nobody knows and is aware of these weaknesses, it becomes challenging to detect and prevent them. For the real-time zero-day vulnerabilities detection, we bring out a novel reinforcement learning (RL) methodology with the help of Deep Q-Networks (DQN). It works by learning the vulnerabilities without any prior knowledge of vulnerabilities, and it is evaluated using rigorous statistical metrics. Traditional methods are surpassed by this one that is able to adjust to changing threats and cope with intricate state spaces while providing scalability to cybersecurity personnel. In this paper, we introduce a new methodology that uses reinforcement learning for zero-day vulnerability detection. Zero-day vulnerabilities are security weaknesses that have never been exposed or published and are considered highly dangerous for systems and networks. Our method exploits reinforcement learning, a sub-type of machine learning which trains agents to make decisions and take actions to maximize an approximation of some underlying cumulative reward signal and discover patterns and features within data related to zero-day discovery. Training of the agent could allow for real-time detection and classification of zero-day vulnerabilities. Our approach will have the potential as a powerful tool of detection and defense against zero-day vulnerabilities and probably brings significant benefits to security experts and researchers in the field of cyber-security. The new method of discovering vulnerabilities that this approach provides has many comparative advantages over the previous approaches. It is applicable to systems with complex behaviour, such as the ones presented throughout this thesis, and can respond to new security threats in real time. Moreover, it does not require any knowledge about vulnerability itself. Because of that, it will discover hidden weak points. In the present paper, we analyzed the statistical evaluation of forecasted values for several parameters in a reinforcement learning environment. We have taken 1000 episodes for training the model and a further 1000 episodes for forecasting using the trained model. We used statistical measures in the evaluation, which showed that the Alpha value was at 0.10, thereby indicating good accuracy in the forecast. Beta was at 0.00, meaning no bias within the forecast. Gamma was also at 0.00, resulting in a very high level of precision within the forecast. MASE was 3.91 and SMAPE was 1.59, meaning that a very minimal percentage error existed within the forecast. The MAE value was at 6.34, while the RMSE was 10.22, meaning a relatively low average difference within actuals and the forecasted values. Results The results demonstrate the effectiveness of reinforcement learning models in solving complex problems and suggest that the model improves in accuracy with more training data added.

PMID:40424227 | DOI:10.1371/journal.pone.0324595

PLoS One. 2025 May 27;20(5):e0323951. doi: 10.1371/journal.pone.0323951. eCollection 2025.

ABSTRACT

[This corrects the article DOI: 10.1371/journal.pone.0305859.].

PMID:40424208 | DOI:10.1371/journal.pone.0323951

J Sci Food Agric. 2025 May 27. doi: 10.1002/jsfa.14386. Online ahead of print.

ABSTRACT

BACKGROUND: In the process of tobacco sorting, removing non-tobacco related materials (NTRMs) is crucial for the quality of tobacco products. Because of the small size of NTRMs and the abundance and stacking of tobacco leaves, detection of NTRMs is still difficult.

RESULTS: Based on YOLOv8s (You Only Look Once, version 8, small), the present study proposed an efficient YSFER-Tobacco (YOLOv8s-SPDConv-FasterNet-EMA-RTDETRDecoder-Tobacco) model for detection of NTRMs. We replaced some down sampling convolutions in the backbone with SPDConv modules and reconstructed the C2f module using FasterNet and EMA to reduce redundant convolution operations and improve feature extraction capabilities. Finally, the RTDETRDecoder from RT-DETR was employed to improve the head component, resulting in more efficient end-to-end target identification. Experimental results demonstrate that YSFER-Tobacco achieved good model performance, with F1, mAP50, recall and precision reaching 96.1%, 97.2%, 95.7% and 96.5%, respectively, compared to YOLOv8s, which increased by 2.5%, 0.9%, 3.2% and 1.7%. YSFER-Tobacco also outperformed other classical object detection models for detection of NTRMs in tobacco sorting process.

CONCLUSION: Our study demonstrates the effectiveness and superiority of YSFER-Tobacco, providing theoretical support for assessing the quality of tobacco sorting, and has promising application prospects. In addition, we replicated the tobacco sorting environment and created the first dataset, Tobacco-2619, containing 2619 clear images with NTRMs. The dataset and code are available online (https://github.com/Ikaros-sc/Tobacco). © 2025 Society of Chemical Industry.

PMID:40424192 | DOI:10.1002/jsfa.14386

Elife. 2025 May 27;13:RP94029. doi: 10.7554/eLife.94029.

ABSTRACT

Despite the recent breakthrough of AlphaFold (AF) in the field of protein sequence-to-structure prediction, modeling protein interfaces and predicting protein complex structures remains challenging, especially when there is a significant conformational change in one or both binding partners. Prior studies have demonstrated that AF-multimer (AFm) can predict accurate protein complexes in only up to 43% of cases (Yin et al., 2022). In this work, we combine AF as a structural template generator with a physics-based replica exchange docking algorithm to better sample conformational changes. Using a curated collection of 254 available protein targets with both unbound and bound structures, we first demonstrate that AF confidence measures (pLDDT) can be repurposed for estimating protein flexibility and docking accuracy for multimers. We incorporate these metrics within our ReplicaDock 2.0 protocol to complete a robust in silico pipeline for accurate protein complex structure prediction. AlphaRED (AlphaFold-initiated Replica Exchange Docking) successfully docks failed AF predictions, including 97 failure cases in Docking Benchmark Set 5.5. AlphaRED generates CAPRI acceptable-quality or better predictions for 63% of benchmark targets. Further, on a subset of antigen-antibody targets, which is challenging for AFm (20% success rate), AlphaRED demonstrates a success rate of 43%. This new strategy demonstrates the success possible by integrating deep learning-based architectures trained on evolutionary information with physics-based enhanced sampling. The pipeline is available at https://github.com/Graylab/AlphaRED.

PMID:40424178 | DOI:10.7554/eLife.94029

Med Biol Eng Comput. 2025 May 27. doi: 10.1007/s11517-025-03381-3. Online ahead of print.

ABSTRACT

Traditional computed tomography (CT) methods for 3D reconstruction face resolution limitations and require time-consuming post-processing workflows. While deep learning techniques improve the accuracy of segmentation, traditional voxel-based segmentation and surface reconstruction pipelines tend to introduce artifacts such as disconnected regions, topological inconsistencies, and stepped distortions. To overcome these challenges, we propose ToPoMesh, an end-to-end 3D mesh reconstruction deep learning framework for direct reconstruction of high-fidelity surface meshes from CT volume data. To address the existing problems, our approach introduces three core innovations: (1) accurate local and global shape modeling by preserving and enhancing local feature information through residual connectivity and self-attention mechanisms in graph convolutional networks; (2) an adaptive variant density (Avd) mesh de-pooling strategy, which dynamically optimizes the vertex distribution; (3) a topology modification module that iteratively prunes the error surfaces and boundary smoothing via variable regularity terms to obtain finer mesh surfaces. Experiments on the LiTS, MSD pancreas tumor, MSD hippocampus, and MSD spleen datasets demonstrate that ToPoMesh outperforms state-of-the-art methods. Quantitative evaluations demonstrate a 57.4% reduction in Chamfer distance (liver) and a 0.47% improvement in F-score compared to end-to-end 3D reconstruction methods, while qualitative results confirm enhanced fidelity for thin structures and complex anatomical topologies versus segmentation frameworks. Importantly, our method eliminates the need for manual post-processing, realizes the ability to reconstruct 3D meshes from images, and can provide precise guidance for surgical planning and diagnosis.

PMID:40423893 | DOI:10.1007/s11517-025-03381-3

Epilepsia. 2025 May 27. doi: 10.1111/epi.18478. Online ahead of print.

ABSTRACT

OBJECTIVE: Epilepsy is a debilitating disorder affecting more than 50 million people worldwide, and one third of patients continue to have seizures despite maximal medical management. If patients' seizures localize to a discrete brain region, termed a seizure onset zone, resection may be curative. Localization is often confirmed with stereotactic electroencephalography; however, this may require patients to stay in the hospital for weeks to capture spontaneous seizures. Automated localization of seizure onset zones could therefore improve presurgical evaluation and decrease morbidity.

METHODS: Using more than 1 000 000 interictal stereotactic electroencephalography segments collected from 78 patients, we performed five-fold cross-validation and testing on a multichannel, multiscale, one-dimensional convolutional neural network to classify seizure onset zones.

RESULTS: Across held-out test sets, our models achieved a seizure onset zone classification sensitivity of .702 (95% confidence interval [CI] = .549-.805), specificity of .741 (95% CI = .652-.835), and accuracy of .738 (95% CI = .687-.795), which was significantly better than models trained on random labels. The models performed well across the entire brain, with top five region performance demonstrating accuracies between 70.0% and 88.4%. When split by outcomes, the models performed significantly better on patients with favorable Engel outcomes after resection or who were responsive neurostimulation responders. Finally, SHAP (Shapley Additive Explanation) value analysis on median-normalized input data assigned consistently high feature importance to interictal spikes and large deflections, whereas similar analyses on histogram-equalized data revealed differences in feature importance assignments to low-amplitude segments.

SIGNIFICANCE: This work serves as evidence that deep learning on brief interictal intracranial data can classify seizure onset zones across the brain. Furthermore, our findings corroborate current understandings of interictal epileptiform discharges and may help uncover novel interictal morphologies. Clinical application of our models may reduce dependence on recorded seizures for localization and shorten presurgical evaluation time for drug-resistant epilepsy patients, reducing patient morbidity and hospital costs.

PMID:40423629 | DOI:10.1111/epi.18478

Radiology. 2025 May;315(2):e259008. doi: 10.1148/radiol.259008.

NO ABSTRACT

PMID:40423544 | DOI:10.1148/radiol.259008

Toxics. 2025 Apr 27;13(5):349. doi: 10.3390/toxics13050349.

ABSTRACT

The operational complexity of wastewater treatment systems mainly stems from the diversity of influent characteristics and the nonlinear nature of the treatment process. Together, these factors make the control of effluent quality in wastewater treatment plants (WWTPs) difficult to manage effectively. To address this challenge, constructing accurate effluent quality models for WWTPs can not only mitigate these complexities, but also provide critical decision support for operational management. In this research, we introduce a deep learning method that fuses multi-source data. This method utilises various indicators to comprehensively analyse and predict the quality of effluent water: water quantity data, process data, energy consumption data, and water quality data. To assess the efficacy of this method, a case study was carried out at an industrial effluent treatment plant (IETP) in Anhui Province, China. Deep learning algorithms including long short-term memory (LSTM) and gated recurrent unit (GRU) were found to have a favourable prediction performance by comparing with traditional machine learning algorithms (random forest, RF) and multi-layer perceptron (MLP). The results show that the R2 of LSTM and GRU is 1.36%~31.82% higher than that of MLP and 9.10%~47.75% higher than that of traditional machine learning algorithms. Finally, the RReliefF approach was used to identify the key parameters affecting the water quality behaviour of IETP effluent, and it was found that, by optimising the multi-source feature structure, not only the monitoring and management strategies can be optimised, but also the modelling efficiency of the model can be further improved.

PMID:40423427 | DOI:10.3390/toxics13050349

Toxics. 2025 Apr 23;13(5):327. doi: 10.3390/toxics13050327.

ABSTRACT

To address the performance degradation in existing PM2.5 prediction models caused by excessive complexity, poor spatiotemporal efficiency, and suboptimal parameter optimization, we employ stacking ensemble learning for feature weighting analysis and integrate the ant colony optimization (ACO) algorithm for model parameter optimization. Combining meteorological and collaborative pollutant data, a model (namely the stacking-ACO-LSTM model) with a much shorter consuming time than that of only long short-term memory (LSTM) networks suitable for PM2.5 concentration prediction is established. It can effectively filter out feature variables with higher weights, thereby reducing the predictive power of the model. The prediction of hourly PM2.5 concentration of the model is trained and tested using real-time monitoring data in Nanchang City from 2017 to 2019. The results show that the established stacking-ACO-LSTM model has high accuracy in predicting PM2.5 concentration, and compared to the same model without considering time and space efficiency and defective parameter optimization, the mean square error (MSE) decreases by about 99.88%, and the coefficient of determination (R2) increases by about 2.39%. This study provides a new idea for predicting PM2.5 concentration in cities.

PMID:40423406 | DOI:10.3390/toxics13050327

Toxins (Basel). 2025 Apr 27;17(5):219. doi: 10.3390/toxins17050219.

ABSTRACT

Cereal grains and nuts are the world's most produced food and the economic backbone of many countries. Food safety in these commodities is crucial, as they are highly susceptible to mold growth and mycotoxin contamination in warm, humid environments. This review explores hyperspectral imaging (HSI) integrated with machine learning (ML) algorithms as a promising approach for detecting and quantifying mycotoxins in cereal grains and nuts. This study aims to (1) critically evaluate current non-destructive techniques for processing these foods and the applications of ML in identifying mycotoxins through HSI, and (2) highlight challenges and potential future research directions to enhance the reliability and efficiency of these detection systems. The ML algorithms showed effectiveness in classifying and quantifying mycotoxins in grains and nuts, with HSI systems increasingly adopted in industrial settings. Mycotoxins exhibit heightened sensitivity to specific spectral bands within HSI, facilitating accurate detection. Additionally, selecting only relevant spectral features reduces ML model complexity and enhances reliability in the detection process. This review contributes to a deeper understanding of the integration of HSI and ML for food safety applications in cereal grains and nuts. By identifying current challenges and future research directions, it provides valuable insights for advancing non-destructive mycotoxin detection methods in the food industry using HSI.

PMID:40423302 | DOI:10.3390/toxins17050219