Search Results (112)

Speech emotion recognition has become increasingly important in a wide range of applications, driven by the development of large transformer-based natural language processing models. However, the large size of these architectures limits their usability, which has led to a growing interest in smaller models. In this paper, we evaluate nineteen of the most popular small language models for the text and audio modalities for speech emotion recognition on the IEMOCAP dataset. Based on their cross-validation accuracy, the best architectures were selected to create ensemble models to evaluate the effect of combining audio and text, as well as the effect of incorporating contextual information on model performance. The experiments conducted showed a significant increase in accuracy with the inclusion of contextual information and the combination of modalities. The results obtained were highly competitive, outperforming numerous recent approaches. The proposed ensemble model achieved an accuracy of 82.12% on the IEMOCAP dataset, outperforming several recent approaches. These results demonstrate the effectiveness of ensemble methods for improving speech emotion recognition performance, and highlight the feasibility of training multiple small language models on consumer-grade computers. Full article

Background: Diabetic retinopathy (DR) is a leading cause of visual impairment worldwide. Manual grading of fundus images is the gold standard in DR screening, although it is time-consuming. Artificial intelligence (AI)-based algorithms offer a faster alternative, though concerns remain about their diagnostic reliability. Methods: A cross-sectional pilot study among patients (≥18 years) with diabetes was established for DR and diabetic macular edema (DME) screening at the Oslo University Hospital (OUH), Department of Ophthalmology, and the Norwegian Association of the Blind and Partially Sighted (NABP). The aim of the study was to evaluate the validity (accuracy, sensitivity, specificity) and reliability (inter-rater agreement) of automated AI-based compared to manual consensus (MC) grading of DR and DME, performed by a multidisciplinary team of healthcare professionals. Grading of DR and DME was performed manually and by EyeArt (Eyenuk) software version v2.1.0, based on the International Clinical Disease Severity Scale (ICDR) for DR. Agreement was measured by Quadratic Weighted Kappa (QWK) and Cohen’s Kappa (κ). Sensitivity, specificity, and diagnostic test accuracy (Area Under the Curve (AUC)) were also calculated. Results: A total of 128 individuals (247 eyes) (51 women, 77 men) were included, with a median age of 52.5 years. Prevalence of any vs. referable DR (RDR) was 20.2% vs. 11.7%, while sensitivity was 94.0% vs. 89.7%, specificity was 72.6% was 83.0%, and AUC was 83.5% vs. 86.3%, respectively. DME was detected only in one eye by both methods. Conclusions: AI-based grading offered high sensitivity and acceptable specificity for detecting DR, showing moderate agreement with manual assessments. Such grading may serve as an effective screening tool to support clinical evaluation, while ongoing training of human graders remains essential to ensure high-quality reference standards for accurate diagnostic accuracy and the development of AI algorithms. Full article

Specific computational methods, such as machine learning algorithms, can assist mining professionals in quickly and consistently identifying and addressing classification issues related to mineralized horizons, as well as uncovering key variables that impact predictive outcomes, many of which were previously difficult to observe. The integration of numerical and categorical variables, which are part of a dataset for defining ore grades, is part of the daily routine of professionals who obtain the data and manipulate the various phases of analysis in a mining project. Several supervised and unsupervised machine learning methods and applications integrate a wide variety of algorithms that aim at the efficient recognition of patterns and similarities and the ability to make accurate and assertive decisions. The objective of this study is the classification of gold ore or gangue through supervised machine learning methods using numerical variables represented by grade and categorical variables obtained through drillholes descriptions. Four groups of variables were selected with different variable configurations. The application of classification algorithms to different groups of variables aimed to observe the variables of importance and the impact of each one on the classification, in addition to testing the best algorithm in terms of accuracy and precision. The datasets were subjected to training, validation, and testing using the decision tree, random forest, Adaboost, XGBoost, and logistic regression methods. The evaluation was randomly divided into training (60%) and testing (40%) with 10-fold cross-validation. The results revealed that the XGBoost algorithm obtained the best performance, with an accuracy of 0.96 for scenario C1. In the SHAP analysis, the variable As was prominent in the predictions, mainly in scenarios C1 and C3. The arsenic class (Class_As), present mainly in scenario C4, had a significant positive weight in the classification. In the Receiver Operating Characteristic (ROC) and Area Under the Curve (AUC) curves, the results showed that XGBoost/scenario C1 obtained the highest AUC of 0.985, indicating that the algorithm had the best performance in ore/gangue classification of the sample set. The logistic regression algorithm together with AdaBoost had the worst performance, also varying between scenarios. Full article

Medical education is inherently demanding, requiring students to balance intense academic workload, clinical training, and emotional resilience. High levels of stress and burnout among medical students have been associated with decreased empathy, poorer academic performance, and increased risk of mental health problems. This cross-sectional, correlational study examined the relationships between emotional intelligence (EI), perceived stress, and burnout among undergraduate medical students at the Alexandria Faculty of Medicine. Participants completed self-report questionnaires: the Mind Tools Emotional Intelligence Test, the Perceived Stress Scale, and the Maslach Burnout Inventory. Descriptive statistics, bivariate correlations, and multivariate regression models were used for analysis. Among the 264 participants (88% response rate), the majority (73.4%) demonstrated average EI with no statistically significant differences across gender and academic year. Higher perceived stress was strongly correlated with emotional exhaustion and depersonalization, and it was also inversely correlated with personal accomplishment. Regression analysis indicated that gender, academic year, and academic grade were not independent predictors of stress or burnout (R² = 0.054). Approximately 30.3% of the students met the criteria for burnout. These findings highlight the complex interplay between emotional functioning and burnout, and they also suggest that interventions targeting emotional regulation and resilience may be beneficial in reducing stress and promoting well-being among medical students. Full article

The aim of this study was to develop a baseline [¹⁸F]FDG PET/CT model to predict immunotherapy response in advanced cutaneous squamous cell carcinoma (cSCC) and noninvasively determine tumor grade, thereby enhancing early patient stratification. We retrospectively analyzed 59 patients with histologically confirmed advanced cSCC submitted to immunotherapy with cemiplimab. All underwent [¹⁸F]FDG PET/CT at baseline and after approximately 12 weeks. Clinical response was assessed through PET findings integrated with clinical and dermatological evaluation, and patients were classified as responders (complete/partial metabolic response or stable disease) or non-responders (progression or toxicity-related discontinuation). Tumors were also classified as low to intermediate (G1–G2) or poorly differentiated (G3). Machine learning models (Random Forest and Extreme Gradient Boosting) were trained to predict treatment response and tumor grade. Clinical benefit was observed in 46/59 patients (77.9%), while 13 (22.1%) were non-responders. Histology showed 64.4% (n = 38) G1–G2 and 35.6% (n = 21) G3 tumors. The PET-based model best predicted clinical benefit (AUC = 0.96, accuracy = 91% cross-validation; AUC = 0.88, accuracy = 82% internal validation). For tumor grade prediction, the CT-based model achieved a higher AUC of 0.80 (accuracy 73%), whereas the PET-based model reached an AUC of 0.78 but demonstrated a slightly higher accuracy of 77%. Radiomic analysis of baseline [¹⁸F]FDG PET enables the discriminative prediction of immunotherapy response and tumor grade in advanced cSCC, with PET-based models outperforming CT-based ones. Full article

Objective: This study aims to apply different deep learning convolutional neural network algorithms to assess the grading of cervical spinal stenosis and to evaluate their consistency with clinician grading results as well as clinical manifestations of patients. Methods: We retrospectively enrolled 954 patients with cervical spine magnetic resonance imaging (MRI) data and medical records from the Fifth Affiliated Hospital of Sun-Yat Sen University. The Kang grading method for sagittal MR images of the cervical spine and the spinal cord compression ratio for horizontal MR images of the cervical spine were adopted for cervical spinal canal stenosis grading. The collected data were randomly divided into training/validation and test sets. The training/validation sets were processed by various image preprocessing and annotation methods, in which deep learning convolutional networks, including classification, target detection, and key point localization models, were applied. The predictive grading of the test set by the model was finally contrasted with the grading results of the clinicians, and correlation analysis was performed with the clinical manifestations of the patients. Result: The EfficientNet_B5 model achieved a five-fold cross-validated accuracy of 79.45% and near-perfect agreement with clinician grading on the test set (κ= 0.848, 0.822), surpassing resident–clinician consistency (κ = 0.732, 0.702). The model-derived compression ratio (0.45 ± 0.07) did not differ significantly from manual measurements (0.46 ± 0.07). Correlation analysis showed moderate associations between model outputs and clinical symptoms: EfficientNet_B5 grades (r = 0.526) were comparable to clinician assessments (r = 0.517, 0.503) and higher than those of residents (r = 0.457, 0.448). Conclusion: CNN models demonstrate strong performance in the objective, consistent, and efficient grading of cervical spinal stenosis severity, offering potential clinical value in automated diagnostic support. Full article

Additive Manufacturing (AM), as a revolutionary breakthrough in advanced manufacturing paradigms, leverages its unique layer-by-layer construction advantage to exhibit significant technological superiority in the fabrication of complex structural components for aerospace, biomedical, and other fields. However, when addressing industrial-grade precision manufacturing requirements, key challenges such as the multi-scale characteristics of surface damage precursors, interference from background noise, and the scarcity of high-quality training samples severely constrain the intelligent transformation of AM quality monitoring systems. This study proposes an innovative microscopic polarization YOLOv11-LSF intelligent inspection framework, which establishes an automated non-destructive testing methodology for AM device micro-nano damage precursors through triple technological innovations, effectively breaking through existing technical bottlenecks. Firstly, a multi-scale perception module is constructed based on the Large Separable Kernel Attention mechanism, significantly enhancing the network’s feature detection capability in complex industrial scenarios. Secondly, the cross-level local network VoV-GSCSP module is designed utilizing GSConv and a one-time aggregation method, resulting in a Slim-neck architecture that significantly reduces model complexity without compromising accuracy. Thirdly, an innovative simulation strategy incorporating physical features for damage precursors is proposed, constructing a virtual and real integrated training sample library and breaking away from traditional deep learning reliance on large-scale labeled data. Experimental results demonstrate that compared to the baseline model, the accuracy (P) of the YOLOv11-LSF model is increased by 1.6%, recall (R) by 1.6%, mAP50 by 1.5%, and mAP50-95 by 2.8%. The model hits an impressive detection accuracy of 99% for porosity-related micro-nano damage precursors and remains at 94% for cracks. Its unique small sample adaptation capability and robustness under complex conditions provide a reliable technical solution for industrial-grade AM quality monitoring. This research advances smart manufacturing quality innovation and enables cross-scale micro-nano damage inspection in advanced manufacturing. Full article

The successful introduction of new cultivars depends on the evaluation of complex parameters essential for the consumers, market, and fruit producers. A new scab-resistant apple cultivar, ‘Wuranda’ (SQ159/Natyra^®/Magic Star^® × Honeycrisp), recently introduced in Norway and managed under the name Fryd^©, is prone to biennial bearing. Therefore, one of the first tasks, investigated in Southwestern Norway by the Norwegian Institute of Bioeconomy Research, NIBIO-Ullensvang in 2021–2024, was the establishment of optimal crop load level based on the combination of productivity, fruit quality, and return bloom. The apple cultivar Fryd (‘Wuranda’) was propagated on ‘M.9’ rootstock and planted in 2019. The trial was performed in the same orchard for four consecutive years, starting three years after planting. Crop load level affected average fruit mass but had no impact on cv. Fryd fruit quality parameters at harvest such as blush, ground color, firmness, soluble solid content, or starch degradation. Fruit size variation was diminished by crop load regulation, and most fruits fell into 2–3 grading classes. Crop load, not the yield per tree, was the determining factor for the return bloom. The optimal crop load level depended on the orchard age. To guarantee a regular bearing mode of cv. Fryd planted on M.9 rootstock at a 3.5 × 1 m distance and trained as slender spindle, crop load of 5.5–6 fruits cm⁻² TCSA (trunk cross-sectional area) in the 3rd year, 7.5–8 fruits cm⁻² TCSA in the 4th year, and 6.5–7 fruits cm⁻² TCSA in the 5th year should be maintained. Full article

Dilution in underground mining refers to the unintended incorporation of waste material into the ore, reducing ore grade, revenue, and operational safety. Unplanned dilution, specifically, occurs due to overbreak caused by blasting inefficiencies or poor rock stability. To mitigate this issue, various factors related to rock quality, stope geometry, drilling, and blasting must be carefully considered. This study introduces a statistically rigorous methodology for the prediction of dilution in underground mining operations. The proposed framework incorporates a 10-fold cross-validation procedure with 30 repetitions, along with the application of nonparametric statistical tests. A total of eight supervised machine learning algorithms were investigated, with their hyperparameters systematically optimized using two distinct genetic algorithm (GA) strategies evaluated under varying population sizes. The models include support vector machines, neural networks, and tree-based approaches. Using a dataset of 120 samples, the results indicate that the GA-ANN model outperforms other approaches, achieving $MAE$, $R^2$, and $RMSE$ values of 0.2986, 0.8457, and 0.3928 for the training dataset, and 0.1882, 0.9508, and 0.2283 for the testing dataset, respectively. Furthermore, four stacking models were constructed by aggregating the top-performing base learners, giving rise to ensemble metamodels applied, for the first time, to the task of dilution prediction in underground mining. Full article

Background: Clear cell renal cell carcinoma (ccRCC) is the most common subtype of renal cancer, and its prognosis is closely linked to the International Society of Urological Pathology (ISUP) grade. While histopathological evaluation remains the gold standard for grading, non-invasive methods, such as radiomics, offer potential for automated classification. This study aims to develop a radiomics-based machine learning model for the ISUP grade classification of ccRCC using nephrographic-phase CT images, with an emphasis on model interpretability through SHAP (SHapley Additive exPlanations) values. Objective: To develop and interpret a radiomics-based machine learning model for classifying ISUP grade in clear cell renal cell carcinoma (ccRCC) using nephrographic-phase CT images. Materials and Methods: This retrospective study included 109 patients with histopathologically confirmed ccRCC. Radiomic features were extracted from the nephrographic-phase CT scans. Feature robustness was evaluated via intraclass correlation coefficient (ICC), followed by redundancy reduction using Pearson correlation and minimum Redundancy Maximum Relevance (mRMR). Logistic regression, support vector machine, and random forest classifiers were trained using 8-fold cross-validation. SHAP values were computed to assess feature contribution. Results: The logistic regression model achieved the highest classification performance, with an accuracy of 82% and an AUC of 0.86. SHAP analysis identified major axis length, busyness, and large area emphasis as the most influential features. These variables represented shape and texture information, critical for distinguishing between high and low ISUP grades. Conclusions: A radiomics-based logistic regression model using nephrographic-phase CT enables accurate, non-invasive classification of ccRCC according to ISUP grade. The use of SHAP values enhances model transparency, supporting clinical interpretability and potential adoption in precision oncology. Full article

The severity of physical disability at leprosy diagnosis reflects the timeliness of case detection and the effectiveness of disease surveillance. This study evaluates machine learning models to predict factors associated with late leprosy diagnosis—defined as grade 2 physical disability (G2D)—in Brazil from 2018 to 2022. Using an observational cross-sectional design, we analyzed data from the Notifiable Diseases Information System and trained four machine learning models: Random Forest, LightGBM, CatBoost, XGBoost, and an Ensemble model. Model performance was assessed through accuracy, area under the receiver operating characteristic curve (AUC-ROC), recall, precision, F1 score, specificity, and Matthew’s correlation coefficient (MCC). An increasing trend in G2D prevalence was observed, averaging 11.6% over the study period and rising to 13.1% in 2022. The Ensemble model and LightGBM demonstrated the highest predictive performance, particularly in the north and northeast regions (accuracy: 0.85, AUC-ROC: 0.93, recall: 0.90, F1 score: 0.83, MCC: 0.70), with similar results in other regions. Key predictors of G2D included the number of nerves affected, clinical form, education level, and case detection mode. These findings underscore the potential of machine learning to enhance early detection strategies and reduce the burden of disability in leprosy, particularly in regions with persistent health disparities. Full article

The accurate prediction of moisture content is crucial for controlling the drying process of agricultural products. While existing studies on drying models often rely on laboratory-scale experiments with limited data, real-time and high-frequency data collection under industrial conditions remains underexplored. This study collected and constructed a multi-dimensional dataset using an industrial-grade data acquisition system specifically designed for heat pump low-temperature circulating dryers. An artificial neural network (ANN) prediction model for moisture content during the rice drying process was developed. To prevent model overfitting, K-fold cross-validation was utilized. The model’s performance was evaluated using the mean squared error (MSE) and the coefficient of determination (R²), which also helped determine the preliminary structure of the ANN model. Bayesian regularization (trainbr) was then employed to train the network. Furthermore, optimization was conducted using neural network weights (RI) analysis and Sobol variance contribution analysis of the input variables to simplify the model structure and improve predictive performance. The experimental results showed that optimizing the model through RI sensitivity analysis simplified its topology to a 5-14-1 structure. The optimized model exhibited not only simplicity but also high prediction accuracy, achieving R² values of 0.969 and 0.966 for the training and testing sets, respectively, with MSEs of 5.6 × 10⁻³ and 6.3 × 10⁻³. Additionally, the residual errors followed a normal distribution, indicating that the predictions were reliable and realistic. Statistical tests such as t-tests, F-tests, and Kolmogorov–Smirnov tests revealed no significant differences between the predicted and actual values of rice moisture content, confirming the high consistency between them. Full article

Given the significant impact of air pollution on global health, the continuous and precise monitoring of air quality in all populated environments is crucial. Unfortunately, even in the most developed economies, current air quality monitoring networks are largely inadequate. The high cost of monitoring stations has been identified as a key barrier to widespread coverage, making cost-effective air quality monitoring devices a potential game changer. However, the accuracy of the measurements obtained from low-cost sensors is affected by many factors, including gas cross-sensitivity, environmental conditions, and production inconsistencies. Fortunately, machine learning models can capture complex interdependent relationships in sensor responses and thus can enhance their readings and sensor accuracy. After gathering measurements from cost-effective air pollution monitoring devices placed alongside a reference station, the data were used to train such models. Assessments of their performance showed that models tailored to individual sensor units greatly improved measurement accuracy, boosting their correlation with reference-grade instruments by up to 10%. Nonetheless, this research also revealed that inconsistencies in the performance of similar sensor units can prevent the creation of a unified correction model for a given sensor type. Full article

Previous studies on highway_rail grade crossing collisions have primarily focused on identifying factors contributing to the frequency and severity of driver injuries. In recent years, increasing attention has been given to modeling driver injury severity at these crossings. Recognizing the variations in injury severity under different road surface conditions, this study investigates the impact of road surface conditions on driver injury severity at highway_rail grade crossings. Using nearly a decade of accident data (2012–2021), thi study employs a LightGBM model to predict factors influencing injury severity and utilizes SHAP values for result interpretation. The symmetry principle of SHAP esures that factors with identical influence receive equal values, enhancing the reliability of predictive outcomes. The findings reveal that driver injury severity at highway_rail grade crossings varies significantly under different road surface conditions. Key factors identified include train speed, driver age, vehicle speed, annual average daily traffic (AADT), driver presence inside the vehicle, weather conditions, and location. The results indicate that collisions are more frequent when either the vehicle or train travels at high speed. Implementing speed limits for both vehicles and trains under varying road conditions could effectively reduce accident severity. Additionally, older drivers are more prone to severe accidents, highlighting the importance of installing control devices, such as warning signs or signals, to enhance driver alertness and mitigate injury risks. Furthermore, adverse weather conditions, such as rain, snow, and fog, exacerbate accident severity on road surfaces like sand, mud, dirt, oil, or gravel. Timely removal of surface obstacles may help reduce the severity of such accidents. Full article

Background: we evaluated regression models based on quantitative ultrasound (QUS) parameters and compared them with a vendor-provided method for calculating the ultrasound fat fraction (USFF) in metabolic dysfunction-associated steatotic liver disease (MASLD). Methods: We measured the attenuation coefficient (AC) and the backscatter-distribution coefficient (BSC-D) and determined the USFF during a liver ultrasound and calculated the magnetic resonance imaging proton-density fat fraction (MRI-PDFF) and steatosis grade (S0–S4) in a combined retrospective–prospective cohort. We trained multiple models using single or various QUS parameters as independent variables to forecast MRI-PDFF. Linear and nonlinear models were trained during five-time repeated three-fold cross-validation in a retrospectively collected dataset of 60 MASLD cases. We calculated the models’ Pearson correlation (r) and the intraclass correlation coefficient (ICC) in a prospectively collected test set of 57 MASLD cases. Results: The linear multivariable model (r = 0.602, ICC = 0.529) and USFF (r = 0.576, ICC = 0.54) were more reliable in S0- and S1-grade steatosis than the nonlinear multivariable model (r = 0.492, ICC = 0.461). In S2 and S3 grades, the nonlinear multivariable (r = 0.377, ICC = 0.32) and AC-only (r = 0.375, ICC = 0.313) models’ approximated correlation and agreement surpassed that of the multivariable linear model (r = 0.394, ICC = 0.265). We searched a QUS parameter grid to find the optimal thresholds (AC ≥ 0.84 dB/cm/MHz, BSC-D ≥ 105), above which switching from a linear (r = 0.752, ICC = 0.715) to a nonlinear multivariable (r = 0.719, ICC = 0.641) model could improve the overall fit (r = 0.775, ICC = 0.718). Conclusions: The USFF and linear multivariable models are robust in diagnosing low-grade steatosis. Switching to a nonlinear model could enhance the fit to MRI-PDFF in advanced steatosis. Full article