Frontiers in Artificial Intelligence最新文献

Background: Concentration difficulty is recognized as a hallmark of various neurologic and neuropsychiatric disorders. However, an accurate estimation of epidemiological risk factors for concentration difficulty remains severely limited.

Aims: The study aimed to develop an interpretable machine-learning (ML) model to predict risk factors of concentration difficulty among adults in the United States.

Methods: A total of 9,971 participants were included from the 2015-2016 cycle of the National Health and Nutrition Examination Survey (NHANES). Six ML algorithms, including Logistic Regression, ExtraTrees classifier, Bagging, Gradient Boosting, Extreme Gradient Boosting (XGBoost), and Random Forest (RF), were applied in this study. Model performance was evaluated using the area under the receiver operating characteristic curve (AUC), accuracy, precision, specificity, decision curve analysis (DCA), and calibration plots. Finally, a nomogram was constructed based on the best performing model.

Results: Of these, 2,146 participants aged 20 years and older were analyzed. Logistic regression exhibited the best clinical predictive value in both internal and external validation sets, with AUCs of 0.881 and 0.818, respectively. The DCA curve revealed that logistic regression exhibited the greatest net benefits in the internal cohort, whereas the RF model provided the largest net benefits in the external cohort (threshold: 0.2-0.3).

Conclusion: Logistic regression exhibited the highest clinical value in predicting concentration difficulty. These findings provide valuable insights for the recognition, management, and effective interference strategies for concentration difficulty.

Introduction: Rectal cancer is a globally prevalent cancer, and accurate segmentation of rectal lesions in abdominal CT images is critical for clinical diagnosis and treatment planning. Existing methods struggle with imprecise boundary delineation due to low tissue contrast, image noise, and varied lesion sizes, prompting the development of a specialized segmentation framework.

Methods: We developed the Hierarchical Hypercolumn-guided Fusion Segment Anything Model (HHF-SAM) with three core components: 1) A Med-Adapter SAM Encoder integrating LoRA and Adapter modules to adapt SAM's natural image understanding capability to medical-specific features; 2) A Multi-scale Hypercolumn Processing Module to capture comprehensive features for lesions of varying sizes and shapes; 3) A Progressive Hierarchical Fusion Decoder with Hierarchical Fusion Module to aggregate multi-scale features and resolve boundary blurring. The model was evaluated on two public abdominal CT datasets (CARE and WORD) using mean Dice coefficient (mDice) and mean Intersection over Union (mIoU) as metrics.

Results: On the CARE dataset, HHF-SAM achieved a mean mDice of 74.05% and mean mIoU of 58.96%, outperforming state-of-the-art methods (U-SAM: 69.28% mDice, 53.11% mIoU; SAM: 65.98% mDice, 49.44% mIoU). For tumor segmentation specifically, it reached 76.42% mDice and 62.03% mIoU. On the WORD dataset, it achieved an average mDice of 85.84% across all organs, with 83.24% mDice for rectal segmentation (surpassing U-SAM's 80.66% and SAM's 72.77%).

Discussion: This study presents an SAM-based framework optimized for the unique characteristics of abdominal CT images, effectively overcoming the limitations of general segmentation models in medical image processing. The proposed HHF-SAM provides a reliable tool for clinical auxiliary diagnosis, reducing inter-reader variability and improving efficiency in lesion delineation.

DeepSeek is a free and self-hostable large language model (LLM) that recently became the most downloaded app across 156 countries. As early academic literature on ChatGPT was predominantly critical of the model, this mini-review is interested in examining how DeepSeek is being evaluated across academic disciplines. The review analyzes available articles with DeepSeek in the title, abstract, or keywords, using the VADER sentiment analysis library. Due to limitations in comparing sentiment across languages, we excluded Chinese literature in our selection. We found that Computer Science, Engineering, and Medicine are the most prominent fields studying DeepSeek, showing an overall positive sentiment. Notably, Computer Science had the highest mean sentiment and the most positive articles. Other fields of interest included Mathematics, Business, and Environmental Science. While there is substantial academic interest in DeepSeek's practicality and performance, discussions on its political or ethical implications are limited in academic literature. In contrast to ChatGPT, where all early literature carried a negative sentiment, DeepSeek literature is mainly positive. This study enhances our understanding of DeepSeek's reception in the scientific community and suggests that further research could explore regional perspectives.

Introduction: Decision-making supported by intelligent systems is being increasingly deployed in ethically sensitive domains. As a result, it is of considerable importance to understand the patterns of moral judgments generated by large language models (LLMs).

Methods: To this end, the current research systematically investigates how two prominent LLMs (i.e., ChatGPT and Claude Sonnet) respond to 12 moral scenarios previously administered to human participants (first language and second language users). The primary purpose was to examine whether the responses generated by LLMs align with either deontological or utilitarian orientations. Our secondary aim was to compare response patterns of these two models to those of human respondents in previous studies.

Results: Contrary to prevailing assumptions regarding the utilitarian tendency of LLMs, the findings revealed subtle response distributions of moral choice that are context-sensitive. Specifically, both models alternated between deontological and utilitarian judgments, depending on the scenario-specific features.

Discussion: These output patterns reflect complex moral trade-offs and may play a significant role in shaping societal trust and acceptance of AI systems in morally sensitive domains.

The rapid integration of Artificial Intelligence (AI) in India's banking sector offers operational benefits but also raises sustainability challenges. This study focuses on "Green AI," defined as AI technologies optimized for energy efficiency and carbon conscious practices, by extending the Technology-Organization-Environment (TOE) and Technology Acceptance Model (TAM) frameworks with sustainability-linked factors. Data were collected from 412 mid- to senior-level professionals across six leading public and private banks, and Structural Equation Modeling (SEM) was employed to test the proposed hypotheses. Findings reveal that Banking Infrastructure (β = 0.419), Financial Investment (β = 0.401), and Competitive Pressure (β = 0.329) are the strongest predictors of Green AI adoption, while Regulatory Influence (β = 0.147), Perceived Usefulness (β = 0.129), and Perceived Ease of Use (β = 0.098) exert weaker but significant effects. Adoption of Green AI demonstrates a positive link to sustainability outcomes (β = 0.446), indicating its potential to convert structural readiness into measurable environmental gains. Although direct energy-consumption data were unavailable, perceptual measures provided valid proxies aligned with emerging-market studies. The results suggest that resource and market drivers outweigh attitudinal factors, offering actionable insights for infrastructure investment, regulatory refinement, and ESG integration, with implications for other emerging economies.

Background: Accurate antenatal prediction of SGA at birth is essential to improve development and delivery of preventative and therapeutic interventions. This study aimed to assess the performance of machine learning (ML) models to predict SGA at birth among Chinese pregnancies classified according to the Chinese birthweight standard and three international birthweight standards.

Methods: We collected multimodal, longitudinal, antenatal surveillance data on 350,135 singleton pregnancies in Wenzhou City, China, between Jan 1, 2014 and Dec 31, 2016. For three pregnancy intervals we developed ML prediction models for newborns classified as SGA using the China, Intergrowth 21st, Fetal Medicine Foundation (FMF), and Gestation-related Optimal Weight (GROW) standards. We applied lasso regression to conduct feature selection, and CatBoost, XGBoost, LightBoost, Artificial Neural Networks, Random Forest, Stacked ensemble model, and logistic regression for predictive modeling in training data sets, with validation in testing data sets.

Results: Among 22,603 singleton pregnancies with complete data, the rate of SGA using the China standard was 6.1%, compared to 4.3, 6.0, and 9.7% for the Intergrowth 21st, GROW, and FMF standards, respectively. This pattern was maintained in the imputed data set (n = 225,523), with corresponding SGA rates of 6.8, 4.8, 7.4, and 10.7%. Late pregnancy models (<37 weeks) had the best power to predict SGA, compared to middle (<26 weeks) and early pregnancy (<18 weeks) models. With the China standard, the logistic regression model in late pregnancy performed best with an area under the receiver operating characteristic curve (ROC-AUC) of 0.74. Logistic regression also performed better than ML algorithms with the Intergrowth-21st and GROW standards at each pregnancy interval, although differences were small. The Random Forest model with the FMF standard achieved superior performance at each pregnancy interval, reaching a ROC-AUC of 0.79 in late pregnancy. Notably, the middle pregnancy Random Forest model with the FMF standard already attained a ROC-AUC of 0.72 at 26 weeks' gestation. Symphysis-fundal height, maternal abdominal circumference, maternal age, maternal height and weight, and parity were consistently identified as key predictors of SGA across the different standards.

Conclusion: There are important differences in the classification of SGA at birth between national and international birthweight standards. Both machine learning models and traditional logistic regression demonstrated comparable predictive performance for SGA identification. These findings hold promise for guiding risk-stratified prenatal care and optimizing resource allocation in clinical settings.

Introduction: Potato foliar diseases, particularly early and late blight, pose a serious threat to yield and food security, yet reliable visual recognition remains challenging due to cultivar heterogeneity, variable symptom expression, and acquisition noise in field-like imagery. To address these issues, we propose PotatoLeafNet, a two-stage deep learning framework that combines a fixed-sequence image-augmentation pipeline with a compact, task-optimized 11-layer convolutional neural network (CNN) using 3 × 3 kernels for robust, data-efficient classification of potato leaf conditions (Healthy, Early Blight, Late Blight).

Methods: We construct a dataset of 4,072 labeled potato leaf images from the PlantVillage-Potato subset and standardize all inputs to 224 × 224 RGB tensors with pixel intensities normalized to [0,1]. A balanced, fixed-order augmentation policy-comprising rotation, translation, shear, zoom, horizontal flipping, brightness adjustment, and channel jitter-is applied exclusively to the training split, increasing it to 6,000 images (2,000 per class) while keeping the validation and test sets free of synthetic samples. The second stage consists of an 11-layer CNN implemented in TensorFlow/Keras and trained with categorical cross-entropy loss and the Adam optimizer under a unified training and evaluation protocol. Performance is benchmarked against strong CNN and hybrid baselines, including ResNet-50 + VGG-16, VGG-16 + MobileNetV2, MobileNetV2, and Inception-V3.

Results: On the PlantVillage-Potato test set, PotatoLeafNet achieves 98.52% accuracy, 98.67% macro-precision, 99.67% macro-recall, 99.16% macro-F1, and 1.00 macro-AUC, outperforming all baseline models under identical preprocessing and training conditions. In particular, PotatoLeafNet surpasses ResNet-50 + VGG-16 (97.10% accuracy, AUC 0.98), VGG-16 + MobileNetV2 (94.80% accuracy, AUC 0.93), MobileNetV2 (93.20% accuracy, AUC 0.92), and Inception-V3 (92.50% accuracy, AUC 0.91). Short 10-epoch runs yield stable convergence (training accuracy 88.22%, validation accuracy 86.91%, test accuracy 88.15%), indicating efficient learning from the augmented distribution.

Discussion: The results demonstrate that explicitly coupling a fixed sequential augmentation stage with a lightweight 3×3-kernel CNN enables high tri-class accuracy, strong recall for disease classes, and improved generalization relative to deeper or fused architectures, without incurring substantial computational cost. By emphasizing disease-relevant structure while limiting overfitting, PotatoLeafNet provides a practical and resource-efficient solution for automated screening of potato leaf health in real-world agronomic settings, supporting timely and data-driven disease management.

Superiority of images in low light is necessary in the case of medical image as well as autonomous systems but there is still a challenge of balancing between brightness and natural appearance. The presented paper elaborates a new improvement model that combines Interval-Valued Intuitionistic Fuzzy Set as well as Reptile Search Algorithm optimization. The proposed approach automatically tunes the fuzzy membership and hesitation factors to adapt to uncertainty in dark areas while preserving significant structural data. The Performance is evaluated using common objective metrics which are Peak Signal-to-Noise Ratio, Absolute Mean Brightness Error, Contrast Improvement Index and entropy. All the reported percentage improvements are computed using the average metric values of the baseline Interval-Valued Intuitionistic Fuzzy Set method on the complete dataset. The results of the investigations indicate significant and consistent increases in the experimental results with a 3.69% percentage gain in entropy, a 21.71% percentage gain in brightness restoration, an 18.73% percentage gain in contrast and a 66.12% percentage gain in Peak Signal to Noise Ratio compared to the baseline method. As these results show, the given technique yields naturally amplified images that have better qualities in clarity, conciseness and structural conservation, which is extremely applicable in real-life situations involving low-light photography.

In the context of smart manufacturing, improving the quality and efficiency of process planning, especially in the processing of complex parts, has become a key factor influencing the level of intelligence in manufacturing systems. However, most current process planning methods still heavily rely on manual expertise, leading to problems such as difficulty in knowledge reuse, low planning efficiency, and slow response times, which are inadequate to meet the diverse and changing needs of engineering applications. To address these issues, this paper proposes an algorithm for Assembly Process Reasoning and Decision-making based on Bidirectional Long Short-Term Memory with Attention (AP-BiLSTM-ATT), which aims to deeply explore the hidden relationships between the multi-dimensional features of parts and process plans, thereby achieving probabilistic modeling of process decisions. Specifically, the attributes, geometric features, and historical process plans of parts are first labeled and vectorized, transforming traditional process knowledge into structured data representations suitable for deep learning models. A BiLSTM network model, integrated with a multi-head attention mechanism, is then constructed to capture contextual dependencies and semantic weight distributions between features, enhancing the model's ability to express complex process relationships. During training, the model learns the mapping distribution between features and processes from a large-scale historical process dataset, enabling intelligent reasoning and recommendation of process plans for new parts. The results show that this method outperforms traditional methods in terms of accuracy, response speed, and generalization ability in process planning, providing effective support for enhancing the intelligence of complex part process planning and laying a foundation for the structured expression and intelligent application of manufacturing process knowledge.

Current AI benchmarks often equate mimicry with genuine intelligence, emphasizing task performance over the underlying cognitive processes that enable human-like understanding. The Machine Perturbational Complexity & Agency Battery (mPCAB) introduces a new, substrate-independent framework that applies neurophysiological methods used initially to assess consciousness in artificial systems. Unlike existing evaluations, it features four key components-perturbational complexity, global workspace assessment, norm internalization, and agency-that link mechanisms with functions. This enables systematic comparisons across digital, neuromorphic, and biological substrates, addressing three research gaps: long-term reasoning with coherent behavior, norm internalization amid distribution shifts, and transformational creativity involving meta-cognitive rule modification. By analyzing theories of consciousness (GNW, IIT, PP, HOT), we identify targets for AI implementation. Our cognitive architecture analysis maps human functions-such as working memory and executive control-to their computational counterparts, providing guiding principles for design. The creativity taxonomy progresses from combinational to transformational, with measurable criteria like changes in conceptual space and the depth of meta-level reasoning. Ethical considerations are integrated into frameworks for monitoring organoid intelligence, reducing bias in creativity, and addressing rights issues. Pilot studies demonstrate mPCAB's feasibility across different substrates and show that its metrics are comparable. This framework moves evaluation away from superficial benchmarks toward mechanism-based assessment, supporting the development of mind-like machines and responsible AI advancements.