Frontiers in Artificial Intelligence最新文献

Introduction: Digital lending, also known as alternative lending, refers to fintech platforms that offer quick and easy loans through digital channels, bypassing many of the limitations of traditional banking. Since the mid-2000s, digital lending has become a major fintech innovation, with rapid growth in India driven by financial inclusion measures. However, the sector continues to face challenges, including fraud, transparency issues, and consumer dissatisfaction. The primary objective of this study was to understand how consumers perceive and assess India's RBI-approved P2P digital lending apps by analyzing a large dataset of customer feedback to identify strengths, weaknesses, and overall satisfaction levels.

Methods: The study analyzed a final dataset of 15,408 user reviews collected from seven RBI-approved digital lending platforms: 5Paisa, Faircent, i2iFunding, LenDenClub, CashKumar, Lendbox, and IndiaMoneyMart derived from an initial 15,537 reviews. The cleaned data was then examined using natural language processing, topic modeling, and supervised machine learning and deep learning models to identify key themes and evaluate predictive performance.

Results: Topic modeling identified 11 recurring topics. Sentiment analysis revealed that 55% of evaluations were positive, 41% were negative, and 4% were neutral. Strengths included loan disbursement, withdrawals, and EMI payments, while weaknesses involved interface design, transparency around rejections, and login functionality. Comparative data revealed that IndiaMoneyMart and i2iFunding received the highest user satisfaction, while 5Paisa and Lendbox trailed due to recurring complaints about transparency, accessibility, and overall user experience. In terms of modeling, the deep learning model VGG16 and ensemble machine learning techniques (XGBoost, CatBoost, and LightGBM) consistently achieved the highest predictive accuracy (up to 0.88), outperforming simpler models such as decision trees and ResNets.

Discussion: The findings indicate that digital lending platforms support financial inclusion but require improvements in user interface and user experience, better transparency in loan decisions, and stronger customer support. Addressing these areas can help strengthen trust and promote long term adoption of digital lending services.

The rise of agentic artificial intelligence (Agentic AI) marks a transition from systems that optimize externally specified objectives to systems capable of representing, evaluating, and revising their own goals. Whereas earlier AI architectures executed fixed task specifications, agentic systems maintain recursive loops of perception, evaluation, goal-updating, and action, allowing them to sustain and adapt purposive activity across temporal and organizational scales. This paper argues that Agentic AI is not an incremental extension of large language models (LLMs) or autonomous agents in the sense we know it from classical AI and multi-agent systems, but a reconstitution of agency itself within computational substrates. Building on the logic of coordination, delegation, and self-regulation developed in early agent-based process management systems, we propose a general theory of synthetic purposiveness, where agency emerges as a distributed and self-maintaining property of artificial systems operating in open-ended environments. We develop the concept of synthetic teleology-the engineered capacity of artificial systems to generate and regulate goals through ongoing self-evaluation-and we formalize its dynamics through a recursive goal-maintenance equation. We further outline design patterns, computational semantics, and measurable indicators of purposiveness (e.g., teleological coherence, adaptive recovery, and reflective efficiency), providing a foundation for the systematic design and empirical investigation of agentic behaviour. By reclaiming agency as a first-class construct in artificial intelligence, we argue for a paradigm shift from algorithmic optimization toward goal-directed reasoning and purposive orchestration-one with far-reaching epistemic, societal, and institutional consequences.

Introduction: The increasing cyber threats targeting industrial control systems (ICS) and the Internet of Things (IoT) pose significant risks, especially in critical infrastructures like the oil and gas sector. Existing machine learning (ML) approaches for cyberattack detection often rely on binary classification and lack computational efficiency.

Methods: This study proposes two optimized stacked ensemble models to enhance attack detection accuracy while reducing computational overhead. The main contribution lies in the strategic selection and integration of diverse base models, such as Logistic Regression, Extra Tree Classifier, XGBoost, and LGBM, with RFC as the final estimator. These models are chosen to address unique characteristics of security datasets, such as class imbalance, noise, and complex attack patterns. This combination aims to leverage different decision boundaries and learning mechanisms.

Results: Evaluations show that the Stacked Ensemble_2 model achieves 97% accuracy with a training and testing computation time of 54 minutes. Stacked Ensemble_2, which excelled over the traditional Stacked Ensemble_1, was also evaluated on the CICIDS 2017 dataset, achieving an impressive 100% accuracy with an AUROC of 99%.

Discussion: The results indicate that the proposed Stacked Ensemble_2 model provides a scalable, real-time detection mechanism for securing ICS and IoT environments. By proving its effectiveness on unseen data, this model demonstrates a significant advancement over traditional methods, offering enhanced accuracy and efficiency in detecting sophisticated cyber threats in critical infrastructure sectors.

Despite recent advancements, modern kitchens, at best, have one or more isolated (non-communicating) "smart" devices. The vision of having a fully-fledged ambient kitchen where devices know what to do and when has yet to be realized. To address this, we present RiCoRecA, a novel schema for parsing cooking recipes into a workflow representation suitable for automation, a step toward that direction. Methodologically, the schema requires a number of information extraction tasks, i.e., annotating named entities, identifying relations between them, coreference resolution, and entity tracking. RiCoRecA differs from previously reported approaches in that it learns these different information extraction tasks using one joint model. We also provide a dataset containing annotations that follow this schema. Furthermore, we compared two transformer-based models for parsing recipes into workflows, namely, PEGASUS-X and LongT5. Our results demonstrate that PEGASUS-X surpassed LongT5 on all of the annotation tasks. Specifically, PEGASUS-X surpassed LongT5 by 39% in terms of F-Score when averaging the performance on all the tasks; it demonstrated almost human-like performance.

Bone fractures are among the most prominent injuries in the modern world that affect all ages and races. Traditional treatment involves radiographic imaging that relies heavily on radiologists manually analyzing images. There have been efforts to develop computer-aided diagnosis tools that employ artificial intelligence and deep learning approaches. Existing literature focuses on developing tools that only detect and classify bone fractures, rather than addressing the broader issue of bone fracture management. However, evidence of scholarly works that include treatment recommendations is still lacking. Furthermore, deep learning-based object detectors that achieve state-of-the-art results are computationally expensive and considered as black-box solutions. Developing countries, such as Sub-Saharan Africa, face a shortage of radiologists and orthopedists. For this reason, this paper proposes a methodological approach that uses a more efficient object detection model to diagnose long bone fractures and provide prescription recommendations. An enhanced anchoring process, known as adaptive anchoring, is proposed to improve the performance of the Regional Proposal Network and the object detection model. A Faster R-CNN model with ResNet-50/101 and ResNext-50/101 backbones was used to develop an object detection model that uses X-ray images as input. To understand and interpret the model's decision, a Gradient-based Class Activation Mapping method was used to assess the model's learnability. The results indicate that the proposed adaptive anchoring approach can improve computational efficiency, reducing training time by up to 29% compared to the traditional approach. Model accuracy during training and validation ranged between 94% and 98%. Overall, adaptive anchoring performed better when applied with the ResNet-101 backbone, yielding an Average Precision of 92.73%, an F1 score of 96.01%, a precision of 96.80%, and a recall of 95.23%. The study provides valuable insights into the use of computationally efficient deep learning models for medical recommendation systems. Future studies should develop models to diagnose fractures using input images from various modalities and to provide prescription recommendations.

Introduction: Detecting epileptic seizures remains a major challenge in clinical neurology due to the complex, heterogeneous, and non-stationary characteristics of electroencephalogram (EEG) signals. Although recent machine learning (ML) and deep learning (DL) approaches have improved detection performance, most methods still struggle with limited interpretability, inadequate spatial-temporal modeling, and suboptimal generalization. To address these limitations, this study proposes an enhanced hybrid parallel convolutional-GhostNet framework (HPG-ESD) for robust seizure detection using multimodal EEG and functional Magnetic Resonance Imaging (fMRI) data.

Methods: The experimental data consist of pediatric scalp EEG recordings from 24 subjects in the CHB-MIT dataset (22-channel 10-20 system, 256 Hz sampling, continuous multi-hour recordings) and resting-state 3T fMRI scans from 52 participants in the UNAM TLE dataset (26 epilepsy patients and 26 healthy controls). EEG data underwent Gauss-based median filtering, while fMRI images were denoised using an adaptive weight-based Wiener filter. Spatial, temporal, and spectral EEG features were extracted alongside an enhanced common spatial pattern (E-CSP) representation, whereas fMRI features were obtained using deep 3D CNN embeddings combined with a smoothened pyramid histogram of oriented gradients (S-PHOG) descriptor. These multimodal features were fused within a soft voting hybrid parallel convolutional-GhostNet (S-HPCGN) model, integrating an improved attention based parallel convolutional network (IAPCNet) and GhostNet to capture complementary spatial-temporal patterns.

Results: The proposed HPG-ESD framework achieved an accuracy of 0.941, precision of 0.939, and sensitivity of 0.944, outperforming conventional unimodal and state-of-the-art methods.

Discussion: These results demonstrate the potential of multi-modal learning and lightweight attention-enhanced architectures for reliable and clinically relevant seizure detection.

Background: Healthcare professionals' awareness and handling of artificial intelligence applications in healthcare enhance patient outcomes and improve processes. This study aimed to evaluate the perception, attitude, knowledge, and practice of healthcare professionals regarding the application of artificial intelligence in Egyptian healthcare settings.

Method: A cross-sectional study in which 367 healthcare professionals responded to an electronic questionnaire.

Results: Out of 367 participants (234 female), radiology and lab test specialty (36.2%) was the predominant. The mean age was 27.03 years; 51.8% of respondents showed positive perception, 68.7% experienced sub-optimal knowledge, 52.9% expressed negative attitudes, and 53.4% demonstrated a low practice level of AI tools. Younger age was significantly associated with positive perception (adjusted odds ratio (AOR) = 0.905, p = 0.020) and higher AI practice (AOR = 0.907, p = 0.026). University hospital professionals had 61.4% lower odds of optimal knowledge than private hospital professionals (AOR = 0.386, p = 0.046). Men had higher odds of both positive attitudes (AOR = 1.844, p = 0.010) and high practice level (AOR = 2.92, p < 0.001). Pre-bachelor's holders had lower odds of positive attitudes (AOR = 0.361, p = 0.036), as well as physicians compared to nurses and others (AOR = 0.424, p = 0.005). Bachelor's holders showed lower odds of high AI practice (AOR = 0.388, p = 0.017).

Conclusion: Despite moderate perception, most professionals have knowledge, attitude, and practice defects. Mainly, younger age and men showed higher engagement, indicating a need for targeted AI training, especially for older and female professionals.

Introduction: This study addresses the challenge of improving wavefront correction for Orbital Angular Momentum (OAM) in oceanic turbulence using a physics-constrained Generative Adversarial Network (GAN).

Methods: We integrated physical constraints into a deep learning framework to reconstruct degraded input images (SSIM = 0.62). The model was trained with varied loss settings, including a baseline model, spectral constraints (+Spec), and spatial constraints (+Ortho).

Results: The dual-constraint approach (+Ortho+Spec) reached a near-optimal SSIM of 0.98. Ablation studies revealed that while +Ortho boosted modal purity to 95.7%, the dual-constraints achieved 98.4% purity. Power spectral density analysis via KL divergence confirmed the dual-constraints' superiority (KL = 0.56) over the baseline (KL = 2.47).

Discussion: These results demonstrate that integrating both spatial and spectral constraints effectively optimizes reconstruction, purity, and spectral fidelity, offering a robust solution for OAM correction in underwater optical communication systems.

Graph Neural Networks (GNNs) have transformed multimodal healthcare data integration by capturing complex, non-Euclidean relationships across diverse sources such as electronic health records, medical imaging, genomic profiles, and clinical notes. This review synthesizes GNN applications in healthcare, highlighting their impact on clinical decision-making through multimodal integration, advanced fusion strategies, and attention mechanisms. Key applications include drug interaction and discovery, cancer detection and prognosis, clinical status prediction, infectious disease modeling, genomics, and the diagnosis of mental health and neurological disorders. Various GNN architectures demonstrate consistent applications in modeling both intra- and intermodal relationships. GNN architectures, such as Graph Convolutional Networks and Graph Attention Networks, are integrated with Convolutional Neural Networks (CNNs), transformer-based models, temporal encoders, and optimization algorithms to facilitate robust multimodal integration. Early, intermediate, late, and hybrid fusion strategies, enhanced by attention mechanisms like multi-head attention, enable dynamic prioritization of critical relationships, improving accuracy and interpretability. However, challenges remain, including data heterogeneity, computational demands, and the need for greater interpretability. Addressing these challenges presents opportunities to advance GNN adoption in medicine through scalable, transparent GNN models.