Frontiers in Artificial Intelligence最新文献

Introduction: Auditing methods have been significantly influenced by the combination of automation with artificial intelligence (AI) and have, in part, changed the roles of auditors and the quality of audits. Sampling-based traditional auditing has several challenges with identifying anomalies or new risks in financial information environments that are becoming increasingly complex and more data-rich.

Methods: In this study, AI-based techniques (machine learning and natural language processing) will be applied to a number of the steps involved in auditing financial information. The predictive model will be applied to lead propensity analysis and business volume forecasting, allowing the examination of both structured and unstructured data on financial statements and protecting the privacy of client data.

Results: A predictive model utilizing artificial intelligence (AI) was able to identify leads at an 87% rate of accuracy; forecasted business volume errors were less than 5%; and it explained nearly 94% of the variance between the AI model's predictions and the actual loan disbursement amounts. Using AI for full dataset analysis instead of sample-based methods improved auditors' ability to detect anomalies and allocate resources efficiently.

Discussion: Overall, the research demonstrates that AI provides auditors the capability to evaluate all data for a company, automate routine tasks, and identify specific areas (high-risk/high-value) that may require further review compared to other auditing methods. The new methodology also allows for early identification of potential risks and improves the overall efficiency of audits without compromising the protection of the companies' data.

Introduction: Breast cancer, one of the most life-threatening diseases that commonly affects women, can be effectively diagnosed using breast ultrasound imaging. A hybrid deep learning based ensemble framework combining the effectiveness of different convolutional neural network models has been proposed for breast ultrasound image classification.

Methods: Three distinct deep learning models, namely, EffcientNetB7, DenseNet121, and ConvNeXtTiny, have been independently trained on breast ultrasound image datasets in parallel to capture complementary representations. Local features are extracted using EffcientNetB7 through depthwise separable convolutions, whereas structural details are preserved by DenseNet121 utilizing dense connectivity. Global spatial relationships are modeled using ConvNeXtTiny via large kernel operations. Diverse local, global, and hierarchical features extracted with respect to multiple perspectives are integrated into a high-dimensional unified representation from which non-linear decision boundaries are learned utilizing XGBoost as the feature fusion classifier. Additionally, a soft voting ensemble method averages the predicted probabilities of the individual convolutional network architectures.

Results: The model was evaluated using the BUSI dataset, the BUS-UCLM dataset, and the UDIAT dataset. The accuracy, precision, recall, F1 score, and AUC values obtained on the BUSI data set are 88.46%, 88.49%, 88.46%, 88.45%, and 95.38%, respectively. On the BUS-UCLM dataset, the corresponding values are 90. 51%, 90. 56%, 90. 51%, 90. 51%, and 96. 23%, respectively. The accuracy, precision, recall, F1 score, and AUC values obtained on the UDIAT dataset are 96.97%, 100.00%, 90.91%, 95.24%, and 99.17%, respectively. The decision-making capability of the model has been highlighted using SHAP and Grad-CAM visualizations, further improving the interpretability and transparency of the model, and making it more robust for breast ultrasound image classification.

Discussion: The HED-Net framework exhibits significant potential for clinical application by enhancing diagnostic accuracy and decreasing interpretation time, particularly in resource-limited environments where expert radiologists are in short supply.

[This corrects the article DOI: 10.3389/frai.2025.1593017.].

Introduction: Rapid adoption of Artificial Intelligence (AI) in learning has revolutionized learners' engagement but comprehension of psychological and technological drivers of successful AI-enabled learning remains scarce. This research investigates how students' perceived agency of AI, usefulness, ease of use, trust, autonomy supporting, and self-efficacy collectively impact students' self-learning behavior and motivation. Based on Technology Acceptance Model (TAM), Social Cognitive Theory (SCT), and Self-Determination Theory (SDT) theories, our research model predicts an integrated model of motivational and behavioral processes underlying AI adoption in learning settings.

Methods: We adopted and followed a quantitative research design with a structured questionnaire administered among 280 higher education students in Saudi Arabia. We applied Structural Equation Modeling (SEM) using SmartPLS 4 to analyze data.

Results: Findings indicate that students' perceived agency of AI significantly predicts usefulness, ease of use, and autonomy supporting, while ease of use significantly enhances AI-enabled self-efficacy. Self-efficacy and autonomy supporting significantly impact self-learning motivation, driving self-learning behavior positively. But usefulness and trust in AI failed to influence self-efficacy directly, which reveals cultural and contextual settings.

Discussion: This research adds richness to the fusion of TAM, SCT, and SDT theories in illustrating how AI's perceived autonomy and usability collectively promote self-directed learning motivation. This research also provides guidelines to educators and system designers to design AI tools that promote learner autonomous settings, usability, and confidence. Future research ought to perform longitudinal and cross-cultural validations to fine-tune theoretically.

Introduction: Suicide is one of the leading causes of death among young people, to the extent that in many countries it is considered a public health issue. It is important to attempt to reduce the growth of this trend, especially among susceptible individuals, considering that it increased because of the COVID-19 pandemic. Natural language processing (NLP) provides various tools that allow for the analysis of texts to predict the presence of suicidal ideation. This work aims to conduct a systematic literature review to extract the computational techniques for identifying suicidal ideation in texts written in natural language.

Methods: The PRISMA 2020 method was used, which was divided into nine phases, and three inclusion criteria and two exclusion criteria were established for the selection of studies. The searches were conducted through high-level academic databases such as Scopus, IEEE Xplore, ACM Digital Library, Springer, and Web of Science. The risk of bias was assessed using AMSTAR 2. Potential biases identified include a lack of linguistic and cultural diversity and the predominance of data from social networks. A narrative synthesis was used to analyze and compare the findings qualitatively.

Results: In the end, 25 studies related to computational methods for detecting suicidal ideation in texts written in natural language were identified. The techniques mainly focus on transformer-based models such as BERT and hybrid methods, which combine this architecture with neural networks such as CNN and LSTM. There are also approaches with hierarchical attention mechanisms. Some studies employed additional techniques such as feature extraction with TF-IDF and pre-trained embeddings to improve model performance.

Discussion: Limitations in the evidence include the lack of linguistic and cultural diversity and the predominance of data from social networks. These results indicate that computational techniques have high potential to support early prevention strategies for suicidal ideation. However, expanding the diversity of linguistic contexts and improving understanding of the models among non-experts, such as physicians and other interested individuals, is necessary.

Increasingly many applications of machine vision and artificial intelligence (AI) can be observed in agriculture. Yet, high-quality training data remains a bottleneck in the development of many AI solutions, particularly for image segmentation. Therefore, ARAMSAM (agricultural rapid annotation module based on segment anything models) was developed, a user interface that orchestrates the pre-labelling capabilities of both the segment anything models (SAM 1, SAM 2) and conventional annotation tools. One in silico experiment on zero-shot performance of SAM 1 and SAM 2 on three unseen agricultural datasets and another experiment on hyperparameter optimization of the automatic mask generators (AMG) were conducted. In a user experiment, 14 agricultural experts applied ARAMSAM to quantify the reduction of annotation times. SAM 2 benefited greatly from hyperparameter optimization of its AMG. Based on ground-truth masks matched with predicted masks, the F₂ -score of SAM 2 improved from 0.05 to 0.74, while that of SAM 1 was improved from 0.87 to 0.93. The user interaction time could be reduced to 2.1 s/mask on single images (SAM 1) and to 1.6 s/mask on image sequences (SAM 2) compared to polygon drawing (9.7 s/mask). This study demonstrates the potential of segment anything models as incorporated into ARAMSAM to significantly accelerate the process of segmentation mask annotation in agriculture and other fields. ARAMSAM will be released as open-source software (AGPL-3.0 license) at https://github.com/DerOehmer/ARAMSAM.

The rapid development of robo-advisory and quantitative investment has been accompanied by persistent concerns about limited personalization and the opacity of black-box models operating on multimodal financial information. This paper addresses these issues from a decision-support perspective by constructing FinErva, a multimodal chain-of-thought dataset tailored to financial applications. FinErva comprises 7,544 manually verified question-answer pairs, divided into two economically relevant tasks: contract and disclosure understanding (FinErva-Pact) and candlestick-chart-based technical analysis (FinErva-Price). Building on this dataset, the paper propose a two-stage training framework: Supervised-CoT Learning followed by Self-CoT Refinement, and apply it to eight vision-language models, each with fewer than 0.8 billion parameters. Empirical results show that those lightweight models approach the performance of finance professionals and clearly outperform non-expert investors. Overall, the findings indicate that appropriately designed multimodal chain of thought supervision enables interpretable modeling of key research tasks such as contract review and chart interpretation under realistic computational and deployment constraints, providing new data and methodology for the development of personalized, explainable, and operationally feasible AI systems in investment advisory and risk management.

The aim of this systematic review is to examine and synthesize existing empirical evidence on external variables that influence students' attitudes toward the acceptance of artificial intelligence (AI) in improving English writing skills. This research offers a conceptual framework, AI Constructivist Learning Model (AICLM), based on Technology Acceptance Model (TAM) and Constructivist Learning Theory (CLT). Motivation, engagement, and societal expectations, based on CLT, are identified as external variables in TAM. These three constructs support active, autonomous, and student-centered learning. A systematic search of academic databases was conducted following PRISMA (Preferred Reporting Items for Systematic Reviews and Meta-Analyses) guidelines. Sixteen empirical studies published from 2021 to 2025, indexed in Scopus, Web of Science, and Google Scholar, were included in this review. Articles were selected on the basis of certain keywords such as, AI, English writing, TAM, and CLT. Findings indicate that students perceive the ease of use and usefulness of AI if they have high motivation, more engagement, and positive societal expectations. Therefore, motivation, engagement, and societal expectations are significant external variables that influence the attitudes of students toward AI acceptance in improving English writing. AI integration in English writing development can be successful if the interaction between the constructs of TAM and CLT is understood well. CLT supports why and how students engage actively with AI tools. Students are more likely to accept AI if it increases motivation enhances engagement and fulfils societal expectations. This conceptual framework is significant for future researchers and teachers in designing effective AI-based writing instructional strategies and curricula.

Artificial intelligence has become a crucial tool for effective customer management; therefore, this research aims to design and validate a scale measuring the adoption of artificial intelligence in the customer experience. It is approached from a quantitative methodology perspective and an instrumental design. A survey was conducted among 528 customers who frequently make virtual purchases. Then, an exploratory analysis was conducted to determine the factor structure of the scale, followed by a confirmatory analysis to validate the construct. On the other hand, an invariance analysis was conducted to determine whether the construct varies across groups. The results show a multidimensional scale of 16 items grouped into 4 factors (trust in AI, perception of AI, knowledge of AI, shopping experience). Each factor consists of four items, using a Likert-type response scale where 1 indicates "totally disagree" and 5 indicates "totally agree". In conclusion, the proposed scale is a valid measure. It can be used to continue exploring this concept in other latitudes, serving as a valuable tool for entrepreneurs to make an effective diagnosis of this new technology.

The integration of artificial intelligence (AI) into chemical risk assessment (CRA) is emerging as a powerful approach to enhance the interpretation of complex toxicological data and accelerate safety evaluations. However, the regulatory uptake of AI remains limited due to concerns about transparency, explainability, and trustworthiness. The European Partnership for the Assessment of Risks from Chemicals (PARC) project ReadyAI was established to address these challenges by developing a readiness scoring system to evaluate the maturity and regulatory applicability of AI-based models in CRA. The project unites a multidisciplinary consortium of academic, regulatory, and legal experts to define transparent and reproducible criteria encompassing data curation, model development, validation, explainability, and uncertainty quantification. Current efforts focus on identifying key priorities, including harmonized terminology, rigorous data quality standards, case studies, and targeted training of regulatory scientists. ReadyAI aims to deliver a practical, evidence-based scoring system that enables regulators to assess whether AI tools are sufficiently reliable for decision-making and guides developers toward compliance with regulatory expectations. By bridging the gap between AI innovation and regulatory applicability, ReadyAI contributes to the responsible integration of AI into chemical safety assessment frameworks, ultimately supporting human and environmental health protection.