Applied AI letters最新文献

This review evaluates classical machine learning-based hypertension prediction models, emphasizing their role in addressing global health burdens, particularly in low- and middle-income countries. Hypertension affects over 1.28 billion people globally and contributes to cardiovascular disease and mortality. The review compares machine-learning techniques with traditional methods, focusing on key datasets, evaluation metrics, and model development to advance early detection and effective hypertension management. The review used the PRISMA framework, using databases such as Google Scholar, PubMed, and IEEE explorer to identify studies published between 2020 and 2024 on machine learning techniques, predictive models, and early detection of hypertension based on relevance, methodological rigor, and inclusion criteria. The study analyzed hypertension prediction models across various countries, including the US, England, Korea, Japan, China, Indonesia, Thailand, India, Bangladesh, Nepal, and several African countries. The models' performance varied with AUC statistic values ranging from 0.6 to 0.9, indicating a wide range of predictive accuracy. Machine learning techniques generally reported higher performance metrics than traditional statistical methods. Risk factor heterogeneity was evident, with models like random forest, logistic regression, and gradient-boosted trees showing high predictive accuracy. Emerging techniques like SMOTE (Synthetic Minority Oversampling Technique) and ensemble methods improved unbalanced data set performance. The review explores the potential of machine learning-based hypertension prediction models in healthcare, highlighting their ability to accurately predict hypertension risk, tailor interventions to specific populations, and optimize healthcare resources in low- and middle-income countries. However, challenges include data quality, model explainability, and ethical considerations. Despite these, ML integration offers scalable and cost-effective solutions, especially in resource-limited settings. Future research should focus on diverse datasets, advanced feature integration, and longitudinal validations.

The sterile insect technique (SIT) represents a highly effective and promising method for combating tsetse fly-related infections, which involves the release of sterilized male tsetse flies in the assigned zones. However, tsetse fly rearing poses specific challenges, particularly in the tsetse sex separation, as this process is labor-intensive and incurs significant costs. Here, we report a simple model that classifies tsetse flies by sex using an object detection model based on the YOLO algorithm. This paper also conducted a comparative analysis of YOLOv8 and YOLO11 deep learning models, focusing on their efficacy in tsetse fly detection and classification using a range of performance metrics and statistical analysis. The findings reveal that the classification accuracy of YOLO11 stands at 97.6%, whereas YOLOv8 achieves 95.6%. The classification precision of YOLO11 in identifying tsetse flies is 88.6%, while that of YOLOv8 is 85.9%. Additionally, YOLO11 demonstrates an inference speed of 13.0 ms, slightly faster than YOLOv8's 13.4 ms in tsetse sex detection. Moreover, YOLO11 outperformed YOLOv8 in both F1 score and [email protected]–0.9, a success attributed to its enhanced architectural design. However, statistical tests indicate there is no significant difference between the two models, achieving p values ≥ 0.05 for all metrics. This study adds value to tsetse rearing and fly-based disease control by offering automated tsetse sex detection insights into its practical uses in real-world contexts. Furthermore, this research enriches the understanding of the two models with tsetse flies as the focal point and recommends a more effective and accurate detection approach. Finally, integrating the model with the mobile object detection Android app will reduce tsetse sex sorting dependency on experienced technical experts and enhance tsetse rearing productivity.

This study proposes a hybrid methodology for evaluating leading retail companies based on customer perspectives, combining Artificial Intelligence (AI)-driven sentiment analysis with fuzzy multiple criteria decision-making (MCDM). The framework integrates large-scale customer review analysis with expert decision-making to provide a comprehensive assessment of retail performance. The process begins with AI-based text mining to collect and analyze customer reviews, extracting emotional tones and identifying frequently mentioned criteria. Expert judgment is then applied to refine, organize, and assign importance to these criteria. The q-rung orthopair fuzzy set MCDM methodology is employed to address uncertainty, conflicting objectives, and qualitative expert opinions by translating them into a structured quantitative evaluation. This hybrid approach offers a balanced assessment that combines subjective and objective dimensions. As a case study, 2000 customer reviews from each of four major U.S. retailers—Amazon, Walmart, Costco, and Target—were analyzed to derive key evaluation criteria based on user feedback. The proposed method distinguishes itself through its unique integration of sentiment analysis and decision-makers' expert evaluations, enabling a holistic and robust evaluation of alternatives. By bridging customer perceptions with expert analysis, this methodology provides a deeper, more nuanced understanding of retailer performance, contributing to improved supplier selection and business decision-making processes. A second analysis, enabled by this methodology, also highlighted key performance differences among the retailers in areas such as customer service, delivery experience, and return/refund processes. Among the findings, Target and Amazon showed the strongest overall sentiment performance, while Costco excelled in return policies and Walmart exhibited weaker results in customer service and delivery. As a result, this hybrid methodology offers valuable insights for both decision-makers aiming to optimize supplier selection and customers seeking better shopping experiences.

Large Language Models (LLMs) have gained popularity in recent years due to their ability to generate human-like text and conduct context-aware conversations in natural languages. This ability can be greatly beneficial for fields like software development, where LLMs can assist with tasks such as code generation, code review, and debugging. In this paper, thematic analysis has been performed on unstructured opinions obtained from 11 experts about the integration of LLMs in the field of software development to understand their benefits and limitations using two natural language processing (NLP) techniques: sentiment analysis and keyword extraction and analysis. Sentiment analysis suggests that most experts were optimistic and pragmatic about the use of generative artificial intelligence in software development, although some experts engaged in critical reflection. Keyword extraction and analysis mapped several keywords to pre-defined themes, which highlighted benefits of LLMs such as improved code quality and enhanced developer productivity, as well as challenges such as the risk of over-reliance, and privacy and security concerns.

Benford's Law describes the distribution of numerical patterns, specifically focusing on the frequency of the leading digit in a set of natural numbers. It divides these numbers into nine groups based on their first digit, with the largest category comprising numbers beginning with 1, followed by those starting with 2, and so on. Each neuron within a neural network (NN) is associated with a numerical value called a weight, which is updated according to specific functions. This research examines the Degree of Benford's Law Existence (DBLE) across two language model methodologies: (1) recurrent neural networks (RNNs) and (2) long short-term memory (LSTM). Additionally, this study investigates whether models with higher performance exhibit a stronger presence of DBLE. Two neural network language models, namely: (1) simple RNN and (2) LSTM, were selected as the subject models for the experiment. Each model is tested with five different optimizers and four different datasets (textual corpora selected from Wikipedia). This results in a total of 20 different configurations for each model. The neuron weights for each configuration were extracted at each epoch, and the following metrics were measured at each epoch: (1) DBLE, (2) training set accuracy, (3) training set error, (4) test set accuracy, and (5) test set error. The results show that the weights in both models, across all optimizers, follow Benford's Law. Additionally, the findings indicate a strong correlation between DBLE and the performance on the training set in both language models. This means that models with higher performance on the training set exhibit a stronger correlation of DBLE.

Artificial intelligence (AI) is rapidly transforming decision-making in business and entrepreneurship, with particularly significant implications for engineering and product development. This paper reviews existing literature and theoretical models to elucidate AI's role in strategic decision-making, while also identifying critical gaps in current research. To gain a comprehensive perspective, we employed a mixed-methods approach comprising surveys of 105 industry professionals and semi-structured interviews with key stakeholders. Our findings indicate that, although AI integration improves operational efficiency and enhances strategic insights, challenges related to data privacy, ethical concerns, and workforce training persist. These results underscore the need for balanced human–AI collaboration and robust governance frameworks to fully realize AI's potential in complex decision-making environments.

This study seeks to put repetitiveness characteristics into AI. Closer ties between AI and human psychology can enhance the implementation of chatbots. Repetitiveness is a common characteristic of human behavior. Repetitiveness indicates which node is updated frequently and its importance. A chatbot needs to solve a situation regarding how quickly it will access its neural memory to retrieve information. Thus, the ranking of nodes in a neural network is necessary to allocate them to the chatbot's memory. The proposed ranking methodology takes affinity, number of edges, adjacency, average weight, and update time interval parameters into account to calculate the ranked value of each node. After that, a ranking tree is generated. This tree is finally considered the memory navigation path in that neural graph. If a node updates regularly with each clock pulse, which resembles a repetitive task, then its ranked value increases. This node should get preference over other low-ranked nodes. This study provides an approach to convert a neural graph into a ranking tree and a path to navigate through it. Thus, the chatbot can identify which node is more promising and has a shorter path than other nodes for information retrieval.

This study presents a robust age-invariant face recognition framework, addressing challenges posed by age-related facial variations. Evaluated on the FGNet and Morph II datasets, the system integrates Viola-Jones for face detection, SIFT and LBP for feature extraction, and Vision Transformers (ViTs) for global feature representation. Feature fusion and dimensionality reduction (KPCA, IPCA, UMAP) enhance efficiency while retaining key discriminative information. Using Random Forest, KNN, and XGBoost classifiers, the model achieves 96% accuracy, demonstrating the effectiveness of combining traditional and deep learning techniques in advancing age-invariant face recognition.

Artificial Intelligence (AI) is set to become an essential tool for defending against machine-speed attacks on increasingly connected cyber networks and systems. It will allow self-defending and self-recovering cyber-defence agents to be developed, which can respond to attacks in a timely manner. But how can these agents be trusted to perform as expected, and how can they be evaluated responsibly and thoroughly? To answer these questions, a Test and Evaluation (T&E) process has been developed to assess cyber-defence agents. The process evaluates the performance, effectiveness, resilience, and generalizability of agents in both low- and high-fidelity cyber environments. This paper demonstrates the low-fidelity part of the process by performing an example evaluation in the Cyber Operations Research Gym (CybORG) environment on Reinforcement Learning (RL) agents trained as part of Cyber Autonomy Gym for Experimentation (CAGE) Challenge 2. The process makes use of novel Measures of Effectiveness (MoE) metrics, which can be used in combination with performance metrics such as the RL reward. MoE are tailored for cyber defence, allowing a greater understanding of agents' defensive abilities within a cyber environment. Agents are evaluated against multiple conditions that perturb the environment to investigate their robustness to scenarios not seen during training. The results from this evaluation process will help inform decisions around the benefits and risks of integrating autonomous agents into existing or future cyber systems.

Bats are capable of highly dexterous flight maneuvers that rely heavily on highly articulated hand skeletons and malleable wing membranes. To understand the underlying mechanisms, large amounts of detailed data on bat flight kinematics are required. Conventional methods to obtain these data have been based on tracing landmarks and require substantial manual effort. To generate 3D reconstructions of the entire geometry of a flying bat in a fully automated fashion, the current work has developed an approach where the pose of a trainable articulated mesh template that is based on the bat's anatomy is optimized to fit a set of binary silhouettes representing views from different directions of the flying bat. This is followed by post-processing to smooth the reconstructed kinematics and simulate the non-rigid motion of the wing membranes. To evaluate the method, 10 flight sequences that represent several flight maneuvers (e.g., straight flight, takeoff, u-turn) and were recorded in a flight tunnel instrumented with 50 synchronized cameras have been reconstructed. A total of 4975 reconstructions are generated in this fashion and subject to qualitative and quantitative evaluations with promising results. The reconstructions are to be used for quantitative analyses of the maneuvering kinematics and the associated aerodynamics.