Mental health assessment presents unique challenges in healthcare due to its inherently subjective nature and the scarcity of high-quality training data. This research explores the use of large language models (LLMs) to generate synthetic datasets for Human-AI teaming algorithms, focusing on mental health assessments. We created a diverse dataset that simulates human-AI collaboration scenarios in diagnostic processes. Synthetic data are labeled through an innovative approach that involves two human annotators and three LLMs, using majority voting for consensus-based annotations. The dataset initially achieves a similarity score of 83.1%, which is further improved by considering human factors that influence decision-making, with varying performance across different mental health categories, highlighting the need for targeted improvements in data collection and model architecture. Our approach addresses several key challenges in the field, including the lack of real-world training data, privacy concerns, and the need for diverse training datasets. This study serves as a foundation for future work in this critical area, which could lead to more effective and ethically sound AI-assisted mental health assessment tools.

The increasing use of large language models (LLMs) in mental health support necessitates a detailed evaluation of their recommendation capabilities. This study compared four modern LLMs-Gemma 2, GPT-3.5-Turbo, GPT-4o, and Claude 4 Sonnet-in recommending mental health applications. We constructed a structured dataset of 55 mental health apps using RoBERTa-based sentiment analysis and keyword similarity scoring, focusing on depression, anxiety, ADHD, and insomnia. Baseline LLMs demonstrated inconsistent total accuracy (ranging from 60% to 75%) and often relied on outdated or generic information. In contrast, our retrieval-augmented generation (RAG) pipeline enabled all models to achieve 100% accuracy, while maintaining good diversity and recommending apps with significantly better user ratings. These findings demonstrated that dataset-enhanced LLMs, regardless of being open-source or proprietary, can excel in domain-specific applications like mental health resource recommendations, potentially improving accessibility to quality mental health support tools.

This paper presents a case study on predicting daily customer flow in a university’s self-service restaurant. We conduct a systematic comparison of multiple machine learning techniques and diverse feature sets to identify the best-performing model within our experimental scope, aiming to improve forecasting accuracy in this dynamic environment. We analyze real-time data collected via RFID sensors from spring 2019 to 2024. To ensure high data quality, we apply a robust preprocessing pipeline, followed by careful feature engineering to select 10 distinct features, labeled M1 to M10. These features include temporal attributes such as day, month, year, and season, as well as external factors like local weather conditions, public holidays, and menu choices. We conduct a systematic comparison across all feature sets and identify M10 as the optimal combination. A key finding highlights the importance of handling missing data-particularly during the COVID-19 period-as one of the most critical steps in the preprocessing stage. To evaluate the predictive power of our selected features, we tested various machine learning models, including linear regression, random forest, extreme gradient boosting (XGBoost), and long short-term memory (LSTM). Our findings indicate that XGBoost achieves the lowest mean absolute error (MAE) and mean squared error (MSE) values. The XGBoost model outperforms other models across all the feature sets, M1 to M10. XGBoost is particularly effective because it uses past data and a technique called exponential smoothing to understand what customers will do in the short term. Our analysis identifies the most influential features as the previous day’s customer count, exponential smoothing outputs, holidays, day of the week, and weather data. Overall, we recommend XGBoost as the most effective model for predicting daily customer numbers in similar contexts, given its superior accuracy across diverse feature sets.

Although the field of machine learning in health and well-being has experienced significant growth in recent years, the use of sensitive data in these applications is often restricted. While sports and well-being data may appear less sensitive, they are frequently subject to the same standards as health data. Synthetic data generation has emerged as a potential solution to privacy issues, aiming to replicate the properties of real data without disclosing identifiable information. In this study, we generated synthetic recreational runner data using a combined variational autoencoder and neural network model based on baseline measurements and training responses from a three-month training period. We then evaluated the synthetic data by training predictive models and comparing their performance to models trained on real data, with additional metrics measuring statistical similarity and privacy. While some challenges remain, particularly regarding the modeling of rare cases, handling of missing data, and ensuring privacy, our results demonstrate that synthetic data could be used to train predictive models with performance comparable to that of models trained on real data.

In this paper, we present the Enhanced Mountain Gazelle Optimizer (EnMGO), a new variant of the mountain gazelle optimizer developed to address the persistent challenge of balancing exploration and exploitation in high-dimensional optimization. While earlier modifications of the algorithm focused on reformulating the control F-parameter to improve stability, they remained limited in adaptability and convergence efficiency. The proposed EnMGO introduces two modifications: An inertia weight based on a chaotic random technique to improve global exploration, and an exponentially decreasing formulation of the F-parameter to enhance local exploitation. These mechanisms create a more adaptive search process capable of efficiently navigating complex solution spaces. The algorithm was evaluated on a set of standard benchmark functions and engineering design problems, and the results demonstrated that EnMGO consistently outperforms previous variants, achieving faster convergence and higher-quality solutions. It achieved better results in approximately 86.96% of the total 23 benchmark functions considered in the study, thereby highlighting its robustness and effectiveness. Furthermore, in the application to engineering design optimization problems, the EnMGO consistently outperformed the comparative algorithms across both design cases, reaffirming its potential as a highly reliable and efficient optimization tool. Based on the performance, the potential of EnMGO can be explored for application in real engineering optimization problems.

This paper applied a formal method based on timed automata and the Uppaal toolbox to the specification and verification of mutual exclusion algorithms operating on anonymous memory. Anonymous memory is a challenging variation of shared memory where common variables (registers) do not have globally agreed-upon names. Instead, each process has its own view and naming for the shared registers. In addition, the cooperating/competing processes are supposed to be symmetric, that is, they execute the same code, and process identifiers are unknown. A recent proposal of Taubenfeld’s mutual exclusion algorithm with m ≥7 (m odd) shared anonymous registers was chosen as a concrete case study. Such an algorithm was informally studied by the authors together with the establishment of some possibility and impossibility results. This paper's first contribution was to confirm the correctness of the algorithm for N = 2 processes when atomic registers are used. The solution no longer holds for a smaller number of registers, e.g., m = 5, or when non-atomic registers are used. The paper also showed that for N > 2, the algorithm violates the mutual exclusion. To extend the use of the algorithm to N > 2 processes, this paper developed a novel organization where the basic solution for two processes is used as the arbitration unit in a binary, standard state-of-the-art tournament tree. Some timed variations of the model were considered, which permit the study of the algorithm under model checking and/or statistical model checking, for various values of the number N of processes.

Total organic carbon (TOC) is used to determine the total amount of organic compounds in water. It has been used to indicate water purity levels for industrial water for decades. Analyzing TOC in water is often time-consuming and an expensive activity, requiring the use of multiple high-precision sensors. Our main aim of this study was to compare the use of the Random Forest (RF) algorithm, one-dimensional convolutional neural network (1D-CNN), and multilayer perceptron (MLP) in predicting TOC using water quality parameters selected based on their strength of correlation. RF was chosen because it can model complex interactions between the various parameters and is resistant to overfitting. 1D-CNN can handle local dependencies and spatial relationships between input features whereas MLP handles independent numerical features in the dataset. The dataset was obtained from analysis done in Puget Sound marine waters around Seattle King County in the USA at the Duwamish River at three sampling locations. Learning curve analysis demonstrated that the dataset size was sufficient for stable training and generalization, while five-fold cross-validation confirmed consistent model performance across data splits. The effects of wet and dry seasons on the parameter levels were done and their impact on RF model accuracy was assessed. The selection of parameters based on Gini importance ranking was done to evaluate their effect on the accuracy of the RF model. Our results indicated that the prediction of TOC using RF regression was the most accurate with a coefficient of determination (R²) of 0.732. The 1D-CNN and MLP had R² of 0.714 and 0.638, respectively. The mean absolute error (MAE) for RF, 1D-CNN, and MLP were 0.120 mg/L, 0.244 mg/L and 0.270 mg/L, respectively. It was concluded that the RF algorithm would be more feasible in predicting TOC in river water than the two deep learning methods.

Human activity recognition (HAR) refers to the process of identifying human actions and activities using data collected from sensors. Neural networks, such as convolutional neural networks (CNNs), long short-term memory (LSTM) networks, convolutional LSTM, and their hybrid combinations, have demonstrated exceptional performance in various research domains. Developing a multilevel individual or hybrid model for HAR involves strategically integrating multiple networks to capitalize on their complementary strengths. The structural arrangement of these components is a critical factor influencing the overall performance. This study explored a novel framework of a two-level network architecture with dual-stage feature fusion: late fusion, which combines the outputs from the first network level, and intermediate fusion, which integrates the features from both the first and second levels. We evaluated 15 different network architectures of CNNs, LSTMs, and convolutional LSTMs, incorporating late fusion with and without intermediate fusion, to identify the optimal configuration. Experimental evaluation on two public benchmark datasets demonstrated that architectures incorporating both late and intermediate fusion achieve higher accuracy than those relying on late fusion alone. Moreover, the optimal configuration outperformed baseline models, thereby validating its effectiveness for HAR.

Fresh foods are essential products in the global food industry, offering consumers vital nutrients and health benefits worldwide. Despite advancements in freshness classification using image-based data, the literature lacks exploration of quantitative data in this field. In this study, we collected a real-world sensor dataset for food freshness classification during runtime by monitoring and recording environmental and chemical variables that affect food quality, using bananas as a case study. The collected dataset was pre-processed and used to train and test six machine learning models: logistic regression, random forest, support vector machine, K-Nearest Neighbor, decision tree, and gradient boosting. These models were employed for the automatic classification of banana freshness into three health classes: fresh, ripening, and spoiled. The results revealed that the random forest model outperforms other models in predicting banana health class, achieving an average accuracy of 95%. Additionally, we critically analyzed the collected data and provided actionable insights for stakeholders and professionals in the food industry, enabling them to make informed decisions that maintain product quality and reduce food waste.

This paper introduces a multi-agent framework for comprehensive highway scene understanding, designed around a mixture-of-experts strategy. In this framework, a large generic vision-language model (VLM), such as GPT-4o, is contextualized with domain knowledge to generate task-specific chain-of-thought prompts. These fine-grained prompts are then used to guide a smaller, efficient VLM in reasoning over short videos, along with complementary modalities as applicable. This framework simultaneously addresses multiple critical perception tasks including weather classification, pavement wetness assessment, and traffic congestion detection, which achieve robust multi-task reasoning while balancing accuracy and computational efficiency. To support empirical validation, we curated three specialized datasets aligned with these tasks. Notably, the pavement wetness dataset is multimodal, combining video streams with road weather sensor data, highlighting the benefits of multimodal reasoning. Experimental results consistently demonstrate the strong performance across diverse traffic and environmental conditions. From a deployment perspective, the framework can be readily integrated with existing traffic camera systems and strategically applied to high-risk rural locations, such as sharp curves, flood-prone lowlands, and icy bridges. By continuously monitoring the targeted sites, the system enhances situational awareness and delivers timely alerts, even in resource-constrained environments.

Police narrative reports from 911 responses contain valuable signs for early behavioral health intervention, but obtaining them manually is time-consuming and tedious. We present RECAP, a human-AI collaboration framework that fine-tunes three large language models: Mistral-7B, Llama-3-8B, and TinyLlama to classify BH cases in narratives and generate short, sentence-level rationales. The models are trained on an annotated corpus using supervised instruction data and then aligned using direct preference optimization (DPO), allowing patrol officer preferences to continually influence system behavior without disrupting existing workflows. On a held-out test set, Mistral-7B achieves 85.2% weighted accuracy and 84.2% F1-score, matching strong prior baselines while improving interpretability by short, text-span-linked explanations; Llama-3-8B performs similarly, while TinyLlama provides competitive accuracy at a lower compute cost. RECAP is designed to reduce manual effort and identify behavioral health signs earlier in public narratives, while providing rationales and maintaining officer control.

Coordinated truck-drone delivery has emerged as an important extension of the vehicle routing problem (VRP), offering new opportunities to improve logistics efficiency, accessibility, and sustainability. In this review, we synthesize the literature on truck-drone routing by examining three interrelated themes: Vehicle routing and synchronization models, stochastic and dynamic operational constraints, and payload-energy management. Early studies primarily formulate truck-drone delivery as variants of heterogeneous or two-echelon VRP using mixed-integer linear programming (MILP) to capture coupled routing decisions, drone endurance limits, and launch-retrieval feasibility, often supplemented by heuristic or metaheuristic strategies to address computational scalability. More recent research extends these formulations to stochastic and dynamic VRP settings by incorporating uncertain customer availability, time windows, vehicle delays, and weather disruptions, reflecting a shift toward adaptive routing under incomplete information. Parallel work on payload configuration and energy management investigates modular loading and battery-swapping mechanisms, demonstrating how load heterogeneity and energy constraints fundamentally reshape feasible routing structures. Extensions to multimodal VRP variants, including electric vehicle-drone systems and van-robot hybrids, further broaden applicability in dense and constrained urban environments. Despite these advances, most models remain grounded in static or weakly stochastic assumptions, with limited support for real-time updates or predictive decision-making. In this review, we identified the integration of data-driven prediction and learning-based optimization within dynamic VRP frameworks as a critical direction for advancing truck-drone delivery toward large-scale operational deployment.