Corporate entrepreneurship is crucial for fostering adaptability and innovation in today’s dynamic business landscape. This research focuses on enhancing business models (BMs) as catalysts for transformation and innovation, using a systematic approach that integrates Frankenberger’s 4I framework with De Reuver’s business model roadmap. Through a four-phase methodology—ecosystem analysis, ideation, integration, and implementation—this study refines BMs within a defined case context. A case study of Gold Gallery, a traditional gold retail enterprise, demonstrates the methodology’s practical utility. By applying the framework, four tailored business models were developed to address context-specific challenges. The study supports the model’s robustness through expert panel evaluations. This research contributes to corporate entrepreneurship literature by delineating a systematic, repeatable process for BM-driven innovation. While the single industry case provides contextual richness, future research should explore cross sector applicability and long term performance impacts.

Lymphography involvement is a fundamental indicator of cancer progression, making the precise classification of lymphographic findings essential for reliable diagnosis and treatment decision-making. This study presents an embedded approach for lymphography cancer detection that synergistically combines a genetic algorithm (GA) and an instance-based learning classifier. Unlike prior GA hybrids, this approach incorporates GA feature selection with K*, a relative-entropy-based approach that handles noisy, small-sample-size, class-imbalance, and redundant features in lymphography data, rather than the traditional hybrid approach that used KNN with a geometric distance metric (typically Euclidean or Manhattan). Although vital for diagnosing lymphatic diseases, lymphography poses challenges because of the data it generates. To address these challenges, the proposed approach integrates the strengths of GA to optimize the most important features. Instance-based learning (i.e., K* classifier) is used for effective pattern recognition and to classify lymph nodes as benign, normal, metastatic, malignant, or fibrotic. Extensive experimental results on a real-world dataset demonstrated that the combination significantly outperformed state-of-the-art classification algorithms and other hybrid approaches. The efficiency of the proposed approach is evaluated using the most widely used computational metrics, yielding promising results for diagnosing lymphatic diseases. This work highlights the potential of an evolutionary-based embedded classifier for classifying complex biomedical data.

Generative artificial intelligence (AI) is increasingly used in education, but its role in the development of an entrepreneurial mindset remains underexplored. We propose Reciprocal Prompt Co-Adaptation (RPCA), a closed-loop human–AI scaffolding framework that supports the development of an entrepreneurial mindset through iterative prompt refinement between students and GPT-4. In this framework, “reciprocal” refers to interaction-level adaptation: students provide ratings and reflections, and the system revises prompts and coaching strategies accordingly. It does not imply that GPT-4 learns in the human sense or updates its model parameters. Unlike static or one-sided adaptive systems, RPCA integrates three components: (i) a prompt alignment module that incorporates student ratings and reflections through meta-prompting, (ii) an entrepreneurship-specific reward-shaping engine inspired by reinforcement learning from human feedback, and (iii) a confidence tracking layer that adjusts scaffolding and challenge based on textual, behavioral, and brief self-report signals. In an eight-week randomized experiment with 120 Thai undergraduates across four conditions, namely RPCA, static templates, adaptive prompting, and human–AI hybrid, RPCA produced significantly larger pre-to-post gains in opportunity recognition, risk propensity, and creative self-efficacy compared with all baseline conditions. It also achieved higher session-level prompt relevance and faster convergence to effective prompts. An embedded ablation study further suggested that removing reward shaping reduced advantages in opportunity recognition and creative self-efficacy, whereas removing confidence tracking reduced creative self-efficacy and perceived prompt quality. To support transparency and replicability, the interaction protocol, rubric-guided scoring criteria, and illustrative examples of prompt revision are specified in the appendices. These findings provide initial evidence that off-the-shelf large language models, when embedded in structured, feedback-driven tutoring protocols, can serve as adaptive, dialogic scaffolds rather than static content providers, offering a scalable approach to supporting the development of a nonlinear mindset in entrepreneurship education.

Artificial intelligence (AI) is widely assumed to enhance creativity in knowledge work, yet empirical findings reveal substantial variability in its effects. Addressing this inconsistency, the present study advances a metacognitive explanation by integrating metacognitive calibration theory with AI engagement research. We propose that calibration bias, the discrepancy between perceived and actual cognitive competence, influences creativity indirectly through behavioral engagement with AI systems as algorithmic support tools. Drawing on social cognitive theory, we develop a mediation model in which calibration bias predicts AI engagement, which in turn predicts supervisor-rated creativity. We further examine whether AI use intensity moderates this relationship. Using a time-lagged, multi-source design, data were collected from 437 employees nested within 58 supervisors in Islamic banking institutions and analyzed using Mplus-based structural modeling. Calibration bias positively predicts AI engagement, which in turn enhances creativity. Bootstrapped mediation analysis indicates a significant positive indirect effect of calibration bias on creativity through AI engagement, supporting partial mediation. However, AI use intensity does not significantly moderate the relationship between calibration bias and AI engagement. The findings contribute to theory by extending metacognitive calibration research into AI-augmented organizational contexts and reframing AI-assisted creativity as a metacognitively contingent process within AI-enabled work systems rather than a purely technological outcome. By identifying AI engagement as the behavioral mechanism linking metacognitive discrepancy to creative performance, this study advances a process-based understanding of human–AI collaboration.

Innovation teams in regulated digital-service environments routinely encounter hard architectural contradictions, situations in which two non-negotiable requirements are in direct technical conflict. Despite significant scholarly interest in Theory of Inventive Problem Solving (TRIZ)-based contradiction resolution and the institutionalization of Design Thinking in software engineering and management information systems, no existing framework has formally coupled these two paradigms with the generative potential of large language models into a unified, human-centered innovation process. The current study presents TriGenDT, a five-stage Design Science Research artifact that integrates TRIZ contradiction resolution, Design Thinking, and generative artificial intelligence to enable systematic innovation in healthcare management information systems. The framework is illustrated through a constructed clinical decision-support scenario that highlights the tension between real-time alert latency and privacy/compliance constraints. The demonstration illustrates process coherence, stage-to-tool traceability, and design-level feasibility; it does not constitute field validation. Projected values, such as alert latency (~30 ms) and alert-path call-depth reduction (~50%, from six to three hops), are design-level estimates that require empirical confirmation. The study contributes a formally specified TRIZ-to-software-engineering parameter mapping, an integrated innovation methodology grounded in a reproducible demonstration case, a preliminary expert appraisal, representative prompt templates with model parameters, and an openly available artifact repository. Future research should pursue empirical validation through live industry case studies and comparative evaluation against TRIZ-only, Design Thinking-only, and large language model-native TRIZ baselines.

Accurate evaluation of reference evapotranspiration (ETo) is important for effective long-term irrigation planning and water resource management in arid/semi-arid regions. This study evaluates four models: linear regression, Takagi–Sugeno fuzzy inference system, random forest (RF), and support vector regressor (SVR) to estimate daily ETo. The models were trained using ERA5-Land daily climatic variables across 24 meteorological stations in the Batna region (Algeria). To address potential temporal leakage in meteorological time series, we evaluated both random splitting (70% training, 15% validation, and 15% testing) and chronological splitting. The SVR model, utilizing only three automated features—air temperature at 2 m, soil temperature at 0–7 cm, and vapor pressure deficit—demonstrated significantly better performance than the other models at each station, with test root mean square error (RMSE) values ranging from 0.544 mm/day (Tazoult) to 0.780 mm/day (Bitam) and Nash–Sutcliffe efficiency values ≥ 0.91. RF showed severe overfitting, with a 164.93% increase in test RMSE. To further validate SVR’s robustness, we compared it against two additional benchmark models from the literature: a multilayer perceptron (MLP) and a light gradient boosting machine (LightGBM). SVR consistently outperformed both MLP (RMSE: 0.645 mm/day) and LightGBM (RMSE: 0.612 mm/day). Therefore, the results suggest that the SVR model utilizing only three climatic features provides a highly novel, computationally efficient, and accurate platform for predicting daily ETo in arid/semi-arid Mediterranean climates, offering an operational solution for irrigation scheduling in data-scarce environments.

Speech therapy for hearing-impaired children in Sri Lanka is significantly limited by factors such as cost constraints and the lack of automated, language-specific tools to aid in therapy sessions. This study presents a Sinhala auditory training platform for hearing-impaired children that combines adaptive learning with automatic pronunciation evaluation. The platform uses Bayesian knowledge tracing and multi-armed bandit task sequencing to estimate learner competence and personalize training. Its task system is based on four language comprehension task types that serve as blueprints for automatically generating activities of varying difficulty, supported by 353 audio assets. The platform covers phoneme discrimination, syllable processing, and word recognition, while providing analytics for therapists. A pronunciation evaluation module built on a fine-tuned Wav2Vec2 model delivers phoneme-level feedback using forced alignment and goodness of pronunciation scoring. To support this module, 1,534 Sinhala speech recordings from hearing-impaired speakers were collected for training. Experimental results using the model achieved a phoneme error rate of 0.289, demonstrating stable convergence during training. The system highlights the feasibility of combining adaptive tutoring and speech-based feedback for scalable, personalized auditory rehabilitation in Sinhala.

As artificial intelligence (AI) becomes increasingly embedded in university curricula, identifying factors associated with students’ motivation to learn AI has become an important issue in higher education. This study surveyed 1,474 undergraduate students and used structural equation modeling with a validated five-construct framework to examine how AI knowledge and practical use of AI tools were related to students’ perceived demand for AI learning. The results showed that foundational AI knowledge was positively associated with educational demand both directly (β = 0.401, p < 0.001) and indirectly, primarily mediated by learning interest (β = 0.226, p < 0.001). In contrast, practical AI tool competency was not directly associated with educational demand; however, it was indirectly related to learning interest, whereas general optimism about AI was not necessarily associated with a stronger demand for formal AI learning. Multi-group analyses also indicated experience-based differences: students without prior AI learning experience showed a stronger association between hands-on tool interaction and interest, whereas students with prior AI learning experience showed a stronger association between existing AI knowledge and educational motivation. Overall, the findings suggest that higher-education AI literacy programs may benefit from cultivating students’ interest in and engagement with learning, rather than assuming that tool exposure or future-oriented attitudes alone will lead to stronger demand for AI learning.

Skin cancer is the most common malignancy worldwide, and effective early detection remains essential despite advances in prevention and photoprotection strategies. This study introduces an intelligent system for skin cancer diagnosis via deep learning methodologies. The suggested hybrid model combines the best aspects of MobileNet and long short-term memory (LSTM) networks to improve skin lesion image classification. The technology was designed not only to help doctors make accurate diagnoses, but also to support real-world healthcare tasks, including uploading images, automatically analyzing them, suggesting treatments, and scheduling appointments for patients and doctors. The experimental findings indicate that the proposed hybrid model achieves 93% accuracy, surpassing several current models, including support vector machine, convolutional neural network, Visual Geometry Group, ResNet, and MobileNet. Combining MobileNet with LSTM improves the ability to extract and classify features. Early detection of skin cancer can make therapy more effective and lower death rates. The suggested method shows great promise for use in real-world medicine. For future work, the system can be improved by using increasingly diverse datasets, more advanced deep learning architectures, and real-time clinical deployment to make diagnoses more accurate and reliable. The suggested system is intended to facilitate early skin lesion analysis and help medical practitioners with initial diagnosis by acting as an artificial intelligence-assisted screening and decision-support tool.

Cyberattacks targeting Internet of Things (IoT) systems, particularly zero-day exploits, are escalating due to inherent vulnerabilities in IoT networks. Traditional intrusion detection systems (IDSs) use machine learning, such as deep learning (DL), to improve cyberattack detection. Nevertheless, DL-based IDSs require well-balanced datasets with abundant labeled data, which is often not available in IoT networks. In this article, we propose an efficient IDS that incorporates transfer learning (TL), knowledge transfer, and model refinement to accurately identify zero-day attacks. The TL model is based on deep convolutional neural networks adapted to 5G IoT environments with unbalanced and limited labeled datasets. The proposed framework employed three specialized datasets: the University of New South Wales Network-Based 2015 dataset (UNSW-NB15)-Basic for model training, UNSW-NB15-Test+ for evaluating zero-day attack detection, and UNSW-NB15-Test for comprehensive evaluation involving both known and zero-day attacks. The experimental results validate the effectiveness of our approach, achieving high accuracy and low false-prediction rates. In particular, the introduced TL-oriented solution outperforms other DL-based IDSs in detecting various families of known and zero-day attacks, representing a significant step forward in protecting IoT devices against cyber threats.

Accurate long-term solar photovoltaic (PV) power forecasting is essential for renewable energy planning, grid stability, and sustainable energy transition, particularly in regions with high solar potential and climatic variability. This study proposes a weather-integrated hybrid machine learning (ML) and deep learning (DL) framework for long-term solar PV forecasting in Oman using a decade-long daily meteorological dataset (2014–2024) collected from three geographically diverse locations: Izki, Muscat, and Sadah. The framework integrates meteorological variables, including solar radiation, temperature, humidity, rainfall, wind speed, vapor pressure, sunshine duration, and atmospheric pressure. Data preprocessing included missing-value imputation, anomaly correction, normalization, seasonal decomposition, stationarity testing, and cyclical feature engineering to improve forecasting robustness. Conventional statistical approaches (autoregressive integrated moving average, seasonal autoregressive integrated moving average, and exponential smoothing), ML algorithms, and hybrid DL architectures were comparatively evaluated. Results indicate that ensemble regression achieved superior ML performance with validation R² = 0.96, root mean squared error = 38.60, and mean absolute percentage error = 6.16%, while the recursive long short-term memory + convolutional neural network model with residual correction demonstrated enhanced stability for long-term forecasting horizons. Explainability analysis using Shapley additive explanations and partial dependence plots identified latitude, seasonal encodings, sunshine duration, and atmospheric variables as dominant predictors. The proposed framework provides an interpretable and scalable forecasting solution for renewable energy planning in Oman and similar climatic regions.

The increasing adoption of artificial intelligence (AI) has transformed organizational processes across industries, including the auditing profession. AI-powered audit automation enables organizations to enhance audit efficiency, improve analytical accuracy, and reduce operational costs, thereby improving audit quality. Despite the growing adoption of AI technologies in auditing practices, empirical evidence of their impact on audit quality in the United States remains limited. This study examines the effects of AI-powered automation on audit quality by focusing on three key dimensions: efficiency, accuracy, and cost-effectiveness. A quantitative research design was employed using a structured questionnaire distributed to auditors working in organizations across the United States. A total of 207 valid responses were collected and analyzed using correlation and multiple regression analyses to test the proposed hypotheses. The findings reveal that efficiency, accuracy, and cost-effectiveness all have significant positive effects on audit quality. Among these factors, efficiency emerged as the strongest predictor, followed by accuracy and cost-effectiveness. The results indicate that AI-enabled audit automation enhances the reliability, effectiveness, and overall quality of auditing processes by facilitating faster data processing, reducing human error, and optimizing resource utilization. This study contributes to the literature by providing empirical evidence on the role of AI-powered automation in improving audit quality within a highly regulated auditing environment. The findings offer valuable insights for audit practitioners, organizational leaders, and policymakers seeking to leverage AI technologies to strengthen audit performance and improve assurance outcomes. The study also provides a foundation for future research examining AI adoption and audit effectiveness across different industries and institutional contexts.

This study was conducted during the 2025–2026 academic year at Tishk International University in Iraq. It aims to compare teacher-guided ChatGPT feedback and teacher feedback on the writing development of architecture students learning English as a foreign language. Students were divided into control and experimental groups, and a seven-week study was conducted with pre- and post-tests administered. During the study, the control group received teacher feedback, while the experimental group received ChatGPT feedback under continuous teacher guidance and pedagogical supervision, rather than unrestricted artificial intelligence use. Analysis of the data revealed that although both groups showed improvement in writing performance, the experimental group’s post-test scores increased significantly more than those of the control group. Additionally, students in the experimental group reported positive perceptions of ChatGPT feedback. In conclusion, ChatGPT, when used under teacher guidance, is an effective tool for supporting students’ writing skills.

Cross-border e-commerce decisions in emerging corridors often require firms and public partners to choose among platform entry, warehouse investment, payment integration, and local service building before complete country-level online transaction statistics are available. This article develops a systematic innovation corridor design method (SICDM) for China–Central Asia cross-border e-commerce, subject to official data constraints. The method combines auditable trade and digital-readiness evidence with a Theory of the Solution of Inventive Problems-informed contradiction map, Lanchester resource-position logic, and multi-criteria corridor prioritization. Reported partner-country imports from China are used as an official trade benchmark, not as a direct e-commerce turnover metric. The empirical base is a balanced official-data panel covering Kazakhstan, the Kyrgyz Republic, Tajikistan, and Uzbekistan from 2018 to 2023, using the World Integrated Trade Solution/United Nations Comtrade trade data and World Development Indicators. A digital readiness index is constructed from internet use, mobile cellular subscriptions, and secure internet servers; the first principal component explains 78.8% of the variance in information and communication technology. An entropy-weighted technique for order preference by similarity to an ideal solution and a reduced-form panel diagnostic illustrate how official evidence can support corridor comparison without estimating hidden platform sales. Within the 2023 decision matrix, the SICDM classifies Kazakhstan as a consolidation anchor, Uzbekistan as a staged-scaling corridor, the Kyrgyz Republic as a gateway corridor, and Tajikistan as a frontier pilot. The contribution is not a new estimate of cross-border e-commerce turnover, but a reproducible, systematic innovation tool for translating incomplete evidence into corridor-specific contradictions, innovation opportunities, and resource-sequencing actions.

Generative artificial intelligence (AI) is transforming the way music is generated, accompanied, harmonized, pitched, transposed, and generally transformed, created, taught, studied, and evaluated. This is an integrative review that has been structured and synthesized 75 sources from the databases, namely Scopus, Web of Science, IEEE Xplore, ACM Digital Library, Education Resources Information Center (ERIC), APA PsycINFO, Google Scholar, citation searches, and authoritative institutional reports. The literature published between January 1, 2021, and July 7, 2026, was prioritized, while earlier legal, ethical, historical, and theoretical literature was retained when needed for conceptual grounding. The directness of the evidence was coded separately from the evidential weight of the evidence, which included findings from other contiguous technologies, such as automated feedback, intelligent tutoring systems, learning analytics, augmented reality/virtual reality, and online digital audio workstations. The synthesis suggests that activities that are teacher-mediated, process-assessed, culturally responsive, and ethically governed may open up opportunities for generative AI to assist in composition, songwriting, music-theory learning, certain types of feedback on practice, improvisation, and collaborative music-making. Yet, there is a scarcity of direct, controlled, and longitudinal evidence, and issues remain regarding originality, authorship, assessment integrity, cultural bias, data privacy, equitable access, teacher autonomy, and student agency. The review contributes to systematic innovation by connecting the iterative processes of generation, musical evaluation, revision, implementation, and reflection to pedagogical, assessment, and governance frameworks. Generative AI is therefore best understood as a pedagogical and creative tool that is used by a human teacher, not to replace musicianship or the teacher’s judgment.