Purpose: This study aims to address the underutilization of transaction data in the outdoor equipment rental industry by developing an integrated Business Intelligence (BI) dashboard using the FP-Growth algorithm to generate product bundling recommendations and support data-driven decision-making. Methods: A quantitative approach based on the Knowledge Discovery in Databases (KDD) framework was employed using 596 rental transactions from Batas Outdoor Rental recorded between January and May 2026. The data were preprocessed and transformed into a binary matrix using TransactionEncoder. FP-Growth was applied with a minimum support of 2% (0.02), while association rules were generated using a minimum confidence of 30% (0.30) and validated with a lift threshold of 1.20. Results: A quantitative approach based on the Knowledge Discovery in Databases (KDD) framework was employed using 596 rental transactions from Batas Outdoor Rental recorded between January and May 2026. The data were preprocessed and transformed into a binary matrix using TransactionEncoder. FP-Growth was applied with a minimum support of 2% (0.02), while association rules were generated using a minimum confidence of 30% (0.30) and validated with a lift threshold of 1.20. Novelty: This study integrates FP-Growth-based transaction analysis with an interactive BI dashboard specifically for outdoor equipment rentals. Unlike previous studies focusing primarily on product combinations or promotional packages, the proposed approach provides an end-to-end decision-support framework connecting transaction analysis, automated bundling recommendations, and interactive visualization.
Purpose: This study aims to examine the effectiveness of AI-driven influencer marketing, develop an Xception-based Convolutional Neural Network (CNN) for deepfake detection, and analyze ethical and legal responsibility for artificial intelligence use from a rule-of-law perspective. Methods: A mixed-methods approach was employed by integrating Computer Science, Management, and Law perspectives. Quantitative analysis used Partial Least Squares Structural Equation Modeling (PLS-SEM) to examine the relationships among AI Personalization, AI Interaction, User Experience, and Trust. The Xception-based CNN was evaluated using standard classification metrics. Qualitative analysis involved expert interviews and examination of legal principles concerning AI governance, transparency, accountability, and consumer protection. Result: The Xception-based CNN achieved 97.12% accuracy and an AUC of 0.9920 in distinguishing authentic from AI-generated content. PLS-SEM results indicate that AI Interaction significantly influences User Experience and Trust, while User Experience has the strongest effect on Trust (β = 0.520; p < 0.001). The model explains 53.3% of the variance in Trust, and User Experience significantly mediates the relationship between AI Interaction and Trust. The legal analysis highlights the need for stronger AI governance addressing transparency, disclosure of AI-generated content, consent, accountability, and consumer protection. Novelty: The novelty of this study lies in integrating deepfake detection, AI-driven influencer marketing, consumer trust analysis, and rule-of-law perspectives into a unified framework for responsible digital marketing. This integrated approach provides a multidisciplinary perspective for balancing technological innovation with ethical responsibility, legal accountability, and consumer protection.
Purpose: This paper focuses on examining the robustness of a cooperative Multi-Agent Reinforcement Learning (MARL)-based Intrusion Detection System (IDS) for intrusion detection in decentralised 5G security settings. Even though MARL techniques have proven effective against dynamic threats in decentralized 5G networks, current research has not considered any adversarial scenarios at all. Methods: A cooperative MARL-based Intrusion Detection System was developed through the CRISP-DM approach. Radio Access Network (RAN), MEC, and Core agents were trained using Centralised Training with Decentralised Execution (CTDE) and Deep Q-Network (DQN) methods. The algorithm was tested on the NSL-KDD and UNSW-NB15 datasets against Fast Gradient Sign Method (FGSM) evasion attacks (ε = 0.05-0.30) and Byzantine poisoning attacks with 5%, 10%, and 20% compromised agents. Result: The model achieved 96.94% accuracy on NSL-KDD and 85.15% on UNSW-NB15 in clean scenarios. The FGSM attack at ε = 0.20 resulted in substantial performance deterioration, leading to accuracy drops of 50.14 and 45.26 percentage points, respectively, and a simultaneous increase in false positives. Byzantine poisoning produced smaller but persistent decreases in accuracy of 12.03 and 2.62 percentage points, respectively. Novelty: This study provides among the first empirical evaluations of adversarial fragility in cooperative MARL-based intrusion detection within distributed 5G-oriented security abstractions, demonstrating that cooperative intelligence alone does not guarantee adversarial robustness.
Purpose: To evaluate the impact of image preprocessing techniques, specifically contrast enhancement and noise reduction, on improving the CNN performance on Pap smear image classification for early cervical cancer detection. Methods: Three CNN architectures (ResNet34, DenseNet121, and MobileNet-V2) were trained and evaluated on the SIPaKMeD dataset. Two preprocessing techniques were applied: Contrast Limited Adaptive Histogram Equalization (CLAHE) for contrast enhancement and Perona-Malik Diffusion (PMD) filter for noise reduction. Model performance was assessed using a confusion matrix. Results: Preprocessing improved classification performance across all models. CLAHE significantly increased the accuracy of ResNet34 from 76.73% to 84.16% and DenseNet121 from 83.17% to 84.16%, while also providing modest improvement for MobileNet-V2. In contrast, PMD filtering yielded limited improvement and, in some cases, slightly reduced model performance. Novelty: This study provides a systematic comparison of contrast enhancement and noise reduction techniques across multiple CNN architectures. This study demonstrates that contrast enhancement is more effective than noise reduction in improving CNN performance. The study provides new pipelines for improving cervical cancer classification.
Purpose: Since glioma is the most aggressive and infiltrative type of brain tumor, its detection in magnetic resonance imaging (MRI) is especially difficult. Despite the excellent overall accuracy for brain tumor detection with YOLOv8-based object detection, the glioma-specific performance is limited owing to ambiguity of tumor boundaries. This work seeks to elucidate if boundary-aware learning can enhance glioma detection beyond typical bounding box–based approaches. Methods: This study focuses exclusively on glioma detection using the Cheng brain tumor MRI dataset. YOLOv8 is used as the baseline detector, and boundary-aware learning is implemented through segmentation supervision using YOLOv8-Seg by leveraging pixel-level tumor masks. All the experiments are done in a standardized training environment to allow fair and unbiased comparison. Result: Experimental evaluation shows that YOLOv8-Seg achieved a detection precision of 0.899, recall of 0.905, and mAP@50 of 0.940, while segmentation results achieved a mask precision of 0.900, recall of 0.904, and mAP@50 of 0.943. For glioma-specific analysis, the model achieved a box mAP@50 of 0.875 and a mask mAP@50 of 0.877. These results indicate that segmentation supervision improves spatial boundary representation even though improvements in conventional detection metrics remain marginal. Novelty: Unlike the other works that are based on augmentation of the data and performance of better detection, this work has devised a glioma-centric design, and shows bounding box-based detection is insufficient. This work highlights the need for considering boundary aware learning applying the supervision of segmentation in the automated glioma detection system, which can improve the reliability and interpretability of the system.
Purpose: This study aims to develop and evaluate a tree-based survival machine learning framework for predicting graduate study duration and identifying students at risk of delayed graduation. The study addresses the growing need for accurate educational time-to-event prediction to support academic monitoring and data-driven decision-making in higher education institutions, including IPB University. Methods: A quantitative predictive analysis was conducted using data from 3,417 master students from the 2020–2022 cohorts. Four tree-based survival models were evaluated, namely Survival Tree (ST), Extremely Randomized Survival Tree (EST), Random Survival Forest (RSF), and Gradient Boosting Survival (GBS). The analysis used right-censored survival data with a 42-month observation period. Model evaluation was conducted using repeated stratified random split validation (10 repetitions) with Concordance Index (C-index) and Integrated Brier Score (IBS) metrics. Risk stratification was subsequently performed using the best-performing model based on predicted survival probabilities at a 24-month time horizon. Result: GBS achieved the best overall predictive performance with the highest mean C-index (0.659) and the lowest mean IBS (0.194), indicating superior discrimination and prediction accuracy compared to ST, EST, and RSF. The repeated evaluation results also demonstrated stable predictive performance across data partitions. Risk stratification successfully separated students into low-, medium-, and high-risk groups with significantly different survival patterns. High-risk students generally tended to be older, have lower undergraduate GPA, were more often male, and more frequently originate from private undergraduate institutions. Novelty: This study provides a comparative evaluation of multiple tree-based survival machine learning models within an educational time-to-event framework. The integration of repeated survival model evaluation with practical student risk stratification offers both methodological and applied contributions for academic monitoring and early intervention strategies in higher education.
Purpose: Good governance practices in transparency, accountability, responsibility, fairness, and stakeholder participation are core principles that ensure governance mechanisms are not merely procedural, but also creates a sustainable service management environment that aligns governance conditions with public interests. Methods/Study design/approach: Through the integration of COBIT 2019 and ITIL v4 frameworks, along with structured staff interviews, service performance analysis, and direct operational observations, several important domains emerged, indicating the need for management development. Result/Findings: The results show that several COBIT 2019 domains, APO07 (Managed Human Resources), DSS01 (Managed Operations), BAI03 (Managed Solutions Identification and Build), EDM02 (Ensured Benefits Delivery), and MEA01 (Managed Performance and Conformance Monitoring) and ITIL v4 domains, Information Security Management, Organizational Change Management & Workforce, Continual Improvement require immediate improvement with the Capability Index scored 46.875% classified as Partially Achieved (15–49%) and classified in Level 2 (Managed Process). COBIT 2019 Design Factors show urgent areas for improvement in DF-4 (IT Related Issues), DF-6 (Compliance Requirements), and DF-7 (Role of IT). Balanced Scorecard (BSC) was also evaluated governance performance from four perspectives, financial, customer, internal process, and learning & growth, scoring 69,4% and classified as moderate. Key Performance Indicator (KPI) recommendations were also proposed, e.g., staff training increasement, operational IT services availability, user-based IT solutions, IT process measurement, security accidents reporting, effective workforce, and user-oriented IT developments. Novelty/Originality/Value: This study proposes an integrated COBIT 2019 and ITIL v4 approach, embedding good governance principles for sustainable service management in Museum Electronic-Based Government System (SPBE) and addressing the gap between governance conditions and public interests.
Purpose: The purpose of this study is to design a mobile-based attendance application that uses Location-Based Services to support community-based religious institutions, namely mosques, which are simple to understand for all age groups, make it easier for participants to take attendance, and make it easier for administrators to manage participant data, schedules, and attendance. Methods: Interviews and direct observations were conducted to identify problems and determine the new system required. This system utilizes the concept of geolocation by utilizing GPS and geofencing for location validation, which is described using UML and implemented in programming. The completed system was tested through blackbox testing for its features and GPS accuracy testing to determine the stability of the location tracking system. Result: Blackbox testing results show that the functions integrated into the attendance application can run 100% as expected. The GPS accuracy system performance test results obtained are sufficient for semi-outdoor and outdoor spaces with 100% valid results. Novelty: The uniqueness of this research lies in its contribution to community-based religious institutions, namely mosques, in digital attendance recording, as well as providing GPS accuracy evaluation results provided by the system to determine the most appropriate radius range. Further developments that can be made to the system include improving security procedures, as this application directly accesses the user's location, making improvements to the security system and fake location detection very important for attendance applications.
Purpose: This study evaluates the level of Personal Data Protection (PDP) compliance among Electronic System Providers (ESPs) in Indonesia based on the Personal Data Protection Law (PDP Law) and the Government Regulation concerning the Implementation of Electronic Systems and Transactions. Methods: A descriptive-evaluative approach was conducted through observations of website and mobile application interfaces from 20 ESPs registered with the Ministry of Communication and Digital Affairs. Compliance was assessed using a binary scoring system based on six PDP indicators: consent mechanisms, privacy notices, TLS/SSL implementation, data disclosure, malicious libraries, and device data access. Descriptive statistical analysis was used to evaluate compliance levels. Instrument validity was established through content validity and expert judgment. Result: Most ESPs were classified within the moderate compliance category, covering 90% of websites and 80% of mobile applications. Governance-related indicators showed the lowest compliance levels, particularly website consent mechanisms (15%) and website privacy notices (40%) and mobile consent and privacy notice compliance (20%). In contrast, all ESPs complied with technical indicators, including TLS/SSL, malicious library, and device data access requirements. Novelty: Unlike previous studies that focused on single sectors or platforms, this study provides a cross-platform PDP compliance assessment integrating both technical and governance indicators within a single framework. The findings indicate that governance practices remain the primary challenge in PDP implementation, providing practical recommendations for regulators and ESPs in strengthening personal data protection implementation in Indonesia.
Purpose: The objectives of this work are to conduct a comprehensive analysis of the primary influences of both natural and anthropogenic causes on flooding in Kalimantan, utilising a transparent machine learning approach, and to provide policymakers with valuable insights for formulating strategies to mitigate flooding. Methods: The research employs the Random Forest model, integrating SHAP (Shapley Additive Explanations), to examine non-linear multivariate correlations and ascertain the significance of each variable. The data consists of both natural elements (precipitation, elevation, slope gradient, and proximity to the river) and anthropogenic activity (land fire hotspots, planting, mining, and building new infrastructure). The data also includes public sentiment data in text form. We got these data points per year from 2021 to 2025. We used R-squared and SHAP scores to figure out how accurate the model was. Result: The model has a high R² score of 0.81, which shows that it can make accurate predictions. Using SHAP (SHapley Additive exPlanations), we can see that natural factors, such as how far away the river is and how much it rains, are what make the area vulnerable in the first place. Human actions, on the other hand, are what cause the floods to happen again and again. The indicator for hotspots of land burning is the most important factor. Plantation and mining operations make up more than 90% of the overall contributions, followed by other predictors, in the case of flood-induced deforestation. Novelty: The current study proposes a framework for explicable AI that integrates the use of random forests and SHAP to assess the significance of various flood risk indicators via quantitative analysis of geographical and public opinion data.
Purpose: This study evaluates the performance of a Virtual Network Computing (VNC)-based remote monitoring system within a Flight Information Display System (FIDS) environment. The research aims to identify infrastructure factors affecting monitoring performance and to develop a data-driven framework for evaluating monitoring reliability in distributed airport systems. Methods: Correlation analysis, Analysis of Variance (ANOVA), and multiple linear regression were applied to analyze the relationship between system resource utilization, network characteristics, and monitoring performance. The dataset consisted of 1,000 observations collected under various simulated monitoring conditions representing variations in latency, throughput, CPU utilization, and memory usage. Residual analysis and model evaluation were also performed to validate the statistical model. Result: The results showed that most infrastructure variables had very weak correlations (−0.02 to 0.05), indicating minimal multicollinearity. ANOVA testing revealed no statistically significant latency differences across low, medium, and high CPU load categories (F = 0.1625, p = 0.8500), with average latency remaining stable at approximately 52.82 ms. Regression evaluation demonstrated stable residual distribution and acceptable model consistency. The findings indicate that monitoring performance is influenced more by network conditions, particularly latency and throughput variability, than by computational load. Novelty: This study proposes an integrated analytical framework combining correlation analysis, ANOVA, and regression modeling to evaluate VNC-based monitoring performance in distributed systems. The framework provides a practical and reproducible approach for monitoring performance evaluation and infrastructure optimization in airport monitoring environments.
Purpose: This study develops a web-based insurance claim information system at PT Jasaraharja Putera within an integrated port environment. The conventional claim process, which relies on manual communication such as WhatsApp and email, often causes verification delays, document loss, lack of transparency, and risks of duplicate claims or fraud. The system aims to accelerate and simplify claim submission, particularly during reporting and initial verification by field officers, through real-time integration with the company’s core systems. Additionally, digital monitoring features enhance oversight of the claims process, reducing duplication and fraud potential. Methods: This research employed a descriptive qualitative method with an Agile Scrum system development approach. Data collection involved interviews with field officers and the claims team, analysis of existing business processes, system trials, and distribution of questionnaires to users. System development was conducted iteratively over several sprints, with each sprint producing features that were tested and evaluated based on user feedback. Results: Accelerate the claims input and verification process in the field. Increase transparency of claims status through a real-time monitoring dashboard. Reduce the risk of lost documents, human error, and potential duplicate claims. Achieve high user satisfaction levels based on questionnaire results at each development sprint. Novelty: This research introduces a real-time, integrated web-based claims system for port operations, applying Agile Scrum and monitoring features to prevent fraud. It offers practical solutions to public insurance claim inefficiencies and theoretical insights for advancing insurance information systems.
Purpose: The study focuses on the impact of Virtual Reality as a therapeutic technology in the context of Smart Healthcare on the rehabilitation of hand amputees and the ability of out-patients to gain computer access in a cost-effective manner. Methods: A qualitative, descriptive bibliometric survey of the literature was undertaken, utilizing peer-reviewed literature relevant to the integration of virtual medical care (V-Med) with the IoT, EHR, and CDSS. The literature examined was categorized into two analytical domains - physical and clinical opportunities and technical challenges. Results: The evidence suggests that cloud focused Virtual Reality (VR) helps people with Phantom limb pain and reduces the time required for amputees to achieve the adaptation to a prosthesis. The disadvantages of VR technology will only be realized when VR technology, EHR and the Internet of Everything (IoE) systems coalesce. This is where the central issue of technical feasibility lies. Novelty: The presentation here outlines a shift in attitude towards Virtual Reality. Rather than isolating it from a continuum encompassing all virtualization in the cloud with amputee rehabilitation modalities, we embrace it. We propose a framework that integrates cloud-based IoT wearables, EHRs, and CDSS into a Smart Healthcare ecosystem, thus outlining a feasible, affordable scalable virtual rehabilitation path.
Purpose: Epilepsy is a chronic neurological disorder that affects more than 50 million people worldwide, where early detection through EEG signal analysis is crucial for proper management. However, the quality of EEG signals is often affected by noise and artifacts, which can lead to diagnostic errors of up to 30% in the early stages. This study aims to develop an EEG signal preprocessing method to improve the classification performance of epileptic symptoms through preprocessing, segmentation, and seizure interval analysis approaches. Methods: The preprocessing stage involved applying a 50 Hz notch filter and a 0.5–60 Hz bandpass filter. The contribution of this work is in the development of hybrid segmentation based on frequency and amplitude analysis, while seizure intervals were identified using distances criteria between consecutive spikes detected on signals. The method was tested using the CHB-MIT dataset consisting of 23 EEG channels. Result: The results showed that the system successfully identified seizure segments with an average accuracy of 62.09%, and 9 out of 23 channels achieved accuracies above 70%. Channels Ch08 (86.60%), Ch09 (86.36%), and Ch19 (80.51%) achieved the highest accuracies. The results also showed high specificity(99.85%) and low False Positive rate(0.15%) indicating the system’s effectiveness to reduce falase positive. Novelty: This method proved effective in detecting epileptiform activity and shows potential as an EEG-based early detection tool for epilepsy, although further optimization is needed to improve accuracy on channels with low signal-to-noise ratio (SNR).
Purpose: Predicting rainfall using ARIMA, LSTM, and Hybrid ARIMA-LSTM models to obtain accuracy values on data at the Ahmad Yani Semarang station. Methods: This study implements the ARIMA, LSTM, and hybrid ARIMA-LSTM models to determine which of these models produces the most significant predictions using rainfall data at the Ahmad Yani Meteorological Station in Semarang. This method proves whether using the hybrid ARIMA-LSTM, which is a combination of the two models, is able to provide greater accuracy compared to the ARIMA/LSTM model. The results of these predictions can certainly help relevant stakeholders to improve rainfall accuracy, especially at the Ahmad Yani Meteorological Station. Result: By utilizing the power of statistical models (ARIMA) with deep learning (LSTM), the results of these two models provide higher accuracy compared to each model, as seen from the accuracy of the best ARIMA model using RMSE 15.8 and MAE 8.7, the best LSTM model RMSE 14.65 and MAE 9.06, while in the HYBRID ARIMA-LSTM model the best RMSE is 14.1 and MAE 9.06. Novelty: This research adds to the knowledge regarding the accuracy or combination of ARIMA and LSTM models which are rarely used, especially in the world of meteorology or rainfall. By utilizing the ARIMA model which is able to read linear patterns and the LSTM model which reads non-linear patterns, the accuracy of rainfall increases and can help related stakeholders.
Purpose: Sign language recognition systems based on 3D hand keypoints frequently experience generalization issues when trained on limited and homogeneous datasets, particularly under single-subject data collection settings. In BISINDO alphabet recognition, this limitation often leads to significant performance degradation when models are applied to unseen users or different acquisition devices. This study aims to improve cross-domain generalization of BISINDO alphabet recognition models by introducing realistic feature-level augmentation applied directly to 3D hand keypoints. Methods: A realistic 3D keypoint augmentation framework was proposed, consisting of Gaussian Jitter, Anisotropic Scale, Bone Length Scale, and Depth & Tilt Jitter to simulate sensor noise, anatomical variability, and viewpoint changes. Hand keypoints were extracted using MediaPipe Hands and classified using a multilayer perception (MLP). Model performance was evaluated through k-fold cross-validation on a single-subject internal dataset and cross-domain testing on an external dataset involving unseen subjects and different acquisition devices. Results: The experimental results indicate that the proposed augmentation strategy substantially improves generalization performance without degrading in-domain accuracy. The cross-domain F1-score increased from 67.20% in the baseline model to 89.55% after applying realistic 3D keypoint augmentation, while performance variability across validation folds was also reduced, indicating more stable learning behavior. Novelty: This work highlights that controlled geometric manipulation at the 3D keypoint level provides an effective and computationally efficient approach to mitigating overfitting in low-resource BISINDO recognition scenarios. By focusing on feature-level augmentation rather than image-based transformations or algorithm replacement, this study offers a practical strategy for enhancing robustness in real-world sign language recognition systems.
Purpose: This study aims to optimize the determination of the optimal number of clusters in the segmentation of handicraft-based Micro, Small, and Medium Enterprises (MSMEs) in Yogyakarta to support targeted and data-driven development strategies. Approach: A quantitative approach was applied to survey data collected from 145 MSMEs. The analytical pipeline consisted of four stages: (1) data acquisition through structured surveys, (2) preprocessing including encoding, mode imputation for missing values, and Min–Max normalization, (3) model development using the K-Means algorithm integrated with Particle Swarm Optimization (PSO) to automatically search for the optimal cluster number (K = 2–10), and (4) performance evaluation using Silhouette Score, Sum of Squared Error (SSE), and Mean Absolute Error (MAE). Result: The optimization process consistently converged to an optimal configuration of K = 8 clusters. Compared to standard K-Means, the proposed K-Means + PSO model reduced SSE from 54.555 to 51.676 and MAE from 0.124 to 0.116, indicating improved clustering stability and compactness. Semantic centroid analysis further revealed a hierarchical MSME structure consisting of Established Digital Adopters, Developing Potential Enterprises, and Subsistence Micro Enterprises, highlighting disparities in digital maturity and market reach. Novelty: This study contributes by integrating swarm-based optimization with centroid-driven semantic profiling, bridging algorithmic enhancement and policy-relevant interpretation. The proposed framework provides a robust and interpretable clustering model for MSME segmentation in emerging economic contexts.
Purpose: This research aims to design a mobile-based e-commerce system integrated with a payment gateway, as well as a centralized data management dashboard for A3+ Architect Design. The main focus of the research is on the efficiency of architectural design service business processes, including data management, information access, ordering mechanisms, payment integration, and real-time order monitoring for customers. Methods: The research methodology applied a systematic four-stage approach, including needs analysis to identify requirements through interviews with the A3+ studio, system model design, implementation, and testing. The system was built using client-server architecture with Flutter as the mobile application and Laravel as the web service and admin/designer dashboard. Functional validation of the system at the final stage was carried out using the black-box testing method. Result: Black box testing confirmed that all features are operating optimally and in accordance with the requirements specifications. The system successfully provides a comprehensive transaction workflow and an efficient operational dashboard for administrators and designers to support more organized data management. Novelty: The uniqueness of this research lies in the integration of a mobile application system with a web dashboard that is specifically tailored to the characteristics of the architectural services industry, which was previously dominated by manual processes. This research contributes to the digitization of customer services and increased transaction transparency, which directly impacts the strengthening of customer trust and the operational efficiency of service providers.
Purpose: Feature selection is one critical element of healthcare data classification, which directly affects predictive performance, model robustness, and interpretability. Nevertheless, traditional model-based feature importance methods are unstable in robustness and provide random or misleading results on high-dimensional and heterogeneous healthcare data. The purpose of this paper was to assess model-based and SHapley Additive exPlanations (SHAP) - based feature analysis on multimodal healthcare data classification in a comparative manner. Methods: This study employed a quantitative comparative experimental design using real Electronic Medical Records (EMR) data captured from primary care clinics. The dataset comprises 2,158 patient records, with numerical and textual features in a multimodal feature space. The analytical pipeline included data acquisition, preprocessing, multimodal feature integration, and model development using six supervised learning algorithms from ensemble-based and margin-based categories. The model-based feature importance was used with the SHAP-based feature importance. We examined the robustness of this method across different scenarios through systematic feature ablation using baseline, strong, and weak features. Result: Experimental results demonstrate that model-based feature importance exhibits unpredictable behavior when features are removed and is sometimes counterintuitive. On the other hand, feature importance based on SHAP is consistent and linearly proportional across the models under consideration. The highest macro F1-score was achieved by Extra Trees with SHAP-based strong features (0.824), exceeding the baseline (0.811). In robustness testing, SHAP-based weak-feature removal reduced the LightGBM F1-score from 0.764 to 0.575, suggesting a clearer distinction between informative and non-informative features. Overall, SHAP-based feature selection provided a more reliable and interpretable framework for multimodal healthcare classification. Novelty: Our experiment results demonstrate empirically that the SHAP-based feature importance is more robust and reliable than conventional model-based approaches when it comes to extracting features from multimodal medical records. This work shows that SHAP is more than a post hoc explanation by presenting it as an interpretable feature selection criterion guiding feature relevance analysis in healthcare machine learning.
Purpose: Defect detection in industrial bolts is crucial for ensuring product reliability, production safety, and consistent quality control in modern industrial environments. However, visual inspection of metal bolts remains challenging due to low contrast, uneven lighting, and reflective surfaces that often hide subtle defect patterns and reduce detection accuracy. Most existing YOLO-based approaches focus on architectural modifications to improve performance, which may increase model complexity and limit real-time applicability. Methods: This study integrates Contrast-Limited Adaptive Histogram Equalization (CLAHE) with YOLOv8 to improve defect visibility prior to detection. CLAHE enhances local contrast by redistributing pixel intensities while suppressing noise amplification, thereby strengthening feature representation for deep learning-based detection. Experiments were conducted on a publicly available industrial bolt dataset annotated via Roboflow, using a 3-fold cross-validation strategy. Performance was assessed with Precision, Recall, mAP@50, mAP@50–95, FPS, and FLOPs to evaluate accuracy and real-time feasibility. Result: Experimental results based on a 3-fold cross-validation scheme indicate that the proposed CLAHE–YOLOv8 model achieves consistent performance improvements over the baseline YOLOv8 configuration. The method obtains an average Precision of 0.9495±0.0068, Recall of 0.9028±0.0235, mAP@50 of 0.9364±0.0156, and mAP@50–95 of 0.7121±0.0037, while maintaining real-time inference performance at 29.79 FPS. These results demonstrate that contrast-based preprocessing contributes positively to detection stability and localization consistency without increasing model complexity. Novelty: The novelty of this research lies in demonstrating that data-level contrast enhancement using CLAHE effectively improve industrial bolt defect detection performance without architectural modification, offering a practical and computationally efficient solution for real-time industrial inspection systems.
