Journal of Information Systems Engineering and Business Intelligence

Predicting the Volatility of Jakarta Composite Index Using GARCH and LSTM with Volume-Up Strategy Approach

I Made Adhi Dharmaningrat — 2025-10-28

Background: Stock market volatility forecasting is essential for financial decision-making, although the complexity presented significant challenges. This prompted previous studies to identify correlations between the volatilities of international stock indices and Jakarta Composite Index (JKSE), describing the potential of hybrid econometric and deep learning models in the prediction process.

Objective: This study aims to develop an optimized hybrid model for forecasting the volatility of JKSE by integrating Generalized Autoregressive Conditional Heteroskedasticity (GARCH), Long Short-Term Memory (LSTM), and Volume-Up (VU) strategy, in the context of an emerging market recovering from economic disruptions.

Methods: Historical daily data from five major stock indices, namely JKSE, DJI, SPX, N225, and HSI, covering the period from January 1, 2000, to December 31, 2023, were used to formulate eleven datasets. Furthermore, a hybrid model was developed and evaluated by combining GARCH, LSTM, and VU strategy for conditional volatility estimation, sequential prediction, and data transformation, respectively. Hyperparameter tuning was performed to determine the best activation functions, batch sizes, and timesteps. Based on this perspective, Mean Squared Error (MSE) was used to assess predictive accuracy.

Results: GARCH-LSTM exhibited superior performance over a standalone LSTM model, improving RMSE by 11.43%. The incorporation of VU strategy further enhanced accuracy, with an optimal setting (α = 0.5) leading to a total RMSE improvement of 17.35%. The best hyperparameters included SELU + tanh activation function and a batch size of 128 or 256. Meanwhile, a timestep of 1 provided the best predictive performance, depicting the importance of recent market movements in forecasting.

Conclusion: In conclusion, this study proved the effectiveness of hybrid models in predicting stock market volatility in emerging markets. The results outlined the advantage of integrating econometric and deep learning approaches, with VU strategy playing a significant role in refining predictions.

Keywords: GARCH, LSTM, Volatility Prediction, Volume-Up Strategy, Emerging Markets, Economic Recovery

Evaluating the Effectiveness of Mobile Precision Push Services A User-Centric Behavioral Framework

Peng Wang — 2025-10-28

Background: As precision push services (PPS) become increasingly embedded in mobile communication ecosystems, understanding how users perceive and respond to these services has become a pressing research concern. While prior studies have focused on technical accuracy and personalization algorithms, limited attention has been paid to how experiential and perceptual factors collectively influence user engagement and behavioral outcomes.

Objective: This study aims to construct and empirically validate a comprehensive user-centered evaluation model for precision push services. It seeks to identify which experiential dimensions most significantly influence user perception, and how these perceptions translate into behavioral responses.

Methods: A conceptual framework was developed integrating five experiential predictors—message validity and quality, non-interference with user experience, operability, user choice, and information transparency—alongside two perceptual mediators (effect and impact) and one behavioral outcome. A structured questionnaire using an eight-dimensional Likert scale was administered to 279 university students across multiple institutions. Data analysis involved reliability and validity testing, correlation analysis, ANOVA, and multiple regression to examine causal relationships and demographic influences.

Results: The results indicate that user choice is the most influential factor affecting both perceived effect and impact of PPS. Information transparency and message quality also significantly predict perceptual outcomes, while non-interference showed strong correlations but no direct causal influence. The impact of push services emerged as a stronger determinant of user behavior than perceived effectiveness. Gender and geographic differences were statistically controlled and found to have minimal effect on the primary causal pathways.

Conclusion: The study highlights the importance of user autonomy, transparency, and meaningful content delivery in designing effective PPS systems. By validating a full causal model and identifying critical user-centered variables, the research provides actionable insights for improving user engagement, trust, and behavioral response in personalized mobile push environments.

Keywords: Precision Push Service, User Perception, Causal Modeling, Personalization, Personalization

Digital Transformation of Islamic Endowments (Waqf): What Appeals to Generation Z in e-Cash Waqf?

Clarashinta Canggih — 2025-10-28

Background: Cash waqf in Indonesia is optimized through the use of digital media to improve access, transparency, and public participation, particularly among the tech-savvy younger generation. This led to the formulation of effective strategies, which enabled the understanding of factors influencing digital waqf intention, including gender-based differences.

Objective: This present study aims to explore gender differences in respect to the determinants of intention towards participating in digital cash waqf. This was realized by comparing responses between male and female Generation Z individuals.

Methods: This quantitative study adopted purposive sampling method to collect data. Subsequently, a total of 645 respondent data were processed using Partial Least Square Structural Equation Model (PLS-SEM) method with the assistance of SmartPLS 4.0 software.

Results: The male and female respondents stated that cash waqf literacy did not influence trust and behavioral intention. However, perceived ease of e-cash waqf significantly impacted both trust and behavioral intention. Majority of the male respondents reported that religiosity, and trust in nazhir had a significant impact. Both genders stated that religiosity did not moderate the relationship between the variables.

Conclusion: In conclusion, the importance of technological ease of use and religiosity in influencing trust and intention to contribute to digital cash waqf was analyzed. Based on this perspective, both variables impacted trust and behavioral intention. The female respondents perceived trust as an insignificant factor, and recommended nazhir institutions partnered with financial technology (fintech) companies to develop user-friendly platforms. This included the engagement of female donors through religious education. The numerous campaigns should focus on technological literacy and the religious value of digital waqf contributions.

Keywords: E-cash waqf, Generation Z, Multi Group Analysis, Male, Female

CBTi-YOLOv5: Improved YOLOv5 with CBAM, Transformer, and BiFPN for Real-Time Safety Helmet Detection

Tio Dharmawan — 2025-10-28

Background: Some construction workers are often in a situation where injuries can occur from negligence in the use of safety helmets. To avoid this, supervision of the use of safety helmets should be conducted continuously during the work process through the application of computer vision technology. However, the complex background of the construction environment is a challenge to detecting small and densely packed safety helmets accurately.

Objective: The construction environment is complex, and the wide workspace allows workers to be in an area far from supervision. The process makes it difficult for models to detect the use of safety helmets in complex, wide, and very high object density construction environments. Therefore, this study aims to overcome the problem by modifying YOLOv5s (You Only Look Once version 5) architecture.

Methods: Real-time monitoring of the use of safety helmets could be performed using YOLOv5. This study proposed a modified YOLOv5s model called CBTi-YOLOv5s. The model incorporated Convolutional Block Attention Module (CBAM), Transformer, and Bi-directional Feature Pyramid Network (BiFPN) to improve feature extraction, multi-scale object representation, as well as detection accuracy, specifically on small and high-density objects in complex construction environments.

Results: The results showed the modified YOLOv5s architecture had made an improvement of 3.7% in mean average precision (mAP) compared to the base YOLOv5s model. mAP of the base YOLOv5s model was 93.6%, while the modified CBTi-YOLOv5s model achieved 97.3%. The proposed modified YOLOv5s model also achieved an inference speed of 58 frames per second (FPS), and the base model achieved 104 FPS.

Conclusion: CBTi-YOLOv5s improved the accuracy, mAP, and ability to detect objects of varying scales. However, this improvement had drawbacks, namely increased model size and decreased inferential speed due to increased model architectural complexity..

Keywords: Bi-FPN, CBAM, CBTi-YOLOv5s, Helmet Detection, Transformer, YOLOv5

The Moderating Role of Seamless User Experience in Omnichannel Marketing and Customer Retention: A Technology Acceptance Model-Based Study in Emerging Markets

Derrick Nukunu Akude — 2025-10-28

Background: In emerging markets, e-commerce firms increasingly adopt omnichannel marketing to enhance customer retention. A seamless user experience across digital and physical channels strengthens brand loyalty, yet implementation challenges remain significant due to infrastructural and connectivity constraints.

Objective: This paper investigated the moderating role of seamless user experience on the relationship between omnichannel marketing and customer retention among E-commerce users in Ghana.

Methods: The survey utilized a self-administered questionnaire approach, gathering a total of 384 completed responses for data analysis utilizing Smart PLS-SEM (version 4).

Results: The study noted the occurrence of a positively significant relationship between cross-channel customer experience and customer retention. Secondly, there is a positively significant effect between channel service configuration and customer retention. However, the connection between channel integration quality and customer retention is insignificant. Moreover, the relationship between omnichannel personalization and customer retention is insignificant. Furthermore, seamless user experience has a positively significant moderation role in the connection between omni-channel personalization and customer retention. In addition, seamless user experience has a negatively significant moderation role in the relationship between channel integration quality and customer retention. However, seamless user experience has a positively insignificant moderation role in the relationship between cross-channel customer experience and customer retention. Also, seamless user experience has a positively insignificant moderation role in the relationship between channel service configuration and customer retention

Conclusion: This investigation provides insights into the value of integrating seamless user experience to strengthen the relationship between omnichannel marketing as well as customer retention thereby highlighting their implications for theory, managers and business success.

Keywords: Channel Integration Quality, Customer Retention, Seamless User Experience, Channel Service Integration, Cross-channel Customer Experience.

Academic Guidebook Chatbot: Performance Comparison of Fine-Tuned Mistral 7B and LlaMA-2 7B

Davied Indra Rachman — 2025-10-28

Background: Chatbot is recently ranked as the main technological solution due to the high demand for fast and efficient information retrieval. Therefore, this study was carried out to develop a local document-based chatbot that can answer questions related to the contents of PDF documents using open-source AI models such as Mistral 7B and LLaMA-2 7B. Although these models were effective at processing natural language, a major challenge was observed in the tendency to generate hallucinated answers, characterized by having inaccuracies and being out of context.

Objective: This study aims to reduce hallucinatory responses from chatbot models by making their responses more precise and accurate through fine-tuning. The performance of fine-tuned models (Mistral 7B and LLaMA-2 7B) was also compared.

Methods: Fine-tuning of the two models was performed using domain-specific datasets taken from Academic Guidebook. This process was conducted to improve models ability to understand and answer questions relevant to Academic Guidebook context. Performance was evaluated using METEOR Score to measure literal agreement and BERTScore to assess meaning agreement. In addition, response time was measured to assess efficiency, while chatbot system was developed using Streamlit and LangChain for real-time interaction.

Results: Fine-tuned Mistral 7B model achieved the highest METEOR value of 0.40 and F1 of 0.78 based on BERTScore results. Regarding efficiency, fine-tuned Mistral 7B showed a faster response time than LLaMA-2. Meanwhile, the non-fine-tuned Mistral 7B and LLaMA-2 7B showed a longer response time than fine-tuned Mistral 7B and LLaMA-2 7B.

Conclusion: The results showed that the enhancements significantly improved the performance of large language models in specific tasks, reduced hallucinations, and enhanced response quality

Keywords: Chatbot, Large Language Model, Mistral 7B, LLaMA-2 7B, METEOR Score

Hybrid Dual-Stream Deep Learning Approach for Real-Time Kannada Sign Language Recognition in Assistive Healthcare

Gurusiddappa Hugar — 2025-10-28

Background: Recent advances in sign language recognition (SLR) focus on high-resource languages (e.g., ASL), leaving low-resource languages like Kannada Sign Language (KSL) underserved. Edge-compatible, real-time SLR systems for healthcare remain scarce, with most existing methods (CNN-LSTM, 3D ResNet) failing to balance accuracy and latency for dynamic gestures.

Objective: This research work aims to develop a real-time, edge-deployable KSL recognition system for assistive healthcare, addressing gaps in low-resource language processing and spatio-temporal modeling of regional gestures.

Methods: We propose a hybrid dual-stream deep learning architecture combining EfficientNetB0 for spatial feature extraction from RGB frames. A lightweight Transformer with pose-aware attention to model 3D hand keypoints (MediaPipe-derived roll/pitch/yaw angles). We curated a new KSL medical dataset (1,080 videos of 10 critical healthcare gestures) and trained the model using transfer learning. Performance was evaluated quantitatively (accuracy, latency) against baselines (CNN-LSTM, 3D ResNet) and in real-world tests.

Results: The system achieved 97.6% training accuracy and 96.7% validation accuracy, 81% real-world test accuracy (unseen users/lighting conditions). 53ms latency on edge devices (TensorFlow.js, 1.2GB RAM), outperforming baselines by ≥12% accuracy at similar latency. The two-stage output pipeline (Kannada text + synthetic speech) demonstrated 98.2% speech synthesis accuracy (Google TTS API).

Conclusion: Our architecture successfully bridges low-resource SLR and edge AI, proving feasible for healthcare deployment. Limitations include sensitivity to rapid hand rotations and dialect variations.

Keywords: Assistive Healthcare, Edge AI, Kannada Sign Language, Low-resource Language, Real-time Recognition, Transformer.

Generating User Personas for Eliciting Requirements Using Online News Data

Halim Wildan Awalurahman — 2025-10-28

Background: In software development, creating user personas remains challenging despite their recognized value. Cost, adaptability, and data scarcity present obstacles in designing these critical personas. A new perspective and process innovation for generating user personas is essential to overcome this hurdle.

Objective: This study introduces a method for extracting user persona attributes, including names, occupations, workplaces, and goals.

Methods: A framework for extracting user persona information from online news sources is created. Our method employs a comprehensive SpaCy processing pipeline, incorporating NER, SpaCy rule-based matching, and phrase matching.

Results: The evaluation results showcase promising performance metrics, with an average recall value of 0.700, precision of 0.402, and F1-score of 0.506.

Conclusion: This study demonstrates the feasibility of extracting user persona attributes from online news data. Future research could focus on enhancing the method’s performance, investigating its effectiveness in creating relationships, and ensuring that the generated user personas accurately reflect the news text data.

Keywords: Process innovation, Natural Language Processing, Online News, Software Development, User Persona

Transfer Learning-Based Convolutional Neural Network for Accurate Detection of Rice Leaf Disease in Precision Agriculture

Bety Wulan Sari — 2025-10-28

Background: Traditional approaches to rice disease identification depend mainly upon visual examination, which is quite labor-intensive and generally demands a certain skill level from people engaged in this activity. However, these approaches suffer from high time costs and potential errors and are impractical for large-scale daily monitoring. The recent rise of deep learning has offered opportunities for automated detection process improvement, which needs to be fast-accurate as good farmer-centric.

Objective: This study aims to enhance the accuracy of image rice leaf disease classification via feature extraction for rice leaf disease in four instances of pre-trained CNN models and provide an automated solution for early detection ahead of timely care by obtaining insights into crop production through precision agriculture.

Methods: This study combined transfer learning with four pre-trained CNN models - InceptionResNetV2, MobileNetV2, DenseNet121, and VGG16.

Results: The outcome of this research enables the identification of the optimal model to relate datasets where DenseNet121 achieved the highest accuracy, i.e. 99.10%, followed by MobileNetV2, having a precision of 97.10%.

Conclusion: The new framework results in a highly accurate and high-throughput early disease detection element in precision agriculture, better than state-of-the-art approaches based on traditional techniques.

Keywords: Deep Learning, DenseNet121, Image Classification, Rice Leaf Diseases, Transfer Learning

An Empirical Study of In-App Purchase Intention Behavior of Generation Z in Mobile Games

Donny Putratama — 2025-10-28

Background: The rapid evolution of information technology has significantly transformed digital transactions and consumer behavior. Although in-game transactions and the mobile gaming industry are expected to experience significant growth, Generation Z gamers’ purchasing behavior remains underexplored.

Objective: This study aims to investigate the factors influencing Gen-Z’s intention to make in-app purchase of virtual goods within mobile games.

Methods: Partial Least Square (SmartPLS) analysis was conducted to examine whether live streamers, co-branding, good price, and mobile game loyalty affected in-app purchase intention among Gen-Z gamers.

Results: The results showed that live streamers, co-branding, and good price positively influenced gamers’ desire to purchase in-game items. Mobile game loyalty was also found to have the strongest influence on in-app purchase intention.

Conclusion: This study emphasized how game influencers, co-branding, fair pricing, and player loyalty influenced in-app purchase intentions among Indonesian Gen-Z mobile gamers. The findings revealed that using live streamers to showcase game characters, building stronger interactions with players, and offering sales promotions are effective ways to promote more in-app purchases.

Keywords: Co-Branding, Good Price, In-App Purchase Intention, Live Streamers, Mobile Game Loyalty.

Optimizing Tuition Fee Determination with K-Means Cluster Relabeling Based on Centroid Mapping of Principal Component Pattern

Wiyli Yustanti — 2025-10-28

Background: Tuition fee in Indonesian public universities is determined based on the socioeconomic status of prospective students. In this context, students are assigned to tuition fee groups after passing the selection process through achievement-based or computer-based exams. However, the current grouping system shows overlapping distributions, indicating the need for a more precise classification method.

Objective: This research aims to improve the accuracy of tuition fee group assignments by refining the clustering structure and relabeling the classification dataset.

Methods: A total of 13 socioeconomic variables were used to predict tuition fee groups. This research used K-Means clustering algorithm and a relabeling process using centroid mapping of principal components to balance original and newly generated labels. To assess the effectiveness of the relabeling process, six classification algorithms, namely Decision Tree (DT), K-Nearest Neighbors (KNN), Naive Bayes (NB), Logistic Regression (LR), Random Forest (RF), and Support Vector Machine (SVM), were used. Statistical tests at a 5% significance level were conducted to evaluate improvements in classification accuracy.

Results: The relabeling process significantly enhanced prediction accuracy compared to the original dataset. The refined clustering structure reported better classification performance across all six algorithms, showing the effectiveness of the proposed method.

Conclusion: The results showed that robust clustering and a relabeling method improved the precision of tuition fee classification systems. The proposed framework provided a data-driven solution for refining classification models, ensuring a fairer distribution of tuition fee based on socioeconomic indicators. The novelty lies in the centroid-based relabeling, which uses principal component patterns to enhance interpretability and classification accuracy. The method was adaptable for global use in any educational system using socioeconomic-based fee classification. Future research should explore alternative clustering methods and additional socioeconomic factors to enhance classification accuracy.

Keywords: K-Means Clustering, Machine Learning, Relabeling Process, Socioeconomic Indicators, Tuition Fee Classification

Clustering and Mixture Modeling of Schooling Expectancy Trends in Papua Province: A Spatial Analysis Using the Mapping Toolbox

Jonathan Wororomi — 2025-10-28

Background: Persistent educational inequality in Papua Province, particularly in remote highland districts, is driven by limited infrastructure and accessibility. Although Schooling Expectancy (Harapan Lama Sekolah, HLS) is widely recognized as a forward-looking educational metric, existing studies rarely incorporate probabilistic modeling with spatial analysis to examine regional disparities.

ObjectiveThis study aimed to identify spatial and statistical patterns of schooling expectancy across 29 districts in Papua from 2010 to 2023 by combining probabilistic clustering with spatial visualization methods.

Methods: The analysis applied Gaussian Mixture Model (GMM) clustering, which was validated using the Silhouette Index and Davies–Bouldin Index (DBI), to group districts based on HLS trends. Fourteen candidate probability distributions were evaluated using Kolmogorov–Smirnov and Anderson–Darling tests. In addition, five model selection criteria (AIC, BIC, AICc, CAIC, HQC) were applied to refine the fit. Cluster-wise mixture model was constructed, and spatial interpretation was improved through MATLAB’s Mapping Toolbox as well as wind rose diagrams.

Results: During the process of the analysis, four statistically distinct clusters were identified. Cluster 3 (coastal districts) showed the highest and most stable HLS (12.1–14.0 years), while Cluster 4 (remote highlands) signified the lowest (2.4–5.6 years) with high dispersion. Right-skewed distributions (e.g., Weibull, Gamma) modeled high-performing districts, and heavy-tailed, left-skewed ones (e.g., Stable, Inverse Gaussian) modeled marginalized regions. Spatial visualization confirmed a clear coastal–highland divide in educational attainment.

Conclusion: The proposed incorporation of probabilistic modeling and spatial clustering offered a robust analytical tool for capturing intra-regional educational disparities. This framework provided empirical evidence to support geographically differentiated policy interventions in Papua and could be adapted to similar underserved regions in future studies.

Keywords: Schooling Expectancy, Gaussian Mixture Model, Probabilistic Modeling, Silhouette Index, Davies–Bouldin Index, Spatial Clustering, Education Inequality, Papua Province.

Adaptive Multi‑Layer Framework for Detecting and Mitigating Prompt Injection Attacks in Large Language Models

Raden Budiarto Hadiprakoso — 2025-10-28

Background: Prompt injection attacks are methods that exploit the instruction‐following nature of fine‐tuned large language models (LLMs), leading to the execution of unintended or malicious commands. This vulnerability shows the limitation of traditional defenses, including static filters, keyword blocks, and multi‐LLMs cross‐checks, which lack semantic understanding or incur high latency and operational overhead.

Objective: This study aimed to develop and evaluate a lightweight adaptive framework capable of detecting and neutralizing prompt injection attacks in real-time.

Methods: Prompt-Shield Framework (PSF) was developed around a locally hosted Llama 3.2 API. This PSF integrated three modules, namely Context-Aware Parsing (CAP), Output Validation (OV), and Self-Feedback Loop (SFL), to pre-filter inputs, validate outputs, and iteratively refine detection rules. Subsequently, five scenarios were tested, comprising baseline (without any defenses), CAP only, OV only, CAP+OV, and CAP+OV+SFL. The evaluation was performed over a near-balanced dataset of 1,405 adversarial and 1,500 benign prompt, measuring classification performance through confusion matrices, precision, recall, and accuracy.

Results: The results showed that baseline achieved 63.06% accuracy (precision = 0.678; recall = 0.450), while OV only improved performance to 79.28% (precision = 0.796; recall = 0.768). CAP reached 84.68% accuracy (precision = 0.891; recall = 0.779), while CAP+OV yielded 95.25% accuracy (precision = 0.938; recall = 0.966). Finally, integrating SFL over 10 epochs further improved performance to 97.83% accuracy (precision = 0.980; recall = 0.975) and reduced the false-negative count from 48 (CAP+OV) to 35 (CAP+OV+SFL).

Conclusion: The results show the significance of using multiple defenses, such as contextual understanding, OV, and adaptive learning fusion, which are needed for efficient prompt injection mitigation. This shows that PSF framework is an effective solution to protect LLMs against advancing threats. Moreover, further studies should aim to refine the adaptive thresholds in CAP and OV, particularly in multilingual or highly specialized environments, and examine other forms of SFL solutions for better efficiency.

Keywords: Prompt Injection, LLMs Security, Jailbreak, Natural Language Processing

Exploring Service Quality and Consumer Acceptance of Autonomous Convenience Stores

Chin Fei Goh — 2025-10-28

Background: Automation is revolutionizing retail operations, leading consumers to increasingly interact with advanced retail technologies. While there have been studies on the influence of service quality on consumer acceptance, research examining the service quality of hybrid services and consumer acceptance in automated retail is limited.

Objective: This study aims to examine consumer acceptance of automated retail stores.

Methods: This study tested a proposed model by surveying 101 consumers and using a questionnaire for hypothesis testing. Data were analyzed using Partial Least Squares Structural Equation Modeling (PLS-SEM) to explore the effect of e-service quality dimensions on technology acceptance (perceived ease of use, perceived usefulness, and behavior intention) in the context of unmanned automated retail stores.

Results: The findings reveal that information quality positively affects perceived ease of use, while system quality positively affects perceived usefulness.

Conclusion: This study generates new insights by incorporating e-service quality dimensions from the E-Service Quality model into the Technology Acceptance Model. Additionally, the results highlight the growing importance of seamless digital experiences and reliable systems in shaping user perceptions and behavioral intentions. These findings offer practical implications for retailers aiming to enhance customer satisfaction and adoption of unmanned retail technologies through improved service design and digital infrastructure. Future research can further explore other influencing factors such as trust, perceived risk, and user demographics to better understand the evolving dynamics of consumer-technology interaction in automated retail environments.

Keywords: artificial intelligence; autonomous convenience store; consumer acceptance; e-service quality; technology acceptance model

Diabetic Retinopathy Fundus Image Classification Using Self-Organizing Maps

Yulius Denny Prabowo — 2025-10-28

Background: Diabetic retinopathy (DR) is a condition that impairs the blood vessels in the retina, resulting in vision loss ranging from temporary to permanent blindness. It commonly affects individuals diagnosed with diabetes mellitus (DM). Fundoscopy is a technique used to identify DR by examining the fundus of the eye during an eye examination. This process is time-consuming and can be expensive.

Objective: This study aimed to examine the identification of DR using digital image processing methods.

Methods: The self-organizing map (SOM) artificial neural network was employed. Diabetic retinopathy will be categorized according to its severity, including normal, mild, moderate, or severe. This classification considers the quantity of exudates and microaneurysms and the blood vessel structure in the fundus image. The dataset used in this investigation comprised 1000 fundus images acquired from the MESSIDOR ophthalmology database.

Results: The findings indicate that the SOM approach achieves a training accuracy of 72% and a testing accuracy of 62%.

Conclusion: The DR severity classification system can effectively extract DR-related features by segmenting exudates, blood vessels, and microaneurysms from funduscopic images during training, testing, and evaluation.

Keywords: Diabetic Retinopathy, Self-Organizing Map, Fundus Image Classification, Digital Image Processing