Claim Missing Document
Check
Articles

Found 17 Documents
Search

Transformer-based Models for Cardiovascular Disease Predictions from Electronic Health Records: A Systematic Review Chikumo, Onayi Theresa; Ndlovu, Belinda
Journal of Applied Informatics and Computing Vol. 10 No. 1 (2026): February 2026
Publisher : Politeknik Negeri Batam

Show Abstract | Download Original | Original Source | Check in Google Scholar | DOI: 10.30871/jaic.v10i1.11899

Abstract

This systematic literature review (SLR) analyses 16 studies published between 2020 and 2025 that applied transformer-based or other machine learning models to predict cardiovascular disease (CVD) using electronic health records (EHRs). Following the Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) guidelines, the review ensures transparency in the identification, screening, and quality appraisal of eligible studies. The key findings reveal a rapid shift from traditional machine learning models, such as Random Forest, toward transformer architectures like the Bidirectional Encoder Representation from Transformers for Electronic Health Record (BEHRT) and its variants. These models demonstrate a superior discrimination (Area Under Curve:0.84 to 0.93) due to their capacity to model long-term temporal dependencies. Explainable AI (XAI) tools, such as attention visualisation, were frequently employed, yet clinical interpretability and integration into decision support remain underexplored. The review also highlights opportunities in federated and privacy-preserving learning, multimodal data fusion, and hybrid architectures that integrate transformers with traditional machine learning methods. This review addresses a gap in the past literature by being the first SLR to compare transformer variants for the prediction of CVDs. Other SLRs examined general CVD risk models, but the present SLR analyses interpretability, external validation and methodological limitations to transformer models. The findings of the recent SLR reported challenges that include data-shift limitations, model-poor population generalisation and their limitations to clinical adoption, which highlights the need for more evaluation protocols and clinicians’ interpretability frameworks.
Multi-Agent Retrieval Augmented Generation for Clinical Decision Support: A Systematic Review and Integrative Conceptual Framework Mugambiwa , Tarisai; Ndlovu, Belinda
Journal of Applied Informatics and Computing Vol. 10 No. 1 (2026): February 2026
Publisher : Politeknik Negeri Batam

Show Abstract | Download Original | Original Source | Check in Google Scholar | DOI: 10.30871/jaic.v10i1.11900

Abstract

Multi agent retrieval augmented generation (RAG) systems are increasingly explored as advanced architectures for clinical decision support combining information retrieval, reasoning and verification through coordinated agent interactions. This study systematically reviews applications of agentic and multi agent RAG in clinical decision support systems (CDSS) and synthesizes an integrative conceptual framework linking technical design to technology adoption considerations. Following PRISMA guidelines, searches were conducted from PubMed, IEEE Xplore and ScienceDirect using structured Boolean strings combining terms for multi agent architectures, RAG and CDSS.The search yielded 12 studies published between 2020 and 2025 that met the inclusion criteria. The review synthesises evidence on multi agent role configurations retrieval and reasoning strategies, verification mechanisms and reported clinical contexts. Across studies, dominant challenges include data and corpus limitations retrieval quality dependency, limited clinical validation and computational overhead, alongside governance concerns such as privacy, bias and accountability. Building on the synthesis, we propose a four-agent CDSS framework retriever, reasoner, verifier, safety and map its deployment determinants to Technology Acceptance Model constructs perceived usefulness, perceived ease of use, trust and diffusion of Innovations attributes. The review concludes with design-oriented recommendations for safer, explainable, and adoption-ready multi-agent RAG CDSS, particularly for low-resource contexts.
A Personalized Generative AI Model for Diabetes Drug Discovery: Integrating Molecular and Clinical Data Using Variational Autoencoders (VAE) Ndlovu, Belinda
The Indonesian Journal of Computer Science Vol. 15 No. 1 (2026): The Indonesian Journal of Computer Science
Publisher : AI Society & STMIK Indonesia

Show Abstract | Download Original | Original Source | Check in Google Scholar | DOI: 10.33022/ijcs.v15i1.5084

Abstract

Diabetes drug discovery remains slow, costly, and insufficiently personalised, particularly in resource-constrained healthcare settings. This study proposes and empirically evaluates a personalised, generative artificial intelligence framework that integrates molecular and clinical data to generate diabetes drug candidates. Guided by the CRISP-DM framework, a hybrid Clinical–Molecular Variational Autoencoder (VAE) architecture was developed, combining molecular representations with anonymised patient metabolic profiles, including HbA1C, fasting glucose, BMI, cholesterol, and age. Molecular data were sourced from PubChem and ChEMBL, and generated compounds were evaluated using drug-likeness metrics, molecular validity checks, and downstream effectiveness classification. The model successfully generated chemically valid, drug-like molecules with average Quantitative Estimate of Drug-likeness (QED) scores above 0.5. At a fixed decision threshold, effectiveness classification achieved an accuracy of 0.80; however, probability calibration analysis revealed limited discriminative reliability across thresholds (AUC = 0.49), highlighting the impact of class imbalance. Unlike prior molecule-centric generative drug discovery approaches, this study presents one of the first empirically evaluated Clinical–Molecular dual-VAE frameworks for personalised diabetes drug discovery, explicitly integrating patient metabolic profiles while revealing calibration limitations in generative pharmaceutical pipelines.
Developing a graph-based machine learning model for identifying money laundering networks associated with sanctioned entities in a bank in Zimbabwe Ndlovu, Belinda; Kiwa, Fungai Jacqueline; Muduva, Martin; Chipfumbu, Colletor T.; Marambi, Sheltar; Maphosa, Amazing
Indonesian Journal of Innovation and Applied Sciences (IJIAS) Vol. 6 No. 1 (2026): February-May
Publisher : CV. Literasi Indonesia

Show Abstract | Download Original | Original Source | Check in Google Scholar | DOI: 10.47540/ijias.v6i1.2306

Abstract

Money laundering networks associated with sanctioned entities pose a significant risk to financial systems, often operating through complex relational transaction structures that evade traditional rule-based monitoring. While graph neural networks have demonstrated promise in financial crime detection, limited work has formally modelled sanction-linked transaction networks within highly imbalanced banking datasets under consistent comparative evaluation. This study proposes a directed weighted graph-based learning framework for identifying sanction-associated money laundering networks using real-world banking transaction data. Transactions were modelled as relational graphs, with accounts as nodes and transfers as weighted edges, and evaluated using a Graph Convolutional Network (GCN) against classical and ensemble classifiers. The proposed model achieved an accuracy of 88.18%, F1-score of 0.7345, ROC-AUC of 0.8968, and a superior Matthews Correlation Coefficient compared to baseline methods. Results demonstrate that relational graph modelling improves the detection of structurally coordinated laundering behaviours that are not captured by independent transaction classifiers. These findings support the integration of graph neural network architectures into anti-money laundering systems to enhance sanction-linked detection capabilities in complex financial networks.
A Digital Twin–Driven Machine Learning Framework for Diabetes Risk Prediction and Short-Term Health Trajectory Simulation Ndlovu, Belinda
The Indonesian Journal of Computer Science Vol. 15 No. 2 (2026): The Indonesian Journal of Computer Science
Publisher : AI Society & STMIK Indonesia

Show Abstract | Download Original | Original Source | Check in Google Scholar | DOI: 10.33022/ijcs.v15i2.5110

Abstract

Diabetes remains a major global health challenge, requiring early risk detection and proactive management to reduce long-term complications. However, existing approaches are predominantly reactive and rely on static clinical indicators, limiting their ability to support personalized and forward-looking care. This study proposes an integrated framework that combines machine learning (ML) and digital twin (DT) technologies to enable both diabetes risk prediction and short-term health trajectory simulation. Using the CDC Diabetes Health Indicators dataset, a structured CRISP-DM methodology was applied to guide data preprocessing, feature selection, model development, and evaluation. Class imbalance (13.9% minority class) was addressed using the Synthetic Minority Over-sampling Technique (SMOTE). Five machine learning models were evaluated, with Gradient Boosting achieving the best performance (ROC-AUC = 0.797; F1-score = 0.415), indicating acceptable discriminative capability under imbalanced conditions. Building on this predictive layer, a digital twin framework was developed to simulate individual risk trajectories over a 90-day period. The system was operationalized through a web-based architecture that integrates prediction, simulation, and visualization into a unified interface. The results indicate that combining machine learning with digital twin modelling links point-in-time risk estimation with short-term trajectory exploration. While the simulation is based on model-driven assumptions rather than real-time physiological data, it provides an additional analytical layer that supports anticipatory decision-making. This study contributes a scalable, modular framework that bridges predictive analytics and simulation, offering a practical step towards more proactive, data-driven approaches in digital health.
Transformer-Based Abstractive Text Summarisation for Real-Time Web Applications: A Browser-Integrated System with REST API Architecture Revesai, Zvinodashe; Ndlovu, Belinda; Maguraushe, Kudakwashe
Journal of Applied Informatics and Computing Vol. 10 No. 3 (2026): June 2026
Publisher : Politeknik Negeri Batam

Show Abstract | Download Original | Original Source | Check in Google Scholar | DOI: 10.30871/jaic.v10i3.12559

Abstract

The exponential growth of digital textual content has intensified the need for efficient, accessible summarisation tools that support information processing across academic, professional, and research domains. While Transformer-based abstractive summarisation models have demonstrated strong performance in benchmark settings, their real-world deployment remains limited due to computational complexity and lack of user accessibility. This study presents a lightweight Transformer-based abstractive text summarisation system, operationalised as a Google Chrome extension and supported by a REST API, enabling seamless integration into everyday user workflows. The proposed system employs an encoder–decoder framework leveraging a pre-trained Transformer-based encoder and a sequence-to-sequence decoder with attention, fine-tuned on the CNN/Daily Mail dataset. Quantitative evaluation on the benchmark dataset achieved ROUGE-1, ROUGE-2, and ROUGE-L scores of 38.21, 16.54, and 35.12, respectively, demonstrating competitive performance relative to established neural baselines. To address the limitations of lexical evaluation metrics, a complementary human evaluation was conducted using a Likert-scale assessment across coherence, informativeness, and fluency, yielding mean scores above 4.0, thereby confirming the qualitative effectiveness of the generated summaries. In addition to model performance, system-level evaluation assessed functional correctness, latency, scalability, and usability within a real-world deployment context. The system demonstrated stable performance under concurrent usage, with an average response time of 4.2 seconds per request and positive user feedback, validating its practical applicability. The findings demonstrate that high-quality abstractive summarisation can be effectively operationalised within a lightweight, browser-integrated architecture, thereby bridging the gap between research-stage neural models and accessible end-user applications. This work contributes to deployment-oriented natural language processing by emphasising usability, modularity, and real-world integration as critical dimensions of system design.
Deep Learning and XGBoost for Pancreatic Cancer Survival Prediction: A Real-World Evaluation in a Resource-Constrained African Healthcare Setting Revesai, Zvinodashe; Maguraushe, Kudakwashe; Ndlovu, Belinda
Journal of Applied Informatics and Computing Vol. 10 No. 3 (2026): June 2026
Publisher : Politeknik Negeri Batam

Show Abstract | Download Original | Original Source | Check in Google Scholar | DOI: 10.30871/jaic.v10i3.12570

Abstract

Pancreatic cancer remains one of the most lethal malignancies worldwide, with persistently low survival rates and a pressing need for reliable prognostic tools to support treatment planning in resource-constrained healthcare environments. This study presents a structured comparative evaluation of Artificial Neural Network (ANN) and XGBoost classifiers for predicting 12-month survival using real-world clinical data from 569 pancreatic cancer patients treated at a public hospital in Zimbabwe between 2018 and 2023. The Cross-Industry Standard Process for Data Mining (CRISP-DM) framework guided data understanding, preprocessing, model development, and evaluation. A comprehensive preprocessing pipeline incorporating missing value imputation, outlier management, encoding, feature selection, and normalisation was applied, with all transformations derived exclusively from the training set to prevent data leakage. Models were trained using an 80/20 stratified split with cross-validated hyperparameter optimisation and evaluated on a strictly held-out test set using accuracy, precision, recall, F1-score, ROC analysis, and McNemar’s test. On the test dataset, the ANN model achieved 99% overall accuracy and 99% F1-score, outperforming XGBoost, which attained 90% accuracy and 90% F1-score. The performance difference was statistically significant (p < 0.05). Computational analysis demonstrated inference times below 3 milliseconds per sample, supporting feasibility for clinical deployment. While results indicate strong discriminative capacity within this single-centre dataset, external validation across multi-institutional cohorts is necessary to confirm generalisability. These findings suggest that supervised machine learning can provide clinically meaningful support for survival prediction in African tertiary healthcare settings. This study uniquely contributes a deployment-oriented, real-world evaluation of machine learning models within a resource-constrained African healthcare context, addressing a critical gap in the current oncology informatics literature.