Publications

An empirical study of machine learning robustness and scalability for imbalanced tabular clinical data in emergency and critical care

Published in Scientific Reports, 2026

Machine learning can support high-stakes decision-making in emergency and intensive care settings, but severe class imbalance in clinical data limits model reliability and biases predictions toward majority outcomes. We evaluate six model families, including classical methods (Decision Tree, Random Forest, XGBoost), deep learning approaches (TabNet), and tabular foundation models (TabICL, TabPFN v2.6), on MIMIC-IV-ED and eICU datasets across multiple clinical prediction tasks. Models are assessed using Macro F1-score, robustness to increasing imbalance, and computational efficiency. Results show dataset-dependent performance: TabPFN and TabICL perform strongly on MIMIC-IV-ED, while XGBoost leads on eICU. No single model dominates across all settings, but foundation models provide a favorable efficiency–performance trade-off and are increasingly competitive in imbalanced clinical prediction scenarios. Read more

Recommended citation: Brima, Y., Atemkeng, M. An empirical study of machine learning robustness and scalability for imbalanced tabular clinical data in emergency and critical care. Sci Rep 16, 18004 (2026) https://doi.org/10.1038/s41598-026-56413-9

Few-shot Cross-country Generalization of Tabular Machine Learning and Foundation Models for Childhood Anemia Prediction under Distribution Shift

Published in arXiv/preprint, 2026

Childhood anemia affects ~40% of children aged 6–59 months globally and is driven by heterogeneous, context-dependent factors that challenge model generalization. We evaluate a transformer-based tabular foundation model (TabPFN v2.6) against classical methods (Logistic Regression, XGBoost, LightGBM) using DHS data from 16 countries (n=68,856) under cross-country and data-scarce settings. Across LOCO, reverse-LOCO, and few-shot evaluations, TabPFN shows superior performance in low-data regimes, with improved calibration (Brier: 0.042; ECE: 0.203) and strong discrimination. While full-data performance differences are modest, results indicate that population heterogeneity dominates predictive performance, and foundation models offer advantages for robust, low-resource global health prediction. Read more

Recommended citation: Brima Y, Atemkeng M, Kallon LH, Niyukuri D, Vacavant A, Saidu S, Chen DG. Few-shot Cross-country Generalization of Tabular Machine Learning and Foundation Models for Childhood Anemia Prediction under Distribution Shift. arXiv preprint arXiv:2605.26589. 2026 May 26. https://arxiv.org/abs/2605.26589

Bridging visual saliency and large language models for explainable deep learning in medical imaging

Published in arXiv/preprint, 2026

The lack of interpretability in deep learning remains a major barrier to clinical adoption in medical imaging. We propose a multimodal explainability framework for brain tumor classification that combines CNN-based prediction, saliency-based localization (Grad-CAM, Grad-CAM++, ScoreCAM), neuroanatomical mapping via the Harvard-Oxford atlas, and large language models to generate radiology-style diagnostic reports. Evaluated on 4,834 contrast-enhanced T1-weighted MRI scans across three tumor classes, the framework integrates classification, segmentation, and structured report generation into a unified pipeline. InceptionResNetV2 achieved the best classification performance, while Grad-CAM++ produced the most accurate localization. Among LLMs, Grok3 and LLaMA showed complementary strengths in coherence and readability. The results demonstrate that coupling visual attribution with anatomical grounding and language generation improves interpretability and supports more clinically meaningful AI explanations. Read more

Recommended citation: Nguezet PV, Fute ET, Brima Y, Azanguezet BM, Atemkeng M. Bridging visual saliency and large language models for explainable deep learning in medical imaging. arXiv preprint arXiv:2605.06197. 2026 May 7. https://arxiv.org/abs/2605.06197

Artificial intelligence in neglected tropical diseases: current applications, challenges, and opportunities

Published in arXiv/preprint, 2026

Neglected tropical diseases (NTDs) affect over one billion people globally and disproportionately impact low-resource tropical and subtropical regions. While artificial intelligence (AI) offers significant opportunities to strengthen NTD control, the scope, maturity, and translational readiness of existing applications remain unclear. We conducted a systematic review of AI methods applied to NTD-related problems using Scopus and Web of Science, following PRISMA guidelines. Across 289 studies, AI applications were primarily concentrated in diagnosis, surveillance, and prognosis, with most relying on supervised learning from imaging, clinical, or epidemiological data. However, external validation, calibration, and prospective evaluation were limited, and research was geographically concentrated in a small number of countries. Overall, despite rapid growth in the field, substantial gaps remain in methodological rigor, equity, and real-world deployment readiness, highlighting the need for more robust, transparent, and context-aware AI research in NTDs. Read more

Recommended citation: Brima Y, Atemkeng M, Ngueajio MK, Ngueabou Y, Nguembang Fadja A, Bonginkosi N, et al. Artificial intelligence in neglected tropical diseases: current applications, challenges, and opportunities. NPJ Digit Med. 2026. Submitted for publication. https://arxiv.org/abs/2604.00123

An unsupervised XAI framework for dementia detection with context enrichment

Published in Scientific reports, 15, 1 (2025), 2025

This work introduces a framework that evaluates XAI methods by integrating neuroanatomical morphological features with CNN-generated relevance maps for disease classification. We trained a CNN using brain MRI scans from six cohorts: ADNI, AIBL, DELCODE, DESCRIBE, EDSD, and NIFD (N = 3253), including participants that were cognitively normal, with amnestic mild cognitive impairment, dementia due to Alzheimer’s disease and frontotemporal dementia Read more

Recommended citation: Singh, D., Brima, Y., Levin, F., Becker, M., Hiller, B., Hermann, A., Villar-Munoz, I., Beichert, L., Bernhardt, A., Buerger, K. and Butryn, M., 2025. An unsupervised XAI framework for dementia detection with context enrichment. Scientific reports, 15(1), p.39554 https://doi.org/10.1038/s41598-025-26227-2

A systematic review of low-rank and local low-rank matrix approximation in big data medical imaging

Published in Neural Computing and Applications, 2025

Medical imaging datasets pose challenges in storage, transmission, and analysis due to their complexity. Low-rank matrix approximation (LORMA) and its variant, LLORMA, offer efficient data representation, with LLORMA gaining preference since 2015 for capturing structural details. This review examines their applications, limitations in similarity measurement, and the potential of semantic segmentation for improvement. We also explore their extension to structured data, the impact of missing values, and propose a hybrid optimization approach to enhance patch selection and overall applicability in medical imaging. Read more

Recommended citation: Hamlomo, S., Atemkeng, M., Brima, Y. et al. A systematic review of low-rank and local low-rank matrix approximation in big data medical imaging. Neural Comput & Applic (2025). https://doi.org/10.1007/s00521-025-11055-2 https://doi.org/10.1007/s00521-025-11055-2

Understanding Self-Supervised Learning of Speech Representation via Invariance and Redundancy Reduction

Published in MDPI Information, 15(2), 114 (2024), 2024

This paper explores self-supervised learning for speech representation, leveraging redundancy reduction techniques to learn robust representations. It adapts the Barlow Twins framework to speech processing and evaluates its effectiveness in various downstream tasks. Read more

Recommended citation: Brima, Y., et al. (2024). Understanding Self-Supervised Learning of Speech Representation via Invariance and Redundancy Reduction. MDPI Information, 15(2), 114. DOI: 10.3390/info15020114 https://doi.org/10.3390/info15020114

Saliency-driven Explainable Deep Learning in Medical Imaging: Bridging Visual Explainability and Statistical Quantitative Analysis

Published in BioData Mining, 17, 18 (2024), 2024

This work explores explainable deep learning techniques in medical imaging, bridging visual explainability with statistical quantitative analysis. It proposes a novel approach that enhances the interpretability of AI models applied to medical diagnostics. Read more

Recommended citation: Brima, Y., Atemkeng, M. (2024). Saliency-driven explainable deep learning in medical imaging: bridging visual explainability and statistical quantitative analysis. BioData Mining, 17, 18. DOI: 10.1186/s13040-024-00370-4 https://doi.org/10.1186/s13040-024-00370-4

Bridging the Gap: Exploring Interpretability in Deep Learning Models for Brain Tumor Detection and Diagnosis from MRI Images

Published in Information, 15(4), 182 (2024), 2024

This paper investigates the interpretability of deep learning models in detecting and diagnosing brain tumors from MRI images. It evaluates various explainability techniques and their effectiveness in clinical applications. Read more

Recommended citation: Nhlapho, W., Atemkeng, M., Brima, Y., et al. (2024). Bridging the Gap: Exploring Interpretability in Deep Learning Models for Brain Tumor Detection and Diagnosis from MRI Images. Information, 15(4), 182. DOI: 10.3390/info15040182 https://doi.org/10.3390/info15040182

Learning Disentangled Audio Representations through Controlled Synthesis

Published in ICLR Tiny Papers Track 2024, 2023

This paper introduces a novel approach to learning disentangled speech representations through controlled synthesis, enabling better interpretability and generalization in speech processing tasks. Read more

Recommended citation: Brima Y, Krumnack U, Pika S, Heidemann G. (2023). Learning Disentangled Audio Representations through Controlled Synthesis. ICLR Tiny Papers Track 2024. https://openreview.net/forum?id=Fn9ORH8PLl

Learning Disentangled Speech Representations

Published in New in Machine Learning Workshop, NeurIPS 2023, 2023

This study investigates methods for learning disentangled representations of speech signals, focusing on separating different speech-related factors such as speaker identity, linguistic content, and emotional state. The work applies novel deep learning approaches to improve representation learning in speech processing. Read more

Recommended citation: Brima, Y., Krumnack, U., Pika, S., & Heidemann, G. (2023). Learning Disentangled Speech Representations. New in Machine Learning Workshop, NeurIPS 2023. https://openreview.net/forum?id=3ox1TfKeRF

Self-Supervised Learning of Speech Representation via Redundancy Reduction

Published in Gesellschaft für Informatik e.V.. pp. 11-19. Doctoral Consortium at KI 2023. Berlin. 45195, 2023

This work aims to investigate a novel self-supervised learning (SSL) method for speech representation that leverages redundancy reduction techniques to learn robust representations capturing speaker characteristics. Our proposed approach builds upon the Barlow Twins framework, introduced in computer vision, and we adapt it to speech processing. The primary objective is to assess the quality of the learned representations through comprehensive evaluations of various downstream tasks, including speaker identification, gender recognition, emotion recognition, and more. Read more

Recommended citation: Brima, Yusuf (2023): Self-Supervised Learning of Speech Representation via Redundancy Reduction. DC@KI2023: Proceedings of Doctoral Consortium at KI 2023. DOI: 10.18420/ki2023-dc-02. Gesellschaft für Informatik e.V.. pp. 11-19. Doctoral Consortium at KI 2023. Berlin. 45195 https://doi.org/10.18420/ki2023-dc-02

A Mathematical Framework for Understanding Recognition Systems

Published in arXiv preprint biorxiv:2023.06.08.544240, 2023

We propose a theoretical framework for biological recognition, integrating category theory, information theory, dynamical systems modeling, and optimization theory. This framework defines individual recognition (IR) and class-level recognition (CR) as functors between stimuli and responses. We identify five conditions for robust IR systems, termed “signature systems,” and model them as attractor states, offering insights into communication and language evolution. Read more

Recommended citation: Brima, Yusuf. "A Mathematical Framework for Understanding Recognition Systems" biorxiv preprint biorxiv:2023.06.08.544240 (2023). https://doi.org/10.1101/2023.06.08.544240

Deep Transfer Learning for Brain Magnetic Resonance Image Multi-class Classification

Published in Dhaka University Journal of Applied Science and Engineering, 6(2), 14–29 (2022), 2022

This paper presents a deep transfer learning approach for multi-class classification of brain MRI images, evaluating its effectiveness in differentiating various neurological conditions. Read more

Recommended citation: Brima, Y., Kamal Tushar, M. H., Kabir, U., Islam, T. (2022). Deep Transfer Learning for Brain Magnetic Resonance Image Multi-class Classification. Dhaka University Journal of Applied Science and Engineering, 6(2), 14–29. DOI: 10.3329/dujase.v6i2.59215 https://doi.org/10.3329/dujase.v6i2.59215

Transfer Learning for the Detection and Diagnosis of Types of Pneumonia including Pneumonia Induced by COVID-19 from Chest X-ray Images

Published in Diagnostics, 11(8), 1480 (2021), 2021

This study explores the use of transfer learning for detecting and diagnosing different types of pneumonia, including COVID-19-induced pneumonia, from chest X-ray images. The work evaluates the performance of various deep learning models in assisting medical professionals with automated diagnostics. Read more

Recommended citation: Brima, Y., Atemkeng, M., Tankio Djiokap, S., Ebiele, J., & Tchakounté, F. (2021). Transfer Learning for the Detection and Diagnosis of Types of Pneumonia including Pneumonia Induced by COVID-19 from Chest X-ray Images. Diagnostics, 11(8), 1480. DOI: 10.3390/diagnostics11081480 https://doi.org/10.3390/diagnostics11081480

Brain MRI Dataset

Published in Figshare/Dataset, 2021

Published a dataset that was curated in collaboration between the Computer Science and Engineering Department, University of Dhaka and the National Institute of Neuroscience, Bangladesh. It comprise 5,285 T1-weighted contrast- enhanced brain MRI images belonging to 38 categories. Read more

Recommended citation: Yusuf Brima, Mossadek Hossain Kamal, Upama Kabir, and Tariqul Islam (2021). "Brain MRI Dataset." Figshare/Dataset. 1(1). https://doi.org/10.6084/m9.figshare.14778750.v2

Yusuf Brima, Ph.D.

Publications