Accra, Ghana

Samuel
Ankapong.

ML Engineering & Research

Model evaluation, biomedical signals,
and connected sensing systems.

View research CV PDF

Methods, measurements, and their limits.

01 ML engineering & research

Build the pipeline.
Question the result.

I build and evaluate machine-learning pipelines for biomedical data. My work includes neuroimaging foundation-model evaluation, GAN-based EEG denoising, and wearable signal processing.

I hold a BSc in Biomedical Engineering from the University of Ghana, with Second Class Upper honours and a GPA of 3.49/4.0. I also build the software and embedded systems around these projects.

Research interests

Surgical robotics & computer-assisted surgery, rehabilitation robotics, sensing & wearable devices.

Explore the research

02 ML & research

Experiments and findings.

Jun 2026 — Aug 2026Machine learning & model training

Do neuroimaging foundation models generalise to African clinical MRI?

Evaluating the Generalization of Neuroimaging Foundation Models to African Clinical MRI

Implemented the training and evaluation pipeline for BrainIAC, NeuroJEPA, NeuroVFM, and Primus on a Nigerian clinical brain-MRI cohort of 79–80 subjects, classifying Control, Dementia, and Parkinson’s disease.

Benchmarked linear probes across T1w, T2w, T1w+T2w, and FLAIR using five-fold subject-level cross-validation against an end-to-end 3D Vision Transformer.

Reported findings on this cohort
ApproachFinding
Four frozen backbonesCollapsed to majority-class prediction; MCC ≈ 0.
End-to-end ViT3DUp to 53.4% accuracy; MCC = 0.27. Non-trivial recall across all three classes.

These findings describe this small clinical cohort, not the models’ performance on every population or dataset.

  • Foundation-model evaluation
  • Linear probing
  • Subject-level cross-validation
  • ViT3D
Presentation status
Dec 2024 — Sept 2025Undergraduate researcher

GAN-based synthetic data for EEG denoising

Trained a conditional Wasserstein GAN with gradient penalty on EEGdenoiseNet to generate 8,840 noisy–clean EEG pairs. Developed MulSepEEGDenoiser, a multi-branch U-Net with dedicated EMG and EOG paths, Squeeze-and-Excitation attention, and adaptive indicator conditioning.

SNR improvement
4.75 dB
RRMSE
0.746
RRMSE improvement vs baseline
13.3%

Compared with Deep Separator, the model improved correlation by 30.5%; the strongest reported alpha-band correlation was 0.708.

Advisor: Dr. Srinivasan Shankar Balapangu. Results are model-evaluation findings, not evidence of clinical validation.

  • cWGAN-GP
  • U-Net
  • EEG / EOG / EMG
  • Synthetic data
Presentation & proceedings status
Sept 2023 — Nov 2023Undergraduate researcher

BCPRED: a breast-tumour classification application

Reduced the Wisconsin Breast Cancer Database’s 30 features to 16 using correlation heatmaps, swarm plots, and violin plots. Benchmarked Logistic Regression, Random Forest, and XGBoost on the 569-sample dataset.

XGBoost accuracy
0.97
Precision
0.96
F1-score
0.97

Built a Streamlit application for entering cell-nuclei parameters and viewing benign/malignant predictions with confidence scores.

Reported scores are from the project evaluation. The application should not be presented as a clinically validated diagnostic tool.

  • Python
  • scikit-learn
  • XGBoost
  • Streamlit

03 Sensing & systems

From a signal to an application.

Apr 2026Firmware & mobile development

EMGrec: wireless EMG acquisition

Wrote ESP32 C++ firmware to stream muscle-sensor data at 1,000 samples per second over Bluetooth Low Energy, separating sampling and Bluetooth communication across processor cores.

Built the companion React Native app for live signal viewing, recording, and export, with remote start, stop, and reset controls.

  • ESP32
  • C++ / FreeRTOS
  • Bluetooth Low Energy
  • React Native
View EMGrec on GitHub
Jun 2025 — Sept 2025Undergraduate researcher

An ankle-mounted step counter for remote rehabilitation monitoring

Built a Raspberry Pi 5 and MPU6050 wearable with a 100 Hz on-device processing pipeline, including gravity removal and second-order low-pass filtering. Developed a weighted step detector using spectral entropy, dominant frequency, peak-to-peak amplitude, and regularity variance.

Connected the device through TLS-encrypted Wi-Fi with BLE fallback to a cloud application and real-time clinician dashboard.

Validation & limitations

Validation against manual counts showed 43% overcounting at slow gait and 26% at normal gait. Analysed error sources and proposed refractory, feature-weighting, and stationary-gating refinements.

Advisor: Dr. Srinivasan Shankar Balapangu. Proposed refinements are not reported here as implemented improvements.

  • Raspberry Pi 5
  • IMU signal processing
  • Telemetry
  • Ground-truth validation

04 Research outputs

Presentations and research work.

Accepted for oral presentation

Evaluating the Generalization of Neuroimaging Foundation Models to African Clinical MRI

Akinmuleya, O.I., Akano, O.S., Ankapong, S.D., Lawal, O., & Musah, T.

AFRICAI Workshop, MICCAI 2026.

Presented · Proceedings publication pending

Design of a Deep Learning EEG Filter Using GAN-Based Synthetic Data for Simultaneous Removal of EOG and EMG Artifacts

Balapangu, S.S., Ankapong, S.D., Yawuliga, D.B., Asiamah, R., Aidoo, E., Gakpo, H.S., Opoku-Boateng, A.T., & Kyei, J.J.A.

11th IEEE International Conference on Adaptive Science & Technology (ICAST 2026), Accra, Ghana. Presented 22 October 2026.

Conference presentation · November 2023

Developing a Robust Web-Based Breast Tumour Classifier Using Machine Learning

1st Annual Engineering Conference, School of Engineering Sciences, University of Ghana, Legon.

Results, dates, and presentation statuses are taken from my supplied research CV. No proceedings publication is implied where it is still pending.

Teaching & mentorship

Oct 2025 — Apr 2026

Teaching Assistant, Bioinformatics

Supported lectures and tutorials at the University of Ghana and guided Python practical sessions, including ML-based drug-target prediction.

July 2025

Lead Organizer & Tutor, AI Essentials for Engineers

Designed and delivered a five-week workshop covering AI fundamentals, prompt engineering, and introductory AI agents.

Methods & tools

  • PyTorch
  • scikit-learn
  • XGBoost
  • Hugging Face Transformers
  • NumPy
  • Pandas
  • Matplotlib
  • Seaborn
  • Python
  • C/C++
  • Bash
  • Git
  • Linux
  • LaTeX