Research
Four active research tracks turning undergraduate talent into real published work — from decoding brain signals to reading medical scans.
Afrium's research groups work the way real labs do — a research question, a public dataset, a baseline, months of iteration, and a paper aimed at a real venue. Every track below is active work in progress, run by student researchers under mentor supervision.
Methodology & Standards
How we work
Research is evaluated by its rigor and reproducibility, not by the seniority of who wrote it.
Real public datasets
Rigorous research on established open benchmarks (PhysioNet, Temple EEG, BCI Competition, StressID) — never toy or synthetic problems.
Subject-independent, honest evaluation
Cross-subject splits and zero data leakage. No shortcuts in evaluation pipelines or validation splits.
Explainability & robustness
Verification through attribution methods (Grad-CAM, LRP) and artifact rejection before any result is trusted.
Targeting international venues
Every track is engineered from day one toward peer-reviewed submission at premier IEEE, NeurIPS, and ICLR conferences.
Active Investigation Areas
Research Tracks
Four parallel groups executing structured experiments, validation pipelines, and paper drafts.
EEG-Based Brain Decoding
“Decoding thought from signal.”
Working on covert and imagined speech decoding directly from non-invasive EEG, using self-supervised representation learning to extract meaningful structure from noisy neural data, and explainable AI to verify the model is learning real physiological signal rather than artefact. Sits at the intersection of neuroscience, signal processing, and representation learning.
PhysioNet EEG Motor Imagery, GigaScience BCI, BCI Competition IV-2a
Wearable Multimodal Biosignal Intelligence
“Reading the body in real time.”
Fusing signals from wearable sensors — EEG, ECG, PPG, GSR/EDA, EMG, respiration — into models for real-time human activity recognition, rehabilitation monitoring, and physiological state and stress detection. Focused on models that generalize across individuals, not just within one person's data.
StressID, PhysioNet K-EmoPhone, WESAD, CASE Dataset
Self-Supervised and Contrastive Learning
“Learning without needing every label.”
A methods-focused track building representation-learning approaches that transfer across physiological time-series data — bridging what's learned from EEG with what's learned from wearables, using contrastive pretraining to reduce dependence on labeled data. Feeds directly into Tracks A and B.
Massive Multi-Subject Physiological Time-Series Repositories
Medical Imaging and AI
“Extending signal intelligence to sight.”
Applying machine learning to medical imaging problems, carrying the same rigor — real clinical-style datasets, explainability, robustness testing — from biosignals into visual diagnostic data.
NIH ChestX-ray14, ISIC Dermoscopy, CheXpert, BraTS
Active Project
A deeper look at one of the ongoing investigations underway across our cohorts.
A Reliability-Aware, Individual-Adaptive Approach to Multimodal Physiological Stress Detection
Modeling stress detection across individuals using multimodal physiological signals, with reliability-aware adaptation that accounts for how physiological stress responses vary from person to person — aimed at models that hold up outside the lab, not just on one dataset.
Our Standard