NeuroScreen
A hybrid CatBoost + ANN ensemble that detects cognitive impairment in insomniac university students — 95.20% accuracy, ROC-AUC 0.982.
Thesis (CSC 4298, Fall 2025–26). Dual-path training with arithmetic-mean probability blending beats every single model, showing that complementary learners can exceed either alone on survey-scale health data.
[ 01 ] — Results
Metrics
[ 02 ] — Method
How it works
- Built a custom survey dataset of 2,237 Bangladeshi university students (ages 20–35) across 7 cumulative cognitive-symptom categories, engineered and scaled into features.
- Dual-path training: CatBoost (GBDT) on the tabular side, and a 3-layer ANN (128-64-1 MLP, ReLU, Adam, dropout 0.3) on the deep side.
- Final decision = arithmetic mean of the two models' probabilities, evaluated on a stratified 80/20 split.
[ 03 ] — Why it wins
Baseline comparison
| Model | Accuracy | Delta |
|---|---|---|
| NeuroScreen ensemblewinner | 95.20% | baseline |
| CatBoost standalone | 94.12% | +1.08 vs best |
| ANN standalone | 91.25% | +3.95 vs best |
Ensemble blending (arithmetic mean of probabilities) lifts accuracy above both standalone models — complementary learners catch what each alone misses.
[ 04 ] — Interpretability
What predicts impairment
01
Mental / Physical Fatigue
importance 0.095
02
Stress Frequency
importance top-3
03
GPA Impact
importance 0.095
[ 05 ] — Related work
More research
Improving Generalization of Image-Based Water Turbidity Classification Using Fine-Tuning with Small Real-World Datasets
Computer Vision & Pattern Recognition · 2025
EfficientNet-B0 + transfer learning on ~300 self-collected phone photos, with auxiliary cat/dog/panda data to fight overfitting.
Explainable Bangla Toxic Comment Detection using BanglaBERT with SHAP and LIME
Data Mining (IEEE format) · 2025
BanglaBERT with SHAP + LIME explanations, k-fold (5) and 70/15/15 splits across multiple Bangla toxicity datasets.
Financial Sentiment Analysis with Fine-Tuned Transformers
NLP course research · 2025
Fine-tuned FinBERT vs vanilla BERT on financial text — why domain-specific transformers win on jargon-heavy corpora.
[ 06 ] — Want the full write-up?
Get the thesis
The complete thesis report covers the literature review, dataset construction, full feature set, and 16 reference papers. Reach out and I'll share it.
Request the thesis