MLers – (Lasmar, Maésha, Richard, Ahmed)

🧩 Where it started

Alzheimer’s detection still relies heavily on expensive tools like MRI or PET scans.

EEG offers a cheaper, non-invasive alternative - but the signal is noisy, and datasets are small.

That’s what made this challenge interesting.

---

⚙️ What we built

We built a model that classifies EEG recordings as Alzheimer’s Disease (AD) or Cognitively Normal (CN).

Instead of jumping straight into deep learning, we focused on extracting meaningful patterns:

• brain wave frequency bands (delta → gamma)
  
• statistical and spectral features
  
• clinically relevant ratios (like theta/alpha)
  

---

🏗️ How we approached it

With only ~38 training subjects, complexity wasn’t the priority - validity was.

So we:

• segmented signals using sliding windows (30s / 15s overlap)
  
• used XGBoost for fast iteration and interpretability
  
• enforced Leave-One-Subject-Out cross-validation (LOSO-CV) to avoid leakage
  

Every prediction was made at the subject level, not just individual signal snippets.

---

⚠️ Where it got difficult

The hardest part wasn’t building the model - it was trusting the results.

• extremely small dataset
  
• high risk of overfitting
  
• class imbalance
  
• misleading validation if not done properly
  

We constantly had to ask:

“Is this real performance… or are we fooling ourselves?”

---

🏆 What we’re proud of

• achieving ~81.6% subject-level accuracy on a very limited dataset
  
• building a pipeline that avoids data leakage end-to-end
  
• balancing performance with interpretability (important for clinical context)
  
• exploring both classical ML and deep learning approaches
  

---

🧠 What we learned

• evaluation strategy matters more than model choice in small datasets
  
• deep learning isn’t always the answer - simpler models can outperform when data is limited
  
• domain knowledge (EEG biomarkers) is critical for feature design
  
• “good results” mean nothing without proper validation
  

---

🚀 What’s next for MLers

(Lasmar, Maésha, Richard, Ahmed)

• scaling the approach with larger datasets
  
• improving generalization and robustness
  
• exploring hybrid models (XGBoost + deep learning ensembles)
  
• pushing further into AI for healthcare applications
  

Built With

• cuda
• jupyter-notebook
• matplotlib
• mps
• numpy
• pandas
• python
• pytorch
• pywavelets
• scikit-learn
• scipy
• xgboost