🧠 Parkinson’s Disease Classifier (XGBoost)

This project builds a compact, high-performance classifier to detect Parkinson’s disease using biomedical voice features from the UCI Parkinson’s dataset.

The final model uses only the top 6 most important features and achieves over 92% accuracy while maintaining high precision and recall — even with class imbalance.

---

✅ Project Highlights

• Model: Tuned XGBoost Classifier (with baseline and untuned comparisons)
• Feature Set: Top 6 most important voice features (feature selection based on XGBoost importance scores)
• Imbalance Handling: scale_pos_weight based on class ratio
• Model Format: Saved using pickle for portability and reuse

---

🔝 Top 6 Most Important Features

Rank	Feature	Description
1️⃣	mdvp:fhi	Maximum fundamental frequency (Hz) — Highest vocal pitch. Parkinson’s can cause instability, increasing max frequency irregularities.
2️⃣	mdvp:fo	Average fundamental frequency (Hz) — Average vocal pitch. Often reduced or unstable in Parkinson’s speech.
3️⃣	spread1	Nonlinear signal spread — Measures asymmetry in the voice signal; higher values may reflect vocal tremor or breathiness.
4️⃣	rpde	Recurrence Period Density Entropy — Quantifies unpredictability in the signal. Elevated in disordered voices.
5️⃣	d2	Correlation dimension — Measures complexity of the vocal system. Lower complexity often seen in Parkinson’s.
6️⃣	spread2	Second nonlinear spread measure — Complements spread1; also tracks signal deviation and dysphonia symptoms.

---

🧠 Takeaway

The top 6 features identified by XGBoost — primarily nonlinear vocal measures like mdvp:fhi, rpde, and spread1 — capture instability, irregularity, and reduced complexity in voice signals. These characteristics are strongly associated with Parkinson’s-induced dysphonia, making them highly predictive even in a compact model.

---

📊 Final Model Performance

Metric	Class 0 (Healthy)	Class 1 (Parkinson’s)	Macro Avg	Weighted Avg
Precision	0.89	0.93	0.91	0.92
Recall	0.80	0.97	0.88	0.92
F1-score	0.84	0.95	0.90	0.92
Support	10	29	–	39
Accuracy	–	–	–	92.3%

---

🔢 Confusion Matrix

Actual \ Predicted	0 (Healthy)	1 (Parkinson’s)
0 (Healthy)	8	2
1 (Parkinson’s)	1	28

---

📂 Saved Models

File Name	Description
parkinsons_xgb_top6_tuned.pkl	🔧 Tuned XGBoost model using top 6 features (RandomizedSearchCV object)
xgboost_baseline_model.pkl	⚙️ Baseline XGBoost model with default hyperparameters
parkinsons_xgb_top6.pkl	🧪 Untuned XGBoost model using top 6 selected features

---

🔧 Requirements

All dependencies are listed in requirements.txt. Key packages include:

Package	Purpose
xgboost	Gradient-boosted trees for classification
scikit-learn	Model evaluation, cross-validation
ucimlrepo	Load datasets from the UCI ML repository
pandas	Data handling and preprocessing
numpy	Numerical operations
matplotlib	Data visualization
seaborn	Statistical visualizations

---

📦 Installation

Clone the repository and install dependencies:

git clone https://github.com/your-username/parkinsons-xgb.git
cd parkinsons-xgb
pip install -r requirements.txt

import pickle

# Load top 6 feature list
with open('top_features.pkl', 'rb') as f:
    top_features = pickle.load(f)

# Load tuned model (RandomizedSearchCV object)
with open('parkinsons_xgb_top6_tuned.pkl', 'rb') as f:
    rs = pickle.load(f)
best_model = rs.best_estimator_

# Load baseline model
with open('xgboost_baseline_model.pkl', 'rb') as f:
    baseline_model = pickle.load(f)

# Load untuned top-6 model
with open('parkinsons_xgb_top6.pkl', 'rb') as f:
    top6_model = pickle.load(f)

# Example prediction
# X_new is a pandas DataFrame with the same structure as training features
y_pred = best_model.predict(X_new[top_features])