Fetal Health Classification Using SVM & Decision Tree 🤰📊

---

📚 Table of Contents

• Overview
• Key Contributions
• Dataset
• Data Preprocessing
• Evaluation Strategy
• Models
• Dimensionality Reduction & Feature Selection
• Results
• Discussion
• Conclusion

---

Overview

This project builds and compares two classic classifiers—Support Vector Machine (SVM) and Decision Tree (DT)—to predict fetal health status (Normal, Suspect, Pathological) from Cardiotocogram (CTG) measurements. Careful handling of class imbalance and feature engineering maximizes detection of high‑risk cases.

---

Key Contributions

• Balanced a 3‑class CTG dataset via manual oversampling to ensure equal representation.
• Demonstrated the effect of feature scaling, hold‑out evaluation, and stratified splitting on model performance.
• Benchmarked SVM vs. DT across accuracy and macro F1‑score, showing DT + RFE achieved the best results.
• Explored dimensionality reduction (PCA, LDA) and feature selection (Correlation + RFE) to optimize model performance.

---

Dataset

• Source: Kaggle “Fetal Health Classification” (2,126 CTG records).
• Classes (original counts):
  • Normal: 1,655
  • Suspect: 295
  • Pathological: 176

---

Data Preprocessing

1. Balancing
   • Duplicated minority‑class samples to match the Normal class count (1,655 each).
2. Scaling
   • Applied StandardScaler so all features have zero mean and unit variance.
3. Train/Test Split
   • 80/20 stratified hold‑out split to preserve class ratios.

---

Evaluation Strategy

• Hold‑Out Method with stratification
• Metrics:
  • Overall Accuracy
  • Macro F1‑Score (gives equal weight to each class)

---

Models

Model	Key Settings
SVM	Linear kernel; C tuned via grid search
Decision Tree (DT)	Gini criterion; max depth tuned via cross‑validation

---

Dimensionality Reduction & Feature Selection

1. PCA (95% Variance)
   • Reduced 21 features → 14 components; slight drop in accuracy.
2. LDA (2 Components)
   • Reduced to 2 dimensions; further drop in accuracy.
3. Correlation + RFE
   • Filtered out low‑correlation features, then applied Recursive Feature Elimination to pick the top 10.
   • Best performance:
     • DT + RFE → 91.5% accuracy
     • SVM + RFE → 89.4% accuracy

---

Results

Configuration	Accuracy	Macro F1‑Score
SVM (raw features)	88.0%	0.805
DT (raw features)	89.7%	0.812
SVM + PCA	87.1%	0.79
DT + PCA	87.2%	0.79
SVM + LDA	83.6%	0.75
DT + LDA	86.6%	0.80
SVM + Correlation + RFE	89.4%	0.82
DT + Correlation + RFE	91.5%	0.84

---

Discussion

• Feature selection (Correlation + RFE) outperformed dimensionality reduction, highlighting the importance of retaining informative variables rather than compressing them.
• Decision Tree proved more balanced on minority classes than SVM, especially after RFE.
• PCA and LDA helped visualization but removed discriminative power for this multiclass problem.

---

Conclusion

Balancing the dataset, scaling features, and strategically selecting the most relevant CTG measurements enables a simple Decision Tree to reach 91.5% accuracy in fetal health classification—providing a lightweight, interpretable model for early risk detection.