Final Project: Fetal Health Prediction Using CTG Analysis

Project Goal

To predict fetal health using Cardiotocography (CTG) results through machine learning (ML).

Background

Cardiotocography (CTG) is a non-invasive procedure that continuously monitors the fetal heart rate (FHR) and uterine contractions using an ultrasound transducer on the mother's abdomen. CTG is widely used, especially in high-risk pregnancies, to assess fetal well-being. During CTG, several metrics are monitored, including:

• Uterine Contractions
• Baseline Heart Rate
• Variability
• Accelerations and Decelerations

The standard duration of CTG monitoring is 30 minutes, though this may extend if patterns appear suspicious. FHR classifications include:

• Baseline FHR
• Oscillations (oscillation rate and amplitude)

Various factors influence FHR, including maternal, fetoplacental, fetal, and external elements.

Problem Statement

Interpreting CTG data requires trained professionals, posing challenges in regions with limited healthcare resources. This project aims to address this gap by developing a machine learning model to assist in CTG interpretation.

Dataset

The dataset used in this project can be accessed on Kaggle.

Feature Information

• Baseline Value: FHR in beats per minute
• Accelerations: Accelerations per second
• Fetal Movement: Fetal movements per second
• Uterine Contractions: Uterine contractions per second
• Light, Severe, and Prolonged Decelerations: Rates of different deceleration types
• ASTV / MSTV: Short-term variability (abnormal percentage and mean)
• ALTV / MLTV: Long-term variability (abnormal percentage and mean)
• FHR Histogram: Width, minimum, maximum, mode, median, peaks, zeroes, variance, and tendency
• Target: Fetal health status (Normal, Suspected, Pathological)

Data Preprocessing

• Renamed columns for clarity
• Removed 13 duplicate entries

Exploratory Data Analysis (EDA)

• Analyzed target distribution with value counts and proportions
• Visualized feature distributions using box plots and histograms
• Computed a correlation matrix for numerical features
• Analyzed feature importance

Machine Learning Process

• Addressed class imbalance with SMOTE
• Applied Robust Scaling to features
• Encoded labels for the XGBoost model, with decoded interpretations
• Trained and evaluated five models:
  • K-Nearest Neighbors (KNN)
  • AdaBoost
  • Gradient Boosting
  • Random Forest (RF)
  • XGBoost

Additional Resources

• Slides: Project Slides
• Tableau Dashboard: Tableau Visualization
• Streamlit App: Deployed Application