Multi-Class Prediction of Obesity Risk

This project notebook leverages machine learning models to classify individuals based on obesity risk levels. The classification is built upon a Kaggle dataset containing various health and lifestyle attributes.

Dataset

• Source: Obesity or CVD Risk - Classify/Regressor/Cluster Dataset: https://www.kaggle.com/datasets/aravindpcoder/obesity-or-cvd-risk-classifyregressorcluster
• Attributes:
  • Eating habits: Frequency of high-calorie food consumption (FAVC), vegetable consumption (FCVC), meal frequency (NCP), snacking habits (CAEC), water intake (CH20), alcohol consumption (CALC).
  • Physical condition: Calories monitoring (SCC), physical activity frequency (FAF), technology use time (TUE), transportation type (MTRANS).
  • Demographic data: Gender, age, height, and weight.
• Target Variable: Obesity levels defined as follows: -Underweight (<18.5 BMI) -Normal (18.5–24.9 BMI) -Overweight (25.0–29.9 BMI) -Obesity I (30.0–34.9 BMI) -Obesity II (35.0–39.9 BMI) -Obesity III (>40 BMI)

Project Workflow

1. Data Exploration:

   • Basic statistics and data insights (e.g., mean, median).
   • Data quality checks, including handling missing values and data types.

2. Data Visualization:

   • Visualization of key features to reveal potential patterns, correlations, and class imbalances.
   • Use of histograms, scatter plots, and correlation heatmaps to illustrate relationships between health metrics and risk factors.

3. Data Manipulation:

   • Data distribution and outliers detection.
   • Data transformations manually to reduce ouliers and improve subsequent interpretation by the models.

4. Feature Engineering:

   • Encoding categorical features and scaling numerical features using scikit-learn
   • Feature importance analysis to understand the impact of each variable on the predictions (Importance permutation).

5. Modeling:

   • Various machine learning models are evaluated, including:
     • XGBoost
     • LightGBM
     • Deep Learning Model with TensorFlow and Keras
   • Hyperparameter tuning is performed with Optuna for optimization.
   • Probability thresholding is applied to optimize classification outcomes, adjusting the threshold for each class to improve sensitivity or specificity depending on the model's predictions.

6. Model Evaluation:

   • Evaluation metrics such as accuracy and AUC are used.
   • ROC curve analysis to assess the model performance.

Requirements

Python (>= 3.7) Libraries: pandas, numpy, seaborn, matplotlib, scikit-learn, xgboost, lightgbm, tensorflow, optuna & keras_tuner