Deployed on Render 🔗 Calorie Expenditure Predictor App

Calorie Expenditure Prediction

This project predicts the number of calories burned during a workout session using physiological and activity-related features such as Age, Height, Weight, Duration, Heart Rate, and Body Temperature.

By leveraging machine learning models, the system identifies which factors most strongly influence calorie expenditure and selects the best-performing predictive algorithm for fitness planning and performance tracking.

The dataset, sourced from Kaggle’s Calories Burnt Prediction collection, provides the foundation for training a predictive system that supports personalised fitness insights and healthier lifestyle decisions.

Watch the Demo

Dataset Description

• Shape: 15000 rows, 7 columns
• Features:
  • User_ID: unique identifier for each individual
  • Gender: Male/Female
  • Age: individual's age (20 - 79 years)
  • Height: cm
  • Weight: kg
  • Duration: minutes
  • Heart_Rate: average heart rate during workout
  • Body_Temp: body temperature during workout
• Target
  • Calories

Exploratory Data Analysis (EDA) Insights

• Average participant: 174 cm, 75 kg, 96 bpm, 15.5 min workout, ~90 kcal burned.
• Strong correlations: Duration, Heart Rate, Body Temperature with Calories.
• Height & Weight correlation: 0.958 → engineered BMI feature.
• Gender differences: Males are slightly taller/heavier.
• Outliers in height/weight detected.

Feature Engineering and Selection

• Created bmi = (Weight / Height (m) ** 2)
• Selected Features:
  • X = Age, Gender, Duration, Heart_Rate, Body_Temp, bmi
  • y = Calories
• Drop identifier User_ID
• OneHotEncoding Gender as numeric
• Standscale the numeric features; Age, Duration, Heart_Rate, Body_Temp, bmi

Modelling - Training, Evaluation and Selection

Model	Metrics (MAE - RMSE - R²)	Remark
Ridge Regression	8.460 - 11.597 - 0.966	Model with regularization
Linear Regression	8.460 - 11.597 - 0.966	Baseline model
Decision Tree	3.964 - 6.003 - 0.991	Improved accuracy with slight overfitting
K Nearest Neighbor	4.237 - 5.842 - 0.991	Improved accuracy
Gradient Boosting	3.228 - 4.474 - 0.995	Best generalizing model
Random Forest	2.562 - 3.856 - 0.996	Best predictive accuracy

Selected Models: Random Forest Regressor and Gradient Boosting

Hyperparameter Tuning:

To further improve model performance, I applied GridSearchCV to the top two models: Random Forest and Gradient Boosting.

Model	Best Parameters	MAE	RMSE	R²
Random Forest	{'max_depth': 20, 'max_features': 'sqrt', 'min_samples_leaf': 1, 'min_samples_split': 2, 'n_estimators': 300}	2.835	4.413	0.995
Gradient Boosting	{'learning_rate': 0.05, 'max_depth': 5, 'max_features': 'sqrt', 'min_samples_leaf': 1, 'min_samples_split': 2, 'n_estimators': 300, 'subsample': 0.8}	2.491	3.500	0.997

Gradient Boosting achieved lower MAE and RMSE values with better generalisation, demonstrating superior performance and robustness. Final Model Selection: Tuned Gradient Boosting

Model Interpretation

To interpret how features influenced calorie predictions, two complementary explainability methods were applied — Permutation Importance and SHAP (SHapley Additive Explanations). These methods were used on the final Gradient Boosting model, the best-performing estimator after hyperparameter tuning.

Permutation Importance

Feature	Importance
Duration	1.0731
Heart_Rate	0.1691
Age	0.0563
Gender	0.0113
BMI	0.0016
Body_Temp	0.00005

Interpretation: Workout Duration and Heart Rate dominate calorie prediction. Physiological attributes (BMI, Age, Gender) contribute marginally.

Future Improvements

• Deploy advanced models (Polynomial Regression, Neural Networks).
• Improve UI with interactive charts & personalised recommendations.