CalorieMatrix : Predictive Modeling of Calorie Burn Prediction

Personal calorie burn estimation during physical exercise still relies heavily on manual MET-based calculations, which are static, non-personalized, and often inconsistent. Many people today use smartwatches and fitness applications while exercising. These applications usually show the number of calories burned during a workout. However, the values are not always the same because calorie burn depends on several personal factors such as body weight, heart rate, age, and workout duration. While working on this project, I observed that traditional calorie calculation methods often give general estimates and may not represent the actual calories burned by different individuals. To overcome this problem, I developed a system named CalorieMatrix. For building the model, workout records containing personal and exercise-related information were used. Different approaches including LightGBM, TabTransformer, and ACP were tested to check which one could estimate calories more accurately. After comparing their performance, the outputs were combined to form a single ensemble model. The final model produced better results than the individual models. During testing, it achieved an R² score of 94.62%, MAE of 62.23 kcal, and RMSE of 85.05 kcal. It was also observed that workout duration, average heart rate, and body weight had a major influence on calorie expenditure. The system can be used in fitness applications and wearable devices to provide calorie estimates during physical activities. The work shows that data-driven prediction methods can provide more consistent results than traditional calculation techniques.