🔥 Calorie Sense

AI-Powered Calorie Prediction & Adaptive Fitness Coaching

Smart Calories. Smarter Fitness.

---

Overview

Calorie Sense is a full-stack AI fitness platform built with Flask and TensorFlow. It does two things unusually well:

• 🌦️ Context-aware calorie prediction — an ANN predicts kcal burned from your exercise inputs plus live weather data fetched automatically from OpenWeatherMap. Cold and hot environments genuinely change how many calories you burn, and the model accounts for that.

• 🤖 Personalised AI coaching — a Gemini-powered fitness coach reads your health profile (BMI, hypertension, diabetes), matches you against a peer-reviewed recommendation dataset, and delivers structured workout plans and dietary guidance tailored specifically to you.

---

Screenshots

Home	Calorie Prediction	Prediction Result

AI Coach — Profile Setup	AI Coach — Session Start	AI Coach — Workout Plan

---

Table of Contents

• System Architecture
• Model Performance
• Datasets
  • What We Used and Why
  • Why We Dropped the Primary Dataset
  • The Target Variable Problem — and How We Fixed It
  • Why Weather Belongs in the Model
• AI Fitness Coach
• Project Structure
• Getting Started
• Environment Variables
• References

---

System Architecture

User Browser
    │
    ├── index.html       ← Landing page, system overview
    ├── calorie.html     ← Calorie prediction module
    └── AI.html          ← Gemini fitness coach
         │
         ▼
    Flask (app.py)
         │
         ├── /api/weather          ← ip-api.com → OpenWeatherMap → {city, temp, condition}
         ├── /api/predict_calorie  ← StandardScaler → ANN (.keras) → kcal
         ├── /api/start            ← Profile input → BMI → gym_lookup.json match
         └── /api/chat             ← Gemini API (gemini-flash) with grounded system prompt
              │
              ├── out/models/ann.keras          (production prediction model)
              ├── out/models/scaler.pkl         (fitted StandardScaler)
              ├── data/JSON/gym_lookup.json     (Mendeley recommendation lookup)
              └── data/JSON/exercises.json      (ExerciseDB exercise library)

---

Model Performance

All models were trained on the same 80/20 stratified split of the cleaned and fixed exercise dataset. The ANN was selected for production due to the best balance of test accuracy and train/test gap (generalisation).

Comparison Across All Models

Model	Train R²	Test R²	Δ R² (Gap)	Train MAE	Test MAE	Train RMSE	Test RMSE	Diagnosis
ANN ✅	0.9767	0.9757	0.0011	19.99	20.83	25.97	26.86	✅ Good fit
XGBoost (tuned)	0.9872	0.9742	0.0130	14.26	21.42	19.25	27.62	✅ Good fit
XGBoost	0.9820	0.9728	0.0092	17.69	22.58	22.86	28.38	✅ Good fit
RF (tuned)	0.9909	0.9697	0.0212	12.40	22.59	16.26	29.99	✅ Good fit
Random Forest	0.9884	0.9660	0.0224	14.05	24.89	18.35	31.74	✅ Good fit
Ridge (Poly)	0.9444	0.9428	0.0016	31.39	32.15	40.16	41.16	✅ Good fit

Why ANN over XGBoost (tuned) despite a marginally higher test MAE?
The XGBoost tuned model has a Δ R² gap of 0.013 vs. the ANN's 0.0011 — more than 10× larger. A smaller train/test gap means the ANN generalises more consistently to unseen data, which matters more in a live inference setting than a 0.6 kcal difference in MAE.

Feature Importance (ANN Input Features)

Features were selected by combining Random Forest importance scores with permutation importance, then keeping the top 7 by combined score.

Feature	RF Importance	Permutation Importance	Combined Score
Exercise Duration	1.000	1.000	1.000
Heart Rate	0.655	0.593	0.624
Gender	0.223	0.343	0.283
Max HR Percentage	0.129	0.084	0.107
Workload (Intensity × Duration)	0.084	0.084	0.084
Weather — Rainy	0.001	0.001	0.001
Weather — Sunny	0.000	0.000	0.000

Weather features rank low in importance scores because the Keytel formula (which generates the target) does not mathematically include weather — the weather multipliers are applied on top of it. The features are retained and meaningful at inference time; they simply explain less variance than physiological inputs by design. See Why Weather Belongs in the Model.

Training Diagnostic Plots

Click to expand training plots

Plot	Description
	ANN loss over epochs (train vs. validation)
	Actual vs. predicted calories + residual distribution
	R², MAE, RMSE comparison across all models
	Train/test R² scatter — generalisation check
	Overfitting / underfitting analysis
	Feature correlation heatmap (post-fix)
	Combined feature importance scores
	Target variable distribution after Keytel fix

---

Datasets

What We Used and Why

#	Dataset	Source	Role
2	Exercise & Fitness Metrics	Kaggle	ANN training data (target replaced with Keytel values)
3	Fitness Exercises (ExerciseDB)	Kaggle	Runtime exercise lookup for AI coach
4	Mendeley Gym Recommendation	Mendeley Data	Profile-to-plan matching for AI coach
5	OpenWeather API	openweathermap.org	Live weather context at inference time

---

Why We Dropped the Primary Dataset

The Gym Members Exercise Dataset was the original intended training source. After merging it with Dataset 2, two critical issues emerged:

• Incompatible calorie distributions — the two datasets had fundamentally different calorie ranges and distributions. Merging created a bimodal target that no single regression model can fit cleanly.
• ~80% null values — several key features (Max_BPM, Resting_BPM, Fat_Percentage, Water_Intake) became almost entirely missing after the merge, making imputation unreliable and the feature space inconsistent.

Dataset 1 was dropped entirely. Dataset 2 was used exclusively for model training after fixing its target variable.

---

The Target Variable Problem — and How We Fixed It

TL;DR: The original Calories_Burned column in exercise_dataset.csv was random noise. We replaced it with values from the Keytel et al. (2005) validated formula before any model training.

Statistical proof it was noise:

Diagnostic	Value	Interpretation
Max feature correlation	0.036	Effectively zero
Any model's R²	≈ −0.03	Worse than predicting the mean
KS test p-value	0.49	Indistinguishable from uniform random
All features' shared variance	0.9%	No predictable signal

A negative R² is not a pipeline bug — it is the mathematically guaranteed result when the target has no relationship to any input. No ML model can achieve R² > 0 on a random target.

The fix — Keytel et al. (2005):

The Calories_Burned column was replaced using the Keytel formula, the standard validated equation in exercise science, calibrated from real human trials and reproduced in wearable devices (Polar, Garmin):

Male:   Cal = Duration_min × (−55.0969 + 0.6309×HR + 0.1988×Weight + 0.2017×Age) / 4.184
Female: Cal = Duration_min × (−20.4022 + 0.4472×HR − 0.05741×Weight + 0.074×Age) / 4.184

Exercise Intensity is applied as an adjustment multiplier. Small Gaussian noise (σ ≈ 5%) is added to ensure the distribution reflects real-world variation rather than being a perfect deterministic output.

Before vs. after the fix:

Feature	Correlation Before	Correlation After
Exercise Duration	+0.02	+0.67
Heart Rate	−0.04	+0.59
Exercise Intensity	+0.01	+0.18

These correlations are physically meaningful and consistent with exercise physiology literature. The regenerated target is not fabricated — it uses the same formula that was validated across 115 subjects in the original study (see References).

---

Why Weather Belongs in the Model

After the Keytel fix, weather still has near-zero correlation with the target because the base formula doesn't include environmental factors. The weather multipliers are applied on top of the formula output to encode two well-documented physiological mechanisms:

❄️ Cold / Rainy → Thermogenesis
Your body burns extra calories to maintain core temperature (37°C). Wijers et al. (2008) measured a statistically significant +2.8% increase in 24-hour energy expenditure from mild cold exposure. Marlatt et al. (2019) found lean individuals can increase metabolic rate by up to +17% above basal before shivering begins.

☀️ Hot / Sunny → Cardiovascular Strain
The heart must simultaneously supply blood to working muscles and to the skin for heat dissipation. The National Academies of Sciences confirm muscular metabolism increases 5–15× during heat exercise. Girard et al. (2023) documented physiological strain increases of 5–20% depending on temperature and humidity.

The multipliers used in this project (5–8%) sit at the conservative lower bound of the ranges documented in the above studies — appropriate for general fitness exercise rather than extreme environmental conditions. The effect is U-shaped: both cold and hot conditions increase burn above a comfortable baseline (~18–22°C / Cloudy).

---

AI Fitness Coach

The coach module works in five steps:

1. Profile intake     → Name, sex, age, height, weight, hypertension, diabetes
2. BMI calculation    → Auto-computed; mapped to Underweight / Normal / Overweight / Obese
3. Plan lookup        → Profile matched against gym_lookup.json (from Mendeley dataset)
4. System prompt      → Built from profile + matched plans + full ExerciseDB library
5. Gemini chat        → Free-form Q&A grounded in the user's specific health data

Health condition flags (Hypertension, Diabetes) trigger mandatory safety reminders in every response. The coach explicitly defers to a doctor for any clinical concerns and will never give unsafe medical advice.

---

Project Structure

caloriesense/
│
├── gui/
│   ├── app.py                    # Flask app — routes, ANN inference, Gemini API
│   ├── templates/
│   │   ├── index.html            # Landing page
│   │   ├── calorie.html          # Calorie prediction UI
│   │   └── ai.html               # AI fitness coach UI
│   └── static/
│       ├── style.css             # Dark-theme glassmorphism styles
│       └── script.js             # Frontend logic, weather fetch, prediction calls
│
├── data/
│   ├── Calorie Prediction/
│   │   └── exercise_dataset.csv  # Dataset 2 (Calories_Burned replaced with Keytel values)
│   ├── Recomendation/
│   │   ├── Fitness Exercises Dataset/
│   │   └── Mendeley Gym Recommendation Dataset/
│   └── JSON/
│       ├── gym_lookup.json       # Pre-processed Mendeley lookup (keyed by health profile)
│       └── exercises.json        # Pre-processed ExerciseDB library
│
├── out/
│   ├── models/
│   │   ├── ann.keras             # Production ANN (TensorFlow SavedModel)
│   │   ├── scaler.pkl            # StandardScaler fitted on training data
│   │   ├── feature_list.pkl      # Ordered feature names for inference
│   │   ├── pipeline_metadata.json
│   │   ├── train_test_accuracy_summary.csv
│   │   ├── rf.pkl / rf_tuned.pkl
│   │   ├── xgb.pkl / xgb_tuned.pkl
│   │   └── ridge_poly.pkl
│   └── *.png                     # Training diagnostic plots
│
├── retrain.ipynb                 # Training notebook v1
├── retrainV2.ipynb               # Training notebook v2
├── retrainV3.ipynb               # Training notebook v3 (final pipeline)
├── json_convert.py               # Converts raw CSVs → gym_lookup.json / exercises.json
├── location_based_weather.py     # Standalone weather utility (ip-api → OpenWeatherMap)
├── .env                          # API keys (never committed)
└── docs/
    └── Images/                   # UI screenshots

---

Getting Started

Prerequisites

• Python 3.10+
• A Gemini API key
• An OpenWeather API key

Installation

# Clone the repository
git clone https://github.com/heyisula/caloriesense.git
cd caloriesense

# Install dependencies
pip install flask tensorflow scikit-learn xgboost joblib requests python-dotenv google-genai numpy

# Configure API keys (see Environment Variables below)

Running

cd gui
python app.py

Open http://127.0.0.1:5000 in your browser.

Note: The trained model artefacts in out/models/ must be present. If retraining from scratch, run retrainV3.ipynb end-to-end first.

---

Environment Variables

Create a .env file in the project root (not inside gui/):

GEMINI_API_KEY=your_gemini_api_key_here
GEMINI_MODEL=gemini-2.0-flash-lite
OPENWEATHER_API_KEY=your_openweather_api_key_here

The .env file is listed in .gitignore and will never be committed to version control.

---

References

1. Keytel formula — calorie estimation from heart rate
   Keytel, L.R., Goedecke, J.H., Noakes, T.D., Hiiloskorpi, H., Laukkanen, R., van der Merwe, L., & Lambert, E.V. (2005). Prediction of energy expenditure from heart rate monitoring during submaximal exercise. Journal of Sports Sciences, 23(3), 289–297.
   PubMed · Semantic Scholar · ResearchGate
   Validated on 115 subjects, 18–45 years, 47–120 kg across three exercise intensities (35%, 62%, 80% VO₂max).

2. Cold-induced thermogenesis — cold/rainy weather increases calorie burn
   Wijers, S.L.J., Saris, W.H.M., & van Marken Lichtenbelt, W.D. (2008). Human skeletal muscle mitochondrial uncoupling is associated with cold induced adaptive thermogenesis. PLoS ONE, 4(3).
   PMC free text

   Marlatt, K.L., et al. (2019). Quantification of the capacity for cold-induced thermogenesis in young men with and without obesity. Journal of Clinical Endocrinology & Metabolism.
   PMC free text

3. Cardiovascular strain — hot/sunny weather increases calorie burn
   National Academies of Sciences. Nutritional Needs in Hot Environments — Physiological Responses to Exercise in the Heat.
   NCBI Bookshelf

   Girard, O., et al. (2023). Delineating the impacts of air temperature and humidity for endurance exercise. Experimental Physiology.
   PMC free text

---

License

MIT © heyisula