FitSense-XAI: Predicting Stress from Everyday Habits

This is my final-year project where I built a machine learning model that predicts stress levels from daily habits like sleep, physical activity, screen time, and diet.

Unlike most black-box models, this one explains why it made a prediction using SHAP (SHapley Additive exPlanations). The goal is to make AI more transparent and useful in real life.

---

🔗 Live Demo

👉 Try the app here: (https://fitsense-xai.streamlit.app)

Enter your own numbers and see what the model predicts for you.

---

📊 What's Inside

Datasets Used

• Wearable (10,000 records): Sleep, activity, heart rate, stress
• Student (25,500 records): Academic pressure, sleep, social media
• Lifestyle (1,200 records): Diet, mood, physical activity

Models Trained

Model	Wearable	Student	Lifestyle
Logistic Regression	70.3%	79.1%	97.9%
Random Forest	85.9%	86.8%	96.3%
XGBoost	83.2%	85.9%	97.1%

Best overall: Random Forest (Wearable & Student), Logistic Regression (Lifestyle)

---

🧠 Why SHAP?

Most ML projects just give you a number — "your stress level is high". But I wanted to show what factors are driving that prediction.

SHAP helps the app tell you things like:

• "Your sleep quality is below average"
• "Your screen time is higher than usual"
• "You're not physically active enough"

This makes the model more trustworthy and useful in real life.

---

🛠️ Tech Stack

• Python (Pandas, NumPy)
• Scikit-learn (Random Forest, Logistic Regression)
• XGBoost
• SMOTE (for imbalanced data)
• SHAP (for explainability)
• Streamlit (for web app)
• Matplotlib & Seaborn (for visualizations)

---

🚀 How to Run Locally

git clone https://github.com/tahsin1017/FitSense-XAI.git
cd FitSense-XAI
conda env create -f environment.yml
conda activate fitsense-env
streamlit run app/app.py
📁 Project Structure
FitSense-XAI/
├── app/                # Streamlit app
├── data/raw/           # Datasets
├── notebooks/          # Jupyter notebooks (EDA + training)
├── models/             # Saved models & SHAP explainers
├── reports/figures/    # Visualizations & SHAP plots
└── docs/               # Research journal (daily progress)
📌 What I Learned

Working with multiple datasets is messy but teaches you a lot about generalization.
SMOTE actually helps when your data is imbalanced.
SHAP makes ML models more usable and transparent.
Deploying a model as a web app changes how you think about the whole pipeline.
📄 License

MIT License — free for learning and research.

👤 Author

Tahsin Ahmed Rafi
Final Year CSE | Data Science & AI/ML
GitHub