Detecting Stress Levels from PPG Sensor Data using ANN #889

Detecting Stress Levels from PPG Sensor Data using ANN/Dataset/README.md

@@ -0,0 +1,12 @@
+# Dataset
+
+## Heart Rate Prediction to Monitor Stress Level
+
+- **Source:** Kaggle - [vinayakshanawad/heart-rate-prediction-to-monitor-stress-level](https://www.kaggle.com/datasets/vinayakshanawad/heart-rate-prediction-to-monitor-stress-level)
+- **Size:** 369,289 training samples, 3 CSV files merged on uuid
+- **Files used:**
+  - time_domain_features_train.csv - 19 HRV time-domain features
+  - frequency_domain_features_train.csv - 11 frequency-domain features
+  - heart_rate_non_linear_features_train.csv - 4 non-linear features + label
+
+The dataset file is too large to upload directly. Please download from the Kaggle link above.

Detecting Stress Levels from PPG Sensor Data using ANN/Model/README.md

@@ -0,0 +1,38 @@
+# Stress Level Detection from PPG Sensor Data — Model README
+
+## Problem Statement
+Multi-class classification of physiological stress levels (No Stress, Interruption, Time Pressure)
+using HRV features derived from PPG sensor data.
+
+## Dataset
+- Source: [Heart Rate Prediction to Monitor Stress Level](https://www.kaggle.com/datasets/vinayakshanawad/heart-rate-prediction-to-monitor-stress-level)
+- Train samples: 369,289 | Test samples: 73,858
+- Features: 34 HRV features (time-domain, frequency-domain, non-linear)
+- Classes: no stress (54.2%), interruption (28.5%), time pressure (17.3%)
+
+## Models Implemented
+
+| Model   | Test Accuracy | Notes |
+|---------|--------------|-------|
+| ANN     | 99.93%       | Baseline feedforward, 3 Dense layers |
+| 1D-CNN  | **99.99%**   | Best model, 2 Conv1D layers |
+| LSTM    | 99.90%       | Slowest training (~43s/epoch) |
+| GRU     | 99.94%       | Lighter than LSTM, competitive accuracy |
+
+## Best Model: 1D-CNN
+- Architecture: Conv1D(64) → MaxPool → Conv1D(32) → Flatten → Dense(64) → Dense(3)
+- Precision / Recall / F1: 1.00 across all 3 classes
+- Misclassifications: 10 out of 73,858 test samples
+
+## Preprocessing
+- Merged 3 feature CSV files on `uuid` key
+- Dropped `uuid` and `datasetId` columns
+- Label encoded target (`condition`)
+- StandardScaler normalization
+- 80/20 stratified train-test split
+- Reshaped to (samples, 34, 1) for sequence models
+
+## Conclusion
+1D-CNN achieves near-perfect classification while being 4x faster per epoch than LSTM/GRU.
+For tabular HRV features, local pattern detection via convolution outperforms recurrent approaches.
+ANN is a strong baseline given the well-separated feature space visible in the correlation heatmap.

Detecting Stress Levels from PPG Sensor Data using ANN/README.md

@@ -0,0 +1,151 @@
+# Detecting Stress Levels from PPG Sensor Data using ANN
+
+## 🎯 Aim
+
+The goal of this project is to predict stress levels using features derived from **Photoplethysmography (PPG) sensor data** by employing multiple deep learning architectures — ANN, 1D-CNN, LSTM, and GRU — and comparing their performance.
+
+---
+
+## 📂 Dataset
+
+- **Source:** [Heart Rate Prediction to Monitor Stress Level — Kaggle](https://www.kaggle.com/datasets/vinayakshanawad/heart-rate-prediction-to-monitor-stress-level)
+- **Training samples:** 369,289 | **Test samples:** 73,858
+- **Features:** 34 HRV features across three domains:
+  - **Time-domain** (19 features): MEAN_RR, SDRR, RMSSD, HR, pNN50, etc.
+  - **Frequency-domain** (11 features): VLF, LF, HF, LF_HF ratio, etc.
+  - **Non-linear** (4 features): SD1, SD2, sampen, higuci
+- **Target classes:** `no stress`, `interruption`, `time pressure`
+
+---
+
+## 📁 Project Structure
+
+```
+Detecting Stress Levels from PPG Sensor Data using ANN/
+│
+├── Dataset/
+│   └── README.md            ← Dataset source & description
+│
+├── Images/
+│   ├── class_distribution.png
+│   ├── correlation_heatmap.png
+│   ├── accuracy_loss_curves.png
+│   └── confusion_matrix.png
+│
+├── Model/
+│   ├── stress-ppg-detection.ipynb   ← Full pipeline notebook
+│   └── README.md                    ← Model details & results
+│
+├── Web App/
+│   ├── web_app.py           ← Streamlit web application
+│   ├── stress_ppg_cnn_model.h5
+│   ├── scaler.pkl
+│   └── README.md
+│
+├── README.md                ← (this file)
+└── requirements.txt
+```
+
+---
+
+## 🧠 Models Implemented
+
+| Model   | Architecture | Test Accuracy |
+|---------|-------------|--------------|
+| ANN     | 3× Dense layers with Dropout | 99.93% |
+| 1D-CNN  | 2× Conv1D → MaxPool → Dense | **99.99%** Best |
+| LSTM    | 1× LSTM(64) → Dense | 99.90% |
+| GRU     | 1× GRU(64) → Dense | 99.94% |
+
+### Why these models?
+
+| Model | Rationale |
+|-------|-----------|
+| **ANN** | Strong baseline for tabular HRV feature learning |
+| **1D-CNN** | Captures local temporal patterns and short-term signal variations |
+| **LSTM** | Handles long-range dependencies in sequential physiological data |
+| **GRU** | Lighter LSTM alternative — faster training, competitive accuracy |
+
+---
+
+## 📊 Exploratory Data Analysis
+
+### Class Distribution
+![Class Distribution](Images/class_distribution.png)
+
+### Feature Correlation Heatmap
+![Correlation Heatmap](Images/correlation_heatmap.png)
+
+### Training Accuracy & Loss Curves (1D-CNN)
+![Accuracy Loss Curves](Images/accuracy_loss_curves.png)
+
+### Confusion Matrix (1D-CNN)
+![Confusion Matrix](Images/confusion_matrix.png)
+
+---
+
+## ⚙️ Preprocessing Pipeline
+
+1. Merged three feature CSV files on `uuid` key
+2. Dropped `uuid` and `datasetId` identifier columns
+3. Label-encoded target (`condition`: 0 = Interruption, 1 = No Stress, 2 = Time Pressure)
+4. Applied `StandardScaler` normalization
+5. 80/20 stratified train-test split
+6. Reshaped to `(samples, 34, 1)` for sequence-based models (CNN, LSTM, GRU)
+
+---
+
+## 🏆 Results & Conclusion
+
+**1D-CNN** is the best-performing model with **99.99% test accuracy** (only 10 misclassifications out of 73,858 samples). It achieves near-perfect precision, recall, and F1-score across all three stress classes.
+
+Key findings:
+- HRV features are highly discriminative for stress classification — even a simple ANN achieves 99.93%
+- CNN's local pattern detection outperforms recurrent models for tabular HRV data
+- LSTM/GRU are 4× slower per epoch with marginal accuracy difference
+- 1D-CNN is the recommended architecture for deployment
+
+---
+
+## 🌐 Web Application
+
+A **Streamlit** web app allows real-time stress level prediction by entering HRV feature values manually.
+
+**To run:**
+```bash
+cd "Web App"
+streamlit run web_app.py
+```
+
+---
+
+## 📦 Libraries & Dependencies
+
+```
+numpy
+pandas
+matplotlib
+seaborn
+scikit-learn
+tensorflow
+streamlit
+joblib
+kaggle
+```
+
+Install with:
+```bash
+pip install -r requirements.txt
+```
+
+---
+
+## 👤 Author
+
+**Juned Pinjari**
+- GitHub: [@juned-pinjari](https://github.com/juned-pinjari)
+- Contribution: GSSoC 2026
+
+---
+
+*Part of the [DL-Simplified](https://github.com/abhisheks008/DL-Simplified) open-source repository.*

Detecting Stress Levels from PPG Sensor Data using ANN/Web App/README.md

@@ -0,0 +1,50 @@
+# Web App — Stress Level Detection from PPG Sensor Data
+
+## 🌐 Overview
+
+A **Streamlit** web application that predicts stress levels in real-time based on Heart Rate Variability (HRV) features derived from PPG sensor data. The app uses the trained **1D-CNN** model (best performer with 99.99% accuracy).
+
+## 🚀 How to Run
+
+1. **Install dependencies:**
+   ```bash
+   pip install streamlit tensorflow scikit-learn numpy pandas joblib
+   ```
+
+2. **Navigate to the Web App folder:**
+   ```bash
+   cd "Web App"
+   ```
+
+3. **Launch the app:**
+   ```bash
+   streamlit run web_app.py
+   ```
+
+4. Open your browser at `http://localhost:8501`
+
+## 🖥️ App Features
+
+- **34 HRV feature inputs** organized across three domains:
+  - Time-domain (19 features): MEAN_RR, SDRR, RMSSD, HR, pNN25/50, KURT, SKEW, etc.
+  - Frequency-domain (11 features): VLF, LF, HF, LF_HF, LF_NU, HF_NU, TP, etc.
+  - Non-linear (4 features): SD1, SD2, sampen, higuci
+- **One-click prediction** with the trained 1D-CNN model
+- **Confidence scores** displayed as progress bars for all three stress classes:
+  - 😊 No Stress
+  - 😐 Interruption
+  - 😰 Time Pressure
+
+## 📦 Files Required
+
+| File | Description |
+|------|-------------|
+| `web_app.py` | Main Streamlit application |
+| `stress_ppg_cnn_model.h5` | Trained 1D-CNN model (TF/Keras) |
+| `scaler.pkl` | Fitted StandardScaler for feature normalization |
+
+## 📝 Notes
+
+- The app loads the model and scaler with `@st.cache_resource` for efficiency
+- Features are scaled and reshaped to `(1, 34, 1)` before inference (matching the 1D-CNN input shape)
+- Default input values are set to typical resting-state HRV values for quick testing