Blood Pressure Monitoring from PPG Signals

This project provides a machine learning pipeline to estimate Systolic (SBP) and Diastolic (DBP) blood pressure from Photoplethysmogram (PPG) signals extracted from smartphone videos. It implements both a feature-engineered Random Forest (RF) baseline and a deep learning 1D Convolutional Neural Network (CNN).

Quick Start

Prerequisites

• Python 3.12+
• NVIDIA GPU (Recommended for CNN training and inference)
• uv (Recommended for dependency management)

Installation

Clone the repository and install dependencies using uv:

uv pip install -r requirements.txt

Training

Train the models on the UCI dataset parts. The script auto-detects available hardware (CUDA/MPS/CPU).

# Train Random Forest (CPU-based)
uv run python train.py --model rf --parts 1 2 3 4

# Train 1D CNN (Hardware-accelerated)
uv run python train.py --model cnn --parts 1 2 3 4 --epochs 100

Prediction

Predict blood pressure from a recorded video:

# Default prediction using Random Forest
uv run python predict.py your_video.mp4

# Prediction using CNN (Recommended for accuracy)
uv run python predict.py your_video.mp4 --model cnn

Model Performance

The models were evaluated on the UCI Cuff-less Blood Pressure Estimation dataset.

Metric	Random Forest	1D CNN
SBP MAE	13.91 mmHg	7.50 mmHg
DBP MAE	6.67 mmHg	4.18 mmHg
SBP R2	0.24	0.70
DBP R2	0.56	0.74
Hardware	CPU	GPU (CUDA/MPS)

Calibration System

The system includes a calibration mechanism to personalize predictions using reference measurements from clinical devices. This accounts for individual physiological variations and model biases.

Calibration Commands

# Calibrate model with a reference BP reading
uv run python predict.py video.mp4 --calibrate --sbp 120 --dbp 80 --model cnn

# View current calibration status
uv run python predict.py --show-calibration

# Clear all stored calibration data
uv run python predict.py --clear-calibration

Technical Specifications

1D CNN Architecture

The CNN model processes raw PPG signals through a series of convolutional layers designed for temporal feature extraction:

• Input: (1, 625) normalized PPG signal
• Layers:
  • Conv1D (32 filters, kernel=7) + BatchNorm + ReLU + MaxPool
  • Conv1D (64 filters, kernel=5) + BatchNorm + ReLU + MaxPool
  • Conv1D (128 filters, kernel=3) + BatchNorm + ReLU + MaxPool
  • Conv1D (128 filters, kernel=3) + BatchNorm + ReLU
  • GlobalAveragePooling1D
  • Fully Connected (128) + ReLU + Dropout (0.3)
  • Fully Connected (64) + ReLU + Dropout (0.3)
  • Output: Fully Connected (2) -> [SBP, DBP]
• Parameters: ~110K trainable weights

Signal Processing Pipeline

• Windowing: 5-second segments at 125 Hz (625 samples per window)
• Filtering: Butterworth Bandpass filter (0.5–8.0 Hz)
• Normalization: StandardScaler applied per signal (CNN) or per feature (RF)
• Video Processing: Automatic color channel selection based on signal-to-noise ratio; resampling from camera FPS to 125 Hz.

Cross-Platform Support

The project utilizes PyTorch for the CNN implementation and automatically detects the best available backend:

• CUDA: NVIDIA GPUs (via cuda)
• MPS: Apple Silicon GPUs (via mps)
• CPU: Standard fallback for all systems

Project Structure

• src/: Core implementation modules
  • model_cnn.py: CNN architecture and inference logic
  • model_trainer.py: Random Forest training orchestration
  • model_trainer_cnn.py: CNN training orchestration
  • data_loader.py: UCI dataset parsing and windowing
  • features.py: PPG feature extraction using NeuroKit2
  • calibration.py: Persistent calibration management
  • inference_video.py: End-to-end video-to-BP pipeline
• model/: Directory for saved model weights and metadata
• uci_dataset/: Target directory for .mat data files
• train.py: CLI for model training
• predict.py: CLI for running inference on video files

References

• UCI Cuff-less Blood Pressure Estimation Dataset: Archive Link
• Signal analysis facilitated by NeuroKit2