Title
AI-PARI: AI-Enhanced Diagnostics for Pediatric Asthma and Respiratory Irregularities Using Deep Learning and Wearable Sensors

Description
This project implements a deep learning-based system for detecting pediatric asthma and respiratory irregularities using simulated wearable sensor data. The system combines Convolutional Neural Networks (CNNs) and Recurrent Neural Networks (RNNs) to analyze respiratory signals in real-time. Federated learning is incorporated to ensure data privacy while enabling personalized model training across multiple devices.

Dataset Information
The code simulates sensor data for respiratory monitoring, including:

Respiratory Rate (RR)

Oxygen Saturation (SpO2)

Airflow patterns

Heart Rate Variability (HRV)

For real-world implementation, datasets like the Pulmonary Sound Dataset on Kaggle can be used, which contains labeled audio recordings of breathing patterns associated with various lung conditions.

Code Information
The implementation includes:

Sensor data simulation/generation

Signal preprocessing with low-pass filtering

Hybrid CNN-RNN model architecture

Cloud-based predictive analytics

Federated learning simulation

Performance evaluation metrics

Key files:

code.ipynb: Main implementation notebook

Usage Instructions
Install required dependencies (see Requirements section)

Run the Jupyter notebook code.ipynb

The notebook will:

Simulate sensor data

Preprocess the signals

Train the initial model

Simulate federated learning

Evaluate model performance

To use with real data:

Replace the simulate_sensor_data() function with your data loading function

Ensure your data is formatted as (samples, timesteps, features)

Requirements
Python 3.7+

TensorFlow 2.x

NumPy

pandas

scikit-learn

SciPy

Jupyter Notebook

Install with:

pip install tensorflow numpy pandas scikit-learn scipy jupyter
Methodology
Data Generation: Simulates multi-sensor respiratory data

Preprocessing: Applies Butterworth low-pass filter to remove noise

Model Architecture:

CNN layers for spatial feature extraction

LSTM layers for temporal pattern recognition

Dense layers for classification

Federated Learning:

Splits data across simulated clients

Performs local training

Aggregates model weights

Evaluation:

Accuracy, Precision, Recall, F1-score

Classification report

License
This project is provided for research purposes. For specific licensing or collaboration inquiries, please contact the corresponding author Xuewen Qiu at qxw4541@163.com.