Health Sensing: Breathing Irregularity Detection

Sleep study analysis pipeline: visualize overnight signals, build labeled 30-second windows, and train a 1D CNN to classify breathing irregularities (e.g. Hypopnea, Obstructive Apnea) using Leave-One-Participant-Out cross-validation.

Project Structure

Data/                 # Raw data per participant (AP01–AP05)
  AP01/
    Flow - *.txt, Thorac - *.txt, SPO2 - *.txt
    Flow Events - *.txt, Sleep profile - *.txt
  AP02/ ... AP05/

Visualizations/       # PDFs from vis.py (one per participant)
Dataset/              # breathing_dataset.csv/.npz, sleep_stage_dataset.csv/.npz
models/               # cnn_model.py (1D CNN)
scripts/
  data_utils.py       # Load signals, flow events, sleep profile
  vis.py              # 8-hour plots with event overlays
  create_dataset.py   # Filter, window, label → Dataset/
  train_model.py      # LOPO-CV training and metrics
requirements.txt
README.md
report.md

Setup

From the project root:

pip install -r requirements.txt

Requirements: numpy, pandas, matplotlib, scipy, scikit-learn, torch.

Data

Download the dataset of participants (raw signals and annotations for AP01–AP05) and extract it so that the Data/ folder contains one subfolder per participant (e.g. Data/AP01/, Data/AP02/, …).

Download participant dataset (internship.zip) — extract the contents into the project’s Data/ directory.

Usage

Run all steps from the project root (E:\assingment or your repo root).

1. Visualizations (Understanding the Data)

Plot Nasal Airflow, Thoracic Movement, and SpO₂ for the full 8-hour recording with annotated breathing events overlaid. Output is saved as PDF in Visualizations/.

python scripts/vis.py -name "Data/AP01"

Generate for all participants:

python scripts/vis.py -name "Data/AP01"
python scripts/vis.py -name "Data/AP02"
python scripts/vis.py -name "Data/AP03"
python scripts/vis.py -name "Data/AP04"
python scripts/vis.py -name "Data/AP05"

2. Dataset Creation (Preprocessing)

Reads signals and annotations from Data/, applies bandpass filter (0.17–0.4 Hz), splits into 30-second windows with 50% overlap, and labels windows from the flow events file (>50% overlap → event label, else Normal). Writes breathing_dataset and sleep_stage_dataset to Dataset/.

python scripts/create_dataset.py -in_dir "Data" -out_dir "Dataset"

3. Training and Evaluation (1D CNN + LOPO-CV)

Trains a 1D CNN on the breathing dataset using Leave-One-Participant-Out cross-validation and reports Accuracy, Precision, Recall, and Confusion Matrix.

python scripts/train_model.py

Uses Dataset/breathing_dataset.npz by default. Run create_dataset.py first if the dataset is missing.

Data Notes

• Signals: Nasal Airflow and Thoracic Movement at 32 Hz; SpO₂ at 4 Hz. All aligned by timestamp.
• Events: Flow events file lists Hypopnea, Obstructive Apnea, etc., with start/end times (end time may be time-only; the loader infers the date from the start time).
• Labels: Classes include Normal, Hypopnea, Obstructive Apnea, and may include Body event, Mixed Apnea depending on annotations.

Outputs

Output	Description
Visualizations/<ID>_visualization.pdf	8-hour signals + event overlays per participant
Dataset/breathing_dataset.csv	One row per 30s window: participant_id, label, features (f0–f2879)
Dataset/breathing_dataset.npz	X, y, participant_ids for training
Dataset/sleep_stage_dataset.csv / .npz	Same windows labeled by sleep stage from the sleep profile

Training prints per-fold and overall Accuracy, Precision, Recall, and Confusion Matrix to the console.