SleepSense AI

Scoring overnight breathing events by hand is slow specialist work, and the events that matter are rare: in this recording set obstructive apnea accounts for 36 of 1318 windows.

• Parsed the raw signal exports and clinical event annotations, aligning nasal airflow, thoracic movement and SpO2 on their timestamps.
• Applied a Butterworth bandpass at 0.17 to 0.4 Hz with SciPy to isolate the breathing band before segmentation.
• Segmented the recordings into 30-second windows at 50% overlap and labelled each one from the annotated event intervals.
• Trained a 1D CNN, three Conv1d and BatchNorm blocks into adaptive average pooling with dropout, on 3-channel 960-sample windows in PyTorch.
• Trained a Conv-LSTM alternative, two convolutional blocks feeding a single-layer LSTM, on the same windows.
• Evaluated with leave-one-participant-out cross-validation so that no participant appears in both training and test folds.

1D CNN

Three Conv1d and BatchNorm blocks, adaptive average pooling, dropout, then a linear classifier.

Conv-LSTM

Two convolutional blocks feeding a single-layer LSTM, classified from the final timestep.

Butterworth bandpass filter

Isolates the 0.17 to 0.4 Hz breathing band before windowing.

Overnight recordings from 5 participants: nasal airflow, thoracic movement and SpO2. The processed set is 1318 windows, split 886 normal, 396 hypopnea and 36 obstructive apnea.

Both models land near 0.53 accuracy under leave-one-participant-out validation. With 5 participants and 36 apnea examples that is the honest result, and the value of the project is the pipeline and the evaluation protocol rather than the classifier.

• 1D CNN, leave-one-participant-out: accuracy 0.53, precision 0.30, recall 0.33
• Conv-LSTM, leave-one-participant-out: accuracy 0.53, precision 0.33, recall 0.31
• Dataset: 1318 windows from 5 participants (886 / 396 / 36 per class)

These numbers come from held-out-participant validation, which is deliberately harder than a random window split. A random split would leak windows from the same night into both sides and report a much better number that would not survive a new participant.

• Evaluated leave-one-participant-out rather than with a random split, because a random split leaks windows from the same participant into training and test and inflates the score.
• Filtered to 0.17 to 0.4 Hz before windowing to keep the breathing band and drop baseline drift and higher-frequency noise.