Machine Learning Pipeline for SpO₂ Prediction -- 2

The goal is to develop and validate a machine learning pipeline that estimates SpO₂ directly from PPG signals.

Tasks

Data Handling

Read and preprocess .dat/.hea waveform files (PPG signals, 86 Hz).

Align PPG epochs with ground truth values (ABG SaO₂ and device SpO₂).

Implement signal quality checks (SQI) to discard bad/noisy segments.

Feature Extraction

Extract classical features (AC/DC ratios, R-values, waveform morphology, energy, etc.).

Explore deep learning approaches (feedforward NN, CNN, LSTM) using raw epochs.

Model Development

Build ML models (SVM, Random Forest, Gradient Boosting, etc.) for SpO₂ estimation.

Build NN models (1D CNN/LSTM) that take raw PPG epochs as input.

Train/test split: 80/20 with cross-validation.

Validation & Reporting

Compare predicted SpO₂ with ground truth (SaO₂ from ABG).

Report accuracy in MAE, RMSE, Bias, and ARMS, following ISO 80601-2-61 standards.

Provide bin-wise evaluation (90–94, 95–96, 97–100%) and Bland–Altman plots.

Deliverables

Clean, well-documented MATLAB or Python code.

Final trained models (ML + NN).

Validation report with performance metrics and plots.