PICU-Criticality-Prognosis

By: Derek Newman

EEG Criticality and Complexity machine learning analyses for pediatric ICU (PICU) prognosis of functional outcome/recovery.

This repository contains preprocessing, feature extraction, statistical comparison, and classification workflows focused on criticality-informed EEG features.

Overview

The analysis pipeline includes:

• EEG preprocessing.
• Feature extraction across entropy, complexity, fractal, DFA, and spectral power.
• Group-difference and normality analyses.
• Outcome classification with bootstrap-averaged logistic regression and comparative models.
• Visualization scripts and figures.

Repository Layout

PICU_Criticality/
├── figures/                          # Generated figures and comparison tables
├── notebooks/
│   ├── Demos_and_descriptive_statistics.ipynb # cohort and descriptive summaries
│   ├── Group_differences.ipynb                # Recovery, etiology and drug distributed based group analyses.
│   ├── ML_Classification_Bootstrap.ipynb      # bootstrap model training/evaluation.
│   └── Ordinal_analysis.ipynb                 # Data driven identification of group behvaioural recovery score
├── preprocessing/
│   └── Preprocessing_10_ICA.ipynb    # preprocessing workflow
├── scripts/
│   └── feature_extraction/
│       ├── neural_complexity.py       # CLI entrypoint
│       ├── complexity.py              # Main orchestration function
│       ├── feature_analysis.py        # Metric extraction methods
│       ├── parameter_selection.py     # Delay/dimension optimization
│       └── visualize.py               # Diagnostic and summary plots
├── source/                            
├── requirements.txt                   #  Dependencies
└── .gitignore                         # excludes data and source directories.

Installation

1. Create and activate a Python environment.
2. Install dependencies.

python3 -m venv .venv
source .venv/bin/activate
pip install -r requirements.txt

Data Expectations

• The feature extraction CLI accepts either:
  • A path to an MNE epochs file readable by mne.read_epochs(...).
  • An in-memory mne.Epochs object.
• Typical shape assumptions in the pipeline are (n_epochs, n_channels, n_times).

Main Feature Extraction Pipeline

CLI entrypoint:

python scripts/feature_extraction/neural_complexity.py /path/to/epochs-epo.fif \
	--save \
	--out_dir ./ANALYSIS \
	--condition baseline \
	--inspection \
	--entropy \
	--complexity \
	--fractal \
	--dfa \
	--power \
	--PDF

Key behaviors of Complexity_Feature_Extraction(...):

• Optional delay optimization (Fraser method) and embedding dimension optimization (Cao method).
• Extraction families:
  • Entropy metrics (sample entropy, permutation variants, etc.)
  • Complexity metrics (Lempel-Ziv variants, Lyapunov, Hjorth, etc.)
  • Fractal metrics (Higuchi FD, Katz FD, Petrosian FD, etc.)
  • DFA-based metrics
  • Power and spectral metrics
• Diagnostics:
  • Channel traces
  • PSD visualizations
  • Per-metric temporal/channel summaries
  • Heatmaps and correlation maps across metrics

Outputs:

• Pickle file (default): complex_results.pkl
• Condition-specific pickle: {condition}_complex_results.pkl
• PDF report: complexity_analysis_{condition}.pdf (or complexity_analysis.pdf)