VE-ICANS

Code for VE-ICANS — a Visual-EEG-feature-based score for the severity of Immune-effector-Cell-Associated Neurotoxicity Syndrome (ICANS), the neurotoxicity that can follow CAR-T cell therapy. Sibling of the VE-CAM-S delirium-severity score.

What it does

Given expert-rated visual EEG features (delta / theta / alpha frequency content, plus other image-level features) extracted from clinical EEG recordings of CAR-T patients, fit a small, interpretable ordinal model that predicts the patient's ICANS grade (0–4). The model is constrained to produce a clinically usable integer / half-integer score — coefficients are restricted to a fixed decimal grid (e.g. 0.1 or 0.5 increments), can be bounded and ordered, and can be constrained to sum to a target total. This is what turns a logistic-regression-style fit into something that looks like a clinical scoring rubric.

Layout

prepare_new_features.py        clean / curate the input feature spreadsheet
                               (drop out-of-range or "asleep" images,
                               extract numeric ICANS grade from caption,
                               binarise frequency-band features)
fit_model.py                   main modelling code. Defines:
                                 - MyLogisticRegression — logistic regression with
                                   coefficient-decimal / sum / bound / order constraints
                                 - MyCalibrator — ordinal recalibration wrapper around
                                   mord.LogisticAT
                               Plus the cross-validation / hyperparameter search driver.
AUCgraph_new.py                bootstrap ROC + calibration curves
table_info.py                  Table-1 patient summary statistics
                               (median/IQR, p-values via Mann-Whitney)
create_image_captions.py       helper for generating image captions

Required environment

• Python 3 with numpy, pandas, scipy, scikit-learn, mord (ordinal logistic regression), matplotlib, seaborn, tqdm, openpyxl (Excel I/O).
• Input feature workbook (NewFeatures.xlsx) and clinical-data spreadsheets organised as in the original analysis (see the absolute paths in table_info.py for the layout).

Related repos

• bdsp-core/VE-CAM-S — visual-EEG-feature-based delirium-severity score; same constrained- coefficient modelling approach.
• bdsp-core/E-ICANS — fully-automated EEG-based ICANS severity prediction via learning-to-rank (no human EEG-feature grading).

Status

Research code accompanying the VE-ICANS analysis (2021–2022).