Early Prediction of Cerebral Vasospasm After Aneurysmal Subarachnoid Hemorrhage

This repository contains the dataset and source code accompanying the manuscript:

Early Prediction of Cerebral Vasospasm After Aneurysmal SAH Using a Machine Learning Model and Interactive Web Application

Overview

Cerebral vasospasm is one of the most severe complications following aneurysmal subarachnoid hemorrhage (aSAH). Early prediction is critical for timely interventions and improved outcomes. In this work, we developed and validated several machine learning models to predict vasospasm using routinely collected clinical and radiological variables.

Our best-performing model was logistic regression, which provided the most favorable trade-off between accuracy, sensitivity, specificity, and interpretability.

An interactive web application was implemented with Gradio and deployed on HuggingFace. The deployed model was trained on the entire dataset for maximum predictive accuracy in real-time use.

Both the dataset and source code are openly available to support transparency and reproducibility.

Dataset

The published dataset is provided in CSV format:

data/dataset_vasospasm.csv

Features

Feature	Description
id	Patient identifier (ascending integers )
age_diagnose	Age at diagnosis (in years)
sex	Patient sex (male / female)
hunt_hess	Hunt and Hess grade at admission (clinical severity scale)
fisher	Fisher scale (radiological blood load grading)
cns_infection	Presence of central nervous system infection (0 = no, 1 = yes)
EVD	External ventricular drainage placement (0 = no, 1 = yes)
aneurysm	Presence of aneurysm (0 = no, 1 = yes)
aneurysm_location	Anatomical location of the aneurysm (e.g., ACom, ACA, MCA)
aneurysm_diameter	Maximum aneurysm diameter (mm)
aneurysm_height	Aneurysm dome height (mm)
aneurysm_treatment	Treatment modality (e.g., coil, clip)
vasospasm	Binary target variable: occurrence of vasospasm (0 = no, 1 = yes)
vasospasm_severity	Severity of vasospasm (moderate, severe, or empty if none) – not used

Local Modelling (Python)

1. Clone this repository

   git clone https://github.com/RISCSoftware/vasospasm-prediction.git
   cd vasospasm-prediction

2. Set up the environment You can use either Poetry (recommended) or pip.

   • Poetry (reproducible) First install Poetry if it’s not already available. Then run:

     poetry install
     poetry env activate

     This creates and activates a virtual environment with exact dependency versions from the poetry.lock file.

   • Pip (lightweight) First create and activate a virtual environment, e.g.

     python -m venv .venv && source .venv/bin/activate
     

     Then run:

     pip install -r requirements.txt

     This installs dependencies directly; quicker but versions may vary slightly.

3. Run analyses The main entry point for training and evaluation is: vasospasm/main.py

   This script trains the classifiers using nested cross-validation and saves results to output/vasospasm/current/:

   • report.csv: Performance metrics for predictive models for vasospasm.
   • feature_importance.json: Feature importance, coefficients, and odds ratios for logistic regression.

R Environment & Statistical Tables

Generate the baseline characteristics table (with statistical tests) from the paper using R + renv and Quarto.

Prerequisites

• R (≥ 4.2 recommended)
• Quarto CLI
• From the repo root, the dataset is at data/dataset_vasospasm.csv

1) Create/restore the R environment Run the setup script once from the repository root:

Rscript renv-setup.R

This installs/activates renv, restores packages from renv.lock if present, or installs a minimal set and snapshots to create renv.lock.

2) Render the statistics table Render the Quarto document that computes summary statistics and p-values:

quarto render vasospasm/statistics_table.qmd

To restrict the analysis to aneurysmal SAH only, set only_with_aneurysm <- TRUE near the top of vasospasm/statistics_table.qmd before rendering.

Outputs

• An HTML render of the document (default Quarto output).

• A Word table is written by the script to:

  • output/table_1_all.docx (default, all patients)
  • or output/table_1.docx if the document is configured to only include aneurysmal cases.

Web Application

For interactive use, access the deployed HuggingFace space: 🔗 Web App on HuggingFace Spaces

You can also run the Gradio web app locally by running vasospasm/web/gradio_app.py.

License

• The source code of this project is licensed under the MIT License.
• The dataset provided in data/dataset_vasospasm.csv is licensed under the Creative Commons Attribution 4.0 International License. See the data/LICENSE-DATA.txt file for more details.

Citation

If you use this dataset or code, please cite the following paper:

Gollwitzer M, Mazanec V, Steindl M, Atli B, Stroh N, Hauser A, Sardi G, Rossmann T, Aspalter S, Rauch P, Horner E, Sonnberger M, Gruber A, Gmeiner M. Early Prediction of Cerebral Vasospasm After Aneurysmal SAH Using a Machine Learning Model and Interactive Web Application. Brain Sci. 2025.