Binary prediction of DNS using brain MRI scans and clinical variables with machine learning models

• Python version 3.10
• To install the required Python libraries: pip install -r requirements.txt

• Image preprocessing: HD-BET, SPM12 on MATLAB R2021a

• 🔗 ISLES22-model-inference
• Postprocessing to extract four variables from the predicted lesion masks: get_lesion_counts_volumes.ipynb

• We used an in-house MATLAB code running on MATLAB R2021a, which is not publicly available.
• The list of radiomics variables are provided: variables_type_info.xlsx (sheet_name = 'radiomics')
• Please reach out to us via email if you need access to this code.
• Alternatively, this step can be performed using the PyRadiomics package.

/your_project/
├── csv_files_for_training_testing/
|   ├── b1000_coreg_corrected_feature_lobe_1.csv/
|   ├── b1000_coreg_corrected_feature_lobe_2.csv/
|   ├── b1000_coreg_corrected_feature_lobe_3.csv/
|   ├── b1000_coreg_corrected_feature_lobe_4.csv/
|   ├── b1000_coreg_corrected_feature_lobe_5.csv/
|   ├── b1000_coreg_corrected_feature_lobe_6.csv/
|   ├── b1000_coreg_corrected_feature_lobe_7.csv/
|   ├── b1000_coreg_corrected_feature_lobe_8.csv/
|   ├── b1000_coreg_corrected_feature_lobe_9.csv/
|   ├── b1000_coreg_corrected_feature_lobe_10.csv/
|   ├── b1000_coreg_corrected_feature_lobe_11.csv/
|   ├── b1000_coreg_corrected_feature_lobe_12.csv/
|   ├── clinical_features_long.csv/
|   ├── clinical_features_short.csv/
|   ├── DNS_binary_label.csv/
|   ├── lesion_features.csv/
|   └── manual_image_features.csv/
└── variables_type_info.xlsx/

• All .csv files in the csv_files_for_training_testing directory follow the same data format.
• Each row corresponds to a feature, and each column corresponds to a patient, without any row or column headers included.
• The feature order follows that listed in the variables_type_info.xlsx file.

• The clinical data utilized in this study are not publicly accessible due to patient privacy concerns.
• Requests to access the data may be considered upon contact and IRB approval. Please contact us via email for inquiries.
• The imaging features, including radiomics, lesion segmentation, and manually labeled features, are provided.

• Please refer to machine_learning_run.ipynb

• Lee GY, Sohn CH, Kim D, Jeon SB, Yun J, Ham S, Nam Y, Yum J, Kim WY, Kim N. Machine Learning-Based Prediction of Delayed Neurological Sequelae in Carbon Monoxide Poisoning Using Automatically Extracted MR Imaging Features. American Journal of Neuroradiology. 2025 Jun 11.

For questions or inquiries, please contact:

• Grace Yoojin Lee - [email protected]