Development and validation of a machine learning model of radiation induced hypothyroidism with clinical and dose–volume features (Radiotherapy & Oncology 2023)

Paper

This is the official repository for the Radiotherapy and Oncology 2023 paper "Development and validation of a machine learning model of radiation induced hypothyroidism with clinical and dose–volume features".

Authors: Mu-Hung Tsai, Joseph T.C. Chang, Hsi-Huei Lu, Yuan-Hua Wu, Tzu-Hui Pao, Yung-Jen Cheng, Wen-Yen Zhen, Chen-Yu Chou, Jing-Han Lin, Tsung Yu, Jung-Hsien Chiang

Predict-RIHT

In this repository, you will find the dataset to recreate results from the paper. Please note that you will need to install scikit-survival with: conda install -c conda-forge scikit-survival==0.21.0 Other requirements include numpy, pandas, and pickle.

Dataset

We are providing public access to the dataset, including the developmental cohort and external validation cohort.

Training / evaluation: Recreating results from the paper

To recreate results from the paper, ensure data_training.csv and data_validation.csv are in the same directory, and run training.py as below:

python training.py

Results from Table 3 are recreated. Please note this will overwrite the models files in the models/ directory from this repository, however the results should be the same.

Example of using ML models in a treatment planning system

Please see predict_example.py for an example of using ML models in a treatment planning system (TPS).

This script illustrates predicting radiation-induced hypothyroidism risk in a single patient, using a thyroid-only model and thyroid DVH data acquired from the treatment planning system via Varian ESAPI.

This script requires matplotlib and pyesapi libraries. Please ensure that the machine is able to connect to the TPS database (i.e. working ESAPI). A trained model should be in the models/ directory.

First, edit the script to define the patient ID to extract information from; this patient must have an organ named 'Thyroid'.

patient_id = '12345678'

Then, run the script to automatically extract the DVH information from the TPS, and predict freedom from >= grade 1 RIHT:

python predict_example.py

which will show a predicted survival curve for this patient.

Acknowledgements

This study was supported by a research grant from Varian Medical Systems.

Citation

If you found our project helpful, please cite our paper:

@article{
    riht2023, 
    title={Development and validation of a machine learning model of radiation-induced hypothyroidism with clinical and dose–volume features},
    volume={189},
    DOI={10.1016/j.radonc.2023.109911},
    journal={Radiotherapy and Oncology},
    author={Tsai, Mu-Hung and Chang, Joseph T.C. and Lu, Hsi-Huei and Wu, Yuan-Hua and Pao, Tzu-Hui and Cheng, Yung-Jen and Zheng, Wen-Yen and Chou, Chen-Yu and Lin, Jing-Han and Yu, Tsung and et al.}, 
    year={2023},
    pages={109911}
}