menchelab / radipop_scripts

Scripts to train and validate random forest model for HVPG prediction

Radiomics-based prediction of portal hypertension severity and of liver-related events using routine CT scans of patients with cirrhosis

Hepatic venous pressure gradient (HVPG) is the reference standard to diagnose portal hypertension. Elevated HVPG is predictive of hepatic decompensation and mortality [Ripoll, 2007], and its measurement is indicated for diagnosis, therapy monitoring and risk stratification. However, HVPG measurement is invasive, relatively expensive and requires specialized medical infrastructure and expertise. Therefore, a non-invasive alternative is highly desirable.

In this project, we developed a radiomics-based model for the non-invasive determination of HVPG > 10mmHg (clinically significant portal hypertension, CSPH) from abdominal CT scans.

This work is published in <<<>>>

Code base for analysis

This codebase is organized in 3 main folders:

scripts_0preprocessing

• clean the metadata
• preprocess raw images
• extract radiomics features

scripts_1ml

• explore the feature space
• feature selection and batch correction
• train and optimize a random forest classifier to predict for HVPG ≥10 mmHg
• evaluate performance of the model

scripts_cox (Lorenz Balcar/Bernhard Scheiner)

• preform cox regression analysis for prognosis endpoints

References

Ripoll, C. et al. Hepatic venous pressure gradient predicts clinical decompensation in patients with compensated cirrhosis. Gastroenterology 133, 481–488 (2007)