CatBoost C2h-21 model

Stages

1. 1.
   Patient enrollment and eligibility completion(selection)

   This stage defines the final analyzable cohort before measurement and modeling.

   Selection: Patients with metabolic dysfunction were enrolled, then eligibility criteria and study procedure completion determined the analyzed cohort.

2. 2.
   Non-invasive staging and EV feature acquisition(functional_characterization)

   This stage generates the input labels and biomarker features required for downstream ML modeling.

   Selection: Obtain steatosis/fibrosis stage labels by transient elastography and EV size/concentration features by nanoparticle tracking.

3. 3.
   Machine learning model development(broad_screen)

   This stage explores multiple model/task configurations to identify useful EV-based and multimodal classifiers.

   Selection: Develop six models for S0 versus S1-S3 and fourteen models for severe steatosis identification using different feature sets.

4. 4.
   Performance assessment and interpretability analysis(confirmatory_validation)

   This stage identifies the best-performing models and explains feature-stage relationships.

   Selection: Assess models using ROC-AUC, specificity, sensitivity, correlation analysis, and XAI/SHAP.

Steps

1. 1.
   Enroll patients with metabolic dysfunction

   Assemble the initial study population for MASLD-related biomarker analysis.

   Enrollment is required before eligibility filtering and study measurements can occur.

2. 2.
   Apply eligibility criteria and complete study procedures

   Define the final analyzable cohort with complete data collection.

   Eligibility and completion filtering narrows the cohort before feature acquisition and modeling.

3. 3.
   Stage steatosis and fibrosis by transient elastographystaging assay

   Generate steatosis and fibrosis stage information for the classification tasks.

   Stage labels are needed before supervised model development.

4. 4.
   Measure circulating plasma EV characteristics by nanoparticle trackingEV characterization assay

   Generate EV size and concentration features for model input.

   Feature acquisition follows cohort definition and provides the biomarker variables used in modeling.

5. 5.
   Develop EV-only and multimodal ML models for steatosis tasksclassification models

   Train models to distinguish S0 from S1-S3 and to identify severe steatosis.

   Model development requires both stage labels and feature inputs from the prior measurement stage.

6. 6.
   Evaluate model performance by repeated cross-validation

   Estimate predictive performance using ROC-AUC, specificity, and sensitivity.

   Performance evaluation follows model training to identify the strongest classifiers.

7. 7.
   Interpret feature relationships using correlation analysis and SHAP/XAIinterpretability method

   Explain how EV and other features relate to steatosis stages and model predictions.

   Interpretability analysis is performed after model development so feature contributions can be examined on trained models.