A time-dependent explainable radiomic analysis from the multi-omic cohort of CPTAC-Pancreatic Ductal Adenocarcinoma

Gian Maria Zaccaria ¹, Francesco Berloco ¹, Domenico Buongiorno ²

⁰Department of Electrical and Information Engineering (DEI), Polytechnic University of Bari, Via Edoardo Orabona, 4, Bari, 70126, Italy.

Computer Methods and Programs in Biomedicine +

|

September 29, 2024

View abstract on PubMed

Summary

Radiomics improves prognostic models for Pancreatic Ductal Adenocarcinoma (PDA). Survival machine-learning classifiers integrated with radiomics enhance patient risk stratification and provide explainable predictions for overall survival and recurrence.

Area Of Science

Oncology
Bioinformatics
Medical Imaging

Background

Pancreatic Ductal Adenocarcinoma (PDA) presents unmet clinical needs for novel prognostic factors.
Multi-omic models are emerging to address these needs.
This study leverages public multi-omic datasets from the CPTAC-PDA project.

Purpose Of The Study

To develop and validate a pipeline using survival machine-learning (SML) classifiers and explainability methods.
To stratify prognosis in PDA patients using multi-omic data.
To identify quantitative prognostic factors for overall survival (OS) and recurrence (REC).

Main Methods

Analysis of radiologic images, clinical, and mutational data.
Feature selection using univariate (UV) and multivariate (MV) survival analyses.
Comparison of four SML classifiers (Cox, survival random forest, generalized boosted, SVM) with concordance (C) index assessment.
Explainability analysis using SurvSHAP(t) for top-performing models.

Main Results

SML classifiers incorporating radiomics outperformed those using only clinical or mutational data.
For OS, a Cox model with radiomic, clinical, and mutational features achieved a 75% C index.
For REC, an SVM model with radiomics achieved a 68% C index.
Key predictors included radiomic features (e.g., Median Gray Level intensities), gender, and tumor grade.

Conclusions

Radiomics demonstrates significant potential for improving risk stratification in PDA.
The study provides a time-dependent explainability of top multi-omic predictors in PDA prognosis.
This approach enhances the understanding of radiomics' contribution to predicting patient outcomes.

Keywords:

Explainability Machine learning Pancreatic ductal adenocarcinoma Radiomics Survival analysis