Canonical text (the exact bytes fingerprinted)

TRUVACE RECORD VERSION
record: TRV-2026-0907
version: 1
kind: certified
reason: Certified into the record
timestamp: 2026-08-27T06:06:14.348936Z
status: published
lens: trace
sector: health
headline: Predicting Synchronous Liver Metastasis in Pancreatic Cancer Using CT Radiomics and Clinical Features: A Machine Learning Approach
dek: Rationale and objectives To address the challenge of preoperative prediction of synchronous liver metastasis (LM) in pancreatic cancer (PC), we developed and validated machine learning models integrating clinical and computed tomography (CT) radiomics features, and compared the performance and interpretability of linear (linear discriminant analysis [LDA]) versus nonlinear (multilayer perceptron [MLP]) architectures. Materials and methods This retrospective study enrolled 340 patients with pancreatic ductal aden…
gain_title: Integrated machine learning models using CT radiomics and clinical predictors achieved accurate preoperative prediction of synchronous liver metastasis in pancreatic ductal adenocarcinoma in validation, intended to assist decisions on surveillance versus biopsy or neoadjuvant therapy for indeterminate subcentimeter CT-
problem_title: The linear LDA model showed unstable performance and poor discrimination in lymph-node-negative and pancreatic head tumor subgroups despite overall validation AUC.
trace_subject: machine learning prediction of synchronous liver metastasis in pancreatic cancer using CT radiomics and clinical features
gain_reading: Integrated machine learning models using CT radiomics and clinical predictors achieved accurate preoperative prediction of synchronous liver metastasis in pancreatic ductal adenocarcinoma in validation, intended to assist decisions on surveillance versus biopsy or neoadjuvant therapy for indeterminate subcentimeter CT-
gain_evidence: enables accurate preoperative prediction of synchronous LM in PC | assist clinical decision-making in the common dilemma of indeterminate subcentimeter liver lesions on CT
problem_reading: The linear LDA model showed unstable performance and poor discrimination in lymph-node-negative and pancreatic head tumor subgroups despite overall validation AUC.
problem_evidence: LDA performed poorly in the lymph-node-negative (AUC = 0.50) and pancreatic head tumor (AUC = 0.62) subgroups
quick_read: In a retrospective study of 340 pancreatic ductal adenocarcinoma patients, investigators developed machine learning models combining CT radiomics and clinical predictors to predict synchronous liver metastasis preoperatively. In an independent validation cohort of 102 patients, the best linear model (LDA) reached AUC 0.828 and the nonlinear model (MLP) reached AUC 0.822, both showing good calibration with Hosmer-Lemeshow P values of 0.551 and 0.682.

The tool is framed not as a replacement for pathology but as decision support for indeterminate subcentimeter liver lesions on CT, where low-risk scores could support surveillance and spare biopsy and high-risk scores could prompt biopsy or neoadjuvant systemic therapy. Uncertainty remains about generalizability beyond this single retrospective cohort and about stability, as LDA dropped to AUC 0.50 in lymph-node-negative cases while MLP was more stable across subgroups.
limitation: 
tag: Dual reading
key_points: Retrospective study of 340 patients with pancreatic ductal adenocarcinoma (190 with LM, 150 without) with independent validation cohort n=102. | Five radiomics features selected via mRMR and LASSO combined with independent clinical predictors age, CEA, and clinical N stage to build 26 models. | Validation AUC 0.828 for LDA and 0.822 for MLP with good calibration and high net clinical benefit on decision curve analysis.
rundown: Researchers built 26 models from 340 retrospective cases using five selected radiomics features plus age, CEA, and clinical N stage, then evaluated optimal LDA and MLP models in a 102-patient validation cohort with ROC, calibration, decision curve, and SHAP analysis.

SHAP analysis showed clinical N stage dominated the LDA model while CEA and radiomics features were more prominent in MLP, and Friedman's H-statistic identified five significant feature interactions; MLP maintained AUC 0.66-0.77 across subgroups while clinical nomogram alone reached AUC 0.788.
sources:
- peer_reviewed | Academic Radiology | https://doi.org/10.1016/j.acra.2026.08.028 | 2026-08-25
prev: 0000000000000000000000000000000000000000000000000000000000000000

sha256

282f0df419b2cdd323f880ce7a909d3ca0e9ee7a9fd7143e4e955104b296e72d

previous

0000000000000000000000000000000000000000000000000000000000000000