Work

Cross-Cancer Transfer Learning for Gastric Cancer Risk Prediction from Electronic Health Records

Year: 2025

Type: article

Abstract: Background: Timely identification of individuals at elevated risk for gastric cancer (GC) within routine care could enable earlier endoscopy and referral. We posit that cancers within the gastrointestinal/hepatopancreatobiliary spectrum share signals that can be leveraged via transfer learning on electronic health records (EHRs) variables. Methods: We developed a cross-cancer transfer learning framework (TransferGC) on structured EHR data from the MIMIC-IV database, including 508 GC cases in the target cohort, that pretrains on non-gastric gastrointestinal/hepatopancreatobiliary cancers (colorectal, esophageal, liver, pancreatic) and then adapts to GC using only structured variables. We compared transfer variants against strong non-transfer baselines (logistic regression, XGBoost, architecturally matched multilayer perceptron), with area under the receiver operating characteristic curve (AUROC) and average precision (AP) as primary endpoints and F1 and sensitivity/specificity as secondary endpoints. Results: In the full-label setting, Transfer achieved AUROC 0.854 and AP 0.600, outperforming logistic regression (LR), extreme gradient boosting (XGB) and improving over the scratch multilayer perceptron (MLP) in AUROC (+0.024) and F1 (+0.027), while AP was essentially tied (Transfer 0.600 vs. MLP 0.603). As GC labels were reduced, Transfer maintained the strongest overall performance. Conclusions: Cross-cancer transfer on structured EHR data suggests a sample-efficient route to GC risk modeling under label scarcity. However, because all models were developed and evaluated using a single-center inpatient dataset, external validation on multi-center and outpatient cohorts will be essential to establish generalizability before deployment. If confirmed in future studies, the proposed framework could be integrated into EHR-based triage and clinical decision support workflows to flag patients at elevated GC risk for timely endoscopy and specialist referral.

Source: Diagnostics

DOI: 10.3390/diagnostics15243175

Authors: Daeyoung Hong, Jiung Kim, Jiyong Jung

Institution: Myongji University

Language: English

Topic: Machine Learning in Healthcare

Subfield: Artificial Intelligence

Field: Computer Science

Domain: Physical Sciences

SDG: Good health and well-being

Open Access status: gold

Estimated APC paid: $3,137