Leveraging transfer learning for predicting acute graft- versus -host disease in stem cell transplantation

Acute graft- versus -host disease (aGvHD) is a major complication after allogeneic hematopoietic stem cell transplantation (HSCT), occurring when transplanted donor immune cells attack the recipient’s tissues. Despite advances in HSCT, no reliable clinical method currently exists to predict which patients will develop aGvHD. Early identification of high-risk individuals could improve donor selection and guide preventive strategies. Although machine learning techniques have demonstrated promising predictive performance, their practical deployment is limited by data privacy regulations, heterogeneity in healthcare records, and the scarcity of labeled samples across institutions. We propose a machine learning framework for aGvHD prediction that avoids raw-data sharing and addresses these limitations in decentralized data environments. The framework leverages three publicly available datasets and applies a transfer learning strategy. A source model is trained locally using only shared attributes, and its learned representations are transferred to enhance a target model trained on a different dataset. This approach enables knowledge integration across institutions without sharing sensitive patient-level data. Experimental results demonstrate that the proposed method outperforms baseline models, achieving an area under the receiver operating characteristic curve (AUROC) greater than 0.7 on the three datasets. Statistical significance was assessed using Dunnett’s test and bootstrap confidence interval test, revealing significant improvements in AUROC compared to the majority of baselines ( p < 0.01). A post-hoc explainability analysis further identifies key factors associated with aGvHD development. The findings of this study demonstrate that this strategy yields noteworthy outcomes, especially in terms of AUROC, despite the differences in the diagnoses, donor types, and graft manipulation between the datasets.

Authors

Institutions

Publication Details

Journal
PeerJ Computer Science
Published
2026-11-19
DOI
https://doi.org/10.7717/peerj-cs.4163
Primary Topic
Hematopoietic Stem Cell Transplantation
Type
article
Field-Weighted Citation Impact
0.00
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
article

Leveraging transfer learning for predicting acute graft- versus -host disease in stem cell transplantation

Jocelyn Dunstan, Carlos Iván Chesñevar, Axel J. Soto, Ana Gabriela Maguitman et al.
PeerJ Computer Science
Hematopoietic Stem Cell Transplantation
article

Leveraging transfer learning for predicting acute graft- versus -host disease in stem cell transplantation

Jocelyn Dunstan, Carlos Iván Chesñevar, Axel J. Soto, Ana Gabriela Maguitman, Julia Palma, Rodrigo Guerra, Lazaro Gibert-Garcia
article en

Abstract

Acute graft- versus -host disease (aGvHD) is a major complication after allogeneic hematopoietic stem cell transplantation (HSCT), occurring when transplanted donor immune cells attack the recipient’s tissues. Despite advances in HSCT, no reliable clinical method currently exists to predict which patients will develop aGvHD. Early identification of high-risk individuals could improve donor selection and guide preventive strategies. Although machine learning techniques have demonstrated promising predictive performance, their practical deployment is limited by data privacy regulations, heterogeneity in healthcare records, and the scarcity of labeled samples across institutions. We propose a machine learning framework for aGvHD prediction that avoids raw-data sharing and addresses these limitations in decentralized data environments. The framework leverages three publicly available datasets and applies a transfer learning strategy. A source model is trained locally using only shared attributes, and its learned representations are transferred to enhance a target model trained on a different dataset. This approach enables knowledge integration across institutions without sharing sensitive patient-level data. Experimental results demonstrate that the proposed method outperforms baseline models, achieving an area under the receiver operating characteristic curve (AUROC) greater than 0.7 on the three datasets. Statistical significance was assessed using Dunnett’s test and bootstrap confidence interval test, revealing significant improvements in AUROC compared to the majority of baselines ( p < 0.01). A post-hoc explainability analysis further identifies key factors associated with aGvHD development. The findings of this study demonstrate that this strategy yields noteworthy outcomes, especially in terms of AUROC, despite the differences in the diagnoses, donor types, and graft manipulation between the datasets.

PeerJ Computer ScienceVol. 12
St. Jude Children's Research Hospital (US), Pontificia Universidad Católica de Chile (CL), Ministry of Health (CL), Hospital Luis Calvo Mackenna (CL), Centro Científico Tecnológico - Bahía Blanca (AR), Universidad Nacional del Sur (AR), Federico Santa María Technical University (CL)
Hematopoietic Stem Cell Transplantation
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon