Multivariate Augmented Prediction-Powered Survival Inference for EHR Association Studies With Limited Chart Review
Abstract
Objective: To reduce bias and improve precision in survival association studies using electronic health record (EHR) data when computational phenotypes (CPs) for outcomes and covariates are error-prone and chart review is limited. Materials and Methods: We developed Multivariate Augmented Prediction-Powered Survival (MAPS) regression, which combines chart-reviewed and CP-derived labels across different validation patterns. MAPS uses a discrete-time survival model with a complementary log-log link to estimate hazard ratios (HRs). It incorporates data-adaptive augmentation weights, inverse-probability sampling weights, and patient-level cluster-robust variance estimation. We compared MAPS with chart-reviewed data (CRD) and CP-only analyses in simulation studies and a masked-validation study of immune-related adverse events (irAEs) among 1,026 patients receiving immune checkpoint inhibitors. Results: In the main simulations, MAPS produced approximately unbiased estimates and generally near-nominal 95% confidence interval (CI) coverage, with narrower intervals than CRD. CP-only estimates became increasingly biased as CP errors increased. In the EHR analysis, MAPS estimates closely agreed with fully validated estimates and improved precision over CRD, particularly at lower validation rates. CP-only analysis suggested increased irAE risk associated with pre-existing autoimmune disease (AID; HR=1.53; 95% CI, 1.12-2.09), whereas the fully validated analysis did not (HR=0.93; 95% CI, 0.67-1.29). Discussion: Precision gains depended on CP quality and validation rates, while small validation samples may affect confidence interval coverage. Conclusion: MAPS provides a flexible and efficient approach to EHR-based survival inference by combining limited chart-reviewed data with error-prone CPs across different validation patterns.
Related
- MMAD-Risk: Multivariate Mixed Survival Analysis for the Prediction of Age-Dependent Disease Risks from Plasma Proteomes
- QProtFuse: Query-Guided Layer Fusion for Retrieval-Augmented Protein Function Prediction
- Physically aligned forest fire risk prediction: A deep learning framework coupling fuel and climate multivariate factors
- RA-PLA: Retrieval-augmented graph convolutional networks for protein-ligand binding affinity prediction
- Association of BTRC rs10883617 with Survival Outcomes in Surgically Resected Non-Small Cell Lung Cancer
BibTeX
@misc{xie2026multivariate,
title = {{Multivariate Augmented Prediction-Powered Survival Inference for EHR Association Studies With Limited Chart Review}},
author = {Xicheng Xie and Katherine Fiona Simpson and Hyojung Jang and Theresa L. Walunas and Lili Zhao},
howpublished = {arXiv},
year = {2026},
month = oct,
doi = {10.48550/arxiv.2610.34749},
eprint = {2610.34749},
archivePrefix = {arXiv},
url = {https://arxiv.org/abs/2610.34749},
note = {CC BY 4.0},
}
RIS
TY - UNPB
TI - Multivariate Augmented Prediction-Powered Survival Inference for EHR Association Studies With Limited Chart Review
AU - Xicheng Xie
AU - Katherine Fiona Simpson
AU - Hyojung Jang
AU - Theresa L. Walunas
AU - Lili Zhao
PY - 2026
DA - 2026/10/26
PB - arXiv
DO - 10.48550/arxiv.2610.34749
UR - https://arxiv.org/abs/2610.34749
AB - Objective: To reduce bias and improve precision in survival association studies using electronic health record (EHR) data when computational phenotypes (CPs) for outcomes and covariates are error-prone and chart review is limited. Materials and Methods: We developed Multivariate Augmented Prediction-Powered Survival (MAPS) regression, which combines chart-reviewed and CP-derived labels across different validation patterns. MAPS uses a discrete-time survival model with a complementary log-log link to estimate hazard ratios (HRs). It incorporates data-adaptive augmentation weights, inverse-probability sampling weights, and patient-level cluster-robust variance estimation. We compared MAPS with chart-reviewed data (CRD) and CP-only analyses in simulation studies and a masked-validation study of immune-related adverse events (irAEs) among 1,026 patients receiving immune checkpoint inhibitors. Results: In the main simulations, MAPS produced approximately unbiased estimates and generally near-nominal 95% confidence interval (CI) coverage, with narrower intervals than CRD. CP-only estimates became increasingly biased as CP errors increased. In the EHR analysis, MAPS estimates closely agreed with fully validated estimates and improved precision over CRD, particularly at lower validation rates. CP-only analysis suggested increased irAE risk associated with pre-existing autoimmune disease (AID; HR=1.53; 95% CI, 1.12-2.09), whereas the fully validated analysis did not (HR=0.93; 95% CI, 0.67-1.29). Discussion: Precision gains depended on CP quality and validation rates, while small validation samples may affect confidence interval coverage. Conclusion: MAPS provides a flexible and efficient approach to EHR-based survival inference by combining limited chart-reviewed data with error-prone CPs across different validation patterns.
ER -