Abstract
With the increasing availability of electronic health records (EHR) linked with biobank data for translational research, a critical step in realizing its potential is to accurately classify phenotypes for patients. Existing approaches to achieve this goal are based on error-prone EHR surrogate outcomes, assisted and validated by a small set of labels obtained via medical chart review, which may also be subject to misclassification. Ignoring the noise in these outcomes can induce severe estimation and validation bias to both EHR phenotyping and risking modeling with biomarkers collected in the biobank. To overcome this challenge, we propose a novel unsupervised and semiparametric approach to jointly model multiple noisy EHR outcomes with their linked biobank features. Our approach primarily aims at disease risk modeling with the baseline biomarkers, and is also able to produce a predictive EHR phenotyping model and validate its performance without observations of the true disease outcome. It consists of composite and nonparametric regression steps free of any parametric model specification, followed by a parametric projection step to reduce the uncertainty and improve the estimation efficiency. We show that our method is robust to violations of the parametric assumptions while attaining the desirable root-$n$ convergence rates on risk modeling. Our developed method outperforms existing methods in extensive simulation studies, as well as a real-world application in phenotyping and genetic risk modeling of type II diabetes.
Introduction
Biobanks linked with Electronic Health Records (EHR) are a goldmine for medical research, but they present a significant challenge: data noise. While genetic data is available for thousands of patients, we often lack a "ground truth" label for whether they actually have a specific disease. Instead, researchers rely on "surrogates" like billing codes or lab results, which are notoriously error-prone.
Existing phenotyping tools often rely on parametric models (like standard logistic regression). If the relationship between a biomarker and a disease is non-linear—a common biological reality—these models become biased. Conversely, fully non-parametric models are flexible but suffer from the "curse of dimensionality" and produce results that are difficult for clinicians to interpret. To bridge this gap, we introduce TUBE (Three-stage Unsupervised learning approach for Biomarkers linked with Error-prone outcomes), a semiparametric framework that allows for accurate genetic risk modeling even when disease labels are missing or noisy.
The TUBE Framework
The TUBE approach operates in three distinct stages to balance flexibility with statistical efficiency. The underlying data structure is modeled as a Directed Acyclic Graph (DAG) where the true disease status ($Y$) is a latent variable influenced by genetic markers ($G$) and manifested through various noisy surrogates ($X, U, Y^*$).
Methodology
- Stage I: Sieve-Approximated Composite Likelihood. Instead of assuming a specific shape for the data, TUBE uses "sieves" (flexible basis functions) to non-parametrically model how genetic markers and EHR surrogates relate to the unobserved true disease status. This "under-smoothing" step ensures the model captures complex patterns without being biased by rigid assumptions.
- Stage II: Condensing EHR Features. TUBE combines multiple noisy surrogates into a single, high-quality "phenotyping score." It then uses a non-parametric approach to estimate the model's performance (such as the ROC curve) without requiring true disease labels.
- Stage III: Parametric Projection. Finally, the results from the flexible non-parametric stages are "projected" back onto a standard parametric model (e.g., logistic regression). This provides researchers with interpretable coefficients ($\beta$ values) while maintaining high statistical efficiency (the desirable $\sqrt{n}$ convergence rate).
Key Results
Simulation Superiority
In extensive simulation studies involving non-linear data settings, TUBE demonstrated remarkable robustness. It reduced the Mean Squared Error (MSE) of genetic effect estimates by over 90% compared to traditional parametric methods, which failed to account for model misspecification.
Real-World Application: Type II Diabetes
The researchers validated TUBE using the Mass General Brigham (MGB) Biobank, focusing on Type II Diabetes (T2D) in a cohort of 16,963 participants. Using a Polygenic Risk Score (GRS) as the biomarker and ICD codes/A1C lab results as noisy surrogates, TUBE achieved:
- High Accuracy: TUBE’s genetic risk estimates were the closest to the "Validation" estimator derived from manual chart reviews.
- Superior Phenotyping: TUBE achieved an AUC of 0.894 for Hemoglobin A1C as a predictor of T2D, significantly outperforming other unsupervised benchmarks which typically ranged between 0.78 and 0.80.
Conclusion
TUBE provides a robust and efficient solution for disease risk modeling in large-scale biobanks. By starting with a flexible non-parametric foundation and concluding with a parametric projection, it avoids the pitfalls of model misspecification while remaining practical for real-world Genome-Wide Association Studies (GWAS). Future work will focus on extending TUBE to handle time-to-event data and higher-dimensional genetic variants.
Summary updated