WISTERIA: Learning Clinical Representations from Noisy Supervision via Multi-View Consistency in Electronic Health Records

WISTERIA is a machine learning framework that improves clinical AI by treating noisy medical labels as uncertain observations rather than ground truth. By enforcing consistency across multiple weak supervision sources and incorporating medical ontologies, the method achieves better generalization across healthcare institutions and demonstrates robustness to label noise.

WISTERIA addresses a fundamental problem in healthcare AI: clinical labels are inherently unreliable. Traditional representation learning approaches imported from NLP assume labels are accurate, but real medical data comes from inconsistent billing codes, heuristic phenotypes, and incomplete annotations across different institutions. This paper reframes the problem by modeling labels probabilistically and learning representations that reconcile disagreement between multiple noisy labeling sources—an implicit denoising mechanism that recovers clinically meaningful patterns.

The research reflects a broader maturation in machine learning for healthcare. Earlier work focused on achieving high accuracy on benchmark datasets, but practitioners quickly discovered that models trained on noisy institutional data performed poorly in real deployments. Multi-view consistency approaches have proven effective in other domains (computer vision, NLP) where multiple modalities or annotations exist. WISTERIA's innovation lies in adapting this strategy specifically for EHR data while incorporating medical ontology constraints to preserve semantic relationships between clinical concepts.

For healthcare AI developers and institutions, this work has practical implications. Organizations struggling with label quality issues now have a principled framework that doesn't require expensive relabeling efforts. The demonstrated cross-institutional generalization is particularly valuable—models trained on one hospital's data typically fail at others due to different coding practices, but WISTERIA's approach mitigates this problem. The framework suggests that healthcare AI systems should explicitly model and leverage the variability in how different institutions document and code clinical observations rather than treating such variation as noise to eliminate.

The main technical question going forward involves scaling WISTERIA to modern large-scale EHR datasets and integrating it with recent large language model approaches for clinical text.

• →WISTERIA treats medical labels as stochastic observations of latent clinical states rather than fixed ground truth, addressing inherent noise in clinical labeling.
• →Multi-view consistency enforcement across weak supervision operators creates an implicit denoising mechanism that improves robustness to label noise.
• →Ontology-aware regularization preserves semantic structure in medical concepts, enhancing clinical meaningfulness of learned representations.
• →The framework demonstrates superior cross-institutional generalization compared to standard sequence-based pretraining, addressing a major deployment challenge in healthcare AI.
• →Results suggest that modeling the supervision process itself may be more effective for healthcare AI than relying on single supervision signals.