Cross-Institutional Record Linkage Without Global Identifiers

15 concepts14 papers1 proposal

Imagine a patient who calls an emergency hotline, is transported by ambulance to a local hospital, receives a prescription, and later tracks their recovery using a home telehealth device. In an ideal data ecosystem, these events would form a seamless chronological timeline. In reality, this journey is fractured. The hotline, the ambulance service, the hospital, and the telehealth platform each assign a different local identifier to the same human being.

As Ana Szarfman et al. argue in Recommendations for achieving interoperable and shareable medical data in the USA, the lack of a universal, secure patient identifier is a fundamental barrier to data integrity and clinical research. This "global identifier deficiency" is a systemic issue across data science—similarly observed in macro-resource tracking by Martin G. Skjæveland et al. in Publishing the Norwegian Petroleum Directorate's FactPages as Semantic Web Data—but its consequences in medicine directly impact public health and patient safety.

Without a reliable way to link records, researchers cannot easily reconstruct end-to-end patient pathways (Jonas Bienzeisler et al. in Trans-sectoral patient pathways in urgent and emergency care (TRANSPARENT study): protocol for a prospective, mixed-methods study in Germany), evaluate long-term treatment efficacy across distributed networks (Lisa M. Schilling et al. in Scalable Architecture for Federated Translational Inquiries Network (SAFTINet) Technology Infrastructure for a Distributed Data Network), ensure the discoverability of critical research datasets (Daniel Puttmann et al. in Assessing the FAIRness of databases on the EHDEN portal: A case study on two Dutch ICU databases), or track public health crises like tuberculosis across disconnected clinical and laboratory databases (Jessica Taaffe et al. in Advancing TB research using digitized programmatic data).

The challenge is twofold: data is highly siloed across jurisdictions with conflicting privacy mandates (Julian Jang‐Jaccard et al. in Data architecture for telehealth services research: A case study of home tele-monitoring), and unstructured clinical text frequently lacks standardized metadata entirely (James Gardner et al. in An integrated framework for de-identifying unstructured medical data). To bridge these gaps, researchers have developed a spectrum of methodologies designed to resolve identities across disparate systems.

Centralized vs. Probabilistic Alignment

When a formal, centralized infrastructure is legally and logistically feasible, identity resolution can rely on structured master indexes. For example, Raffaella Vaccaroli et al. document in Grand Duchy of Luxembourg: a case study of a national master patient index in production since five years how a national Master Patient Index (MPI) can successfully reconcile cross-border and regional healthcare systems using deterministic matching rules.

However, when linking massive, heterogeneous datasets across populations without a shared key, deterministic matching fails due to typos, missing fields, and varying data formats. To solve this at scale, Clícia Pinto et al. demonstrated a hybrid approach in Probabilistic Integration of Large Brazilian Socioeconomic and Clinical Databases, constructing a longitudinal cohort of 114 million individuals by combining deterministic rules with probabilistic Bloom filter matching.

Raw Demographic Data (Name, DOB) ──► Cryptographic Hashing ──► Bloom Filter (Bit Array)
                                                                      │
                                                                      ▼
                                                          Probabilistic Similarity Score
                                                       (Overlapping bits = Likely Match)

The intuition behind probabilistic Bloom filter matching is elegant: instead of comparing raw text strings (which exposes sensitive information and fails on minor typos), demographic attributes are mapped into a bit array using mathematical hash functions. If two records belong to the same person, their resulting bit patterns will highly overlap, even if there are minor typographical variations. This allows systems to calculate a mathematical probability of a match without ever exposing or directly comparing raw, sensitive personal data.

Cryptographic and Federated Identity Resolution

In many collaborative research scenarios, data cannot be pooled into a single location due to strict privacy regulations. In these federated networks, data must remain resident at its source site (Amin Ardeshirdavani et al. in NGS-Logistics : Federated analysis of NGS sequence variants across multiple locations). This creates a paradox: two institutions want to perform a joint analysis—such as identifying high-risk patients who frequently utilize both systems—but they cannot share their patient lists to find out which individuals they have in common.

To resolve this, Xiao Dong et al. turned to Secure Multi-Party Computation (MPC) in Developing High Performance Secure Multi-Party Computation Protocols in Healthcare: A Case Study of Patient Risk Stratification.. By combining Yao's garbled circuits with Cuckoo hashing, the authors designed a protocol where two institutions can evaluate an identity-matching function over their combined data.

Cuckoo hashing acts as a highly efficient mechanism to organize patient identifiers into predictable table structures. Yao’s garbled circuits then allow these tables to be compared in an encrypted state. The underlying mathematics ensure that both systems only discover the intersection (the matched patients) and their combined analytical outcomes, while the identities of non-matching individuals remain completely hidden and mathematically unreadable to the opposing party.

Metadata-Independent Re-identification

What happens when textual metadata is entirely absent, stripped for anonymity, or corrupted? In fields like neuroimaging, traditional identity resolution is impossible because clinical records are frequently decoupled from the images to protect privacy.

To bridge this gap, Lemuel Puglisi et al. introduced a paradigm shift in DeepBrainPrint: A Novel Contrastive Framework for Brain MRI Re-Identification, demonstrating that the clinical data itself can serve as an identifier. Their framework leverages contrastive deep learning to extract a unique structural "fingerprint" from raw brain MRI scans.

Brain MRI Scan ──► Contrastive Deep Learning ──► Low-Dimensional Vector (Fingerprint)
                                                               │
                                                               ▼
                                                  Invariant to: Scanner Type,
                                                                Image Contrast,
                                                                Patient Aging

The model is trained using self-supervised learning to recognize features of a patient's anatomy that remain invariant across different scanners, image contrasts, and the natural aging process. By mapping these complex 3D biological structures into a compact vector space, the system can re-identify a patient across historical records with near-perfect accuracy based purely on their anatomy, bypassing the need for text-based metadata entirely.

Bypassing Linkage via Confederated Learning

Finally, some modern approaches question whether explicit identity resolution is even necessary to achieve the goals of large-scale research. If the ultimate objective is to train a predictive machine learning model on a comprehensive dataset, do we actually need to link the records first?

Dianbo Liu et al. proposed a solution to this in Confederated learning in healthcare: Training machine learning models using disconnected data separated by individual, data type and identity for Large-Scale health system Intelligence. Their "confederated machine learning" framework is designed for scenarios where data is fragmented not just by institution, but also split across different data types and identities.

Instead of trying to resolve and stitch these identities together prior to training, the model's architecture is designed to learn from disconnected, unlinked data silos simultaneously. The model weights are updated iteratively across the fragmented layers, extracting population-level insights and predictive power without ever requiring a unified patient index or an explicit record linkage step.

Key Foundational Work

An integrated framework for de-identifying unstructured medical data (2009) — contributed 1 concept elements to this direction.
Scalable Architecture for Federated Translational Inquiries Network (SAFTINet) Technology Infrastructure for a Distributed Data Network (2013) — contributed 1 concept elements to this direction.
Data architecture for telehealth services research: A case study of home tele-monitoring (2013) — contributed 1 concept elements to this direction.
History & citationVersion 1 · 2026-10-04

Current · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_patient_identity_resolution_d5c14245,
  title = {patient identity resolution},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/patient-identity-resolution/rdr_d5c142452734452fa0d1d1af63b3aab1},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-19},
}

This exact version (version 1, frozen text):

@misc{lacuna_patient_identity_resolution_d5c14245_v1,
  title = {patient identity resolution},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/patient-identity-resolution/rdr_d5c142452734452fa0d1d1af63b3aab1/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-19},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake