In modern precision medicine, two patients can walk into a clinic with the exact same clinical profile and receive the exact same treatment, yet follow entirely different disease trajectories and outcomes.

Real-World Evidence (RWE) derived from Electronic Health Records (EHR) captures this variation through diagnoses, treatments, and outcomes over time. However, it lacks sufficient biological depth.

While EHR data robustly captures the longitudinal trajectory of patient care, it often falls short in explaining the underlying biological mechanisms driving these observations. Without integrated molecular context, such as genomic variants, transcriptomic dysregulation, or pathway-level alterations, the underlying drivers of this clinical variability remain a black box.

For pharmaceutical R&D and translational teams, this creates a critical gap in identifying and characterizing clinically meaningful patient subgroups.

We solve this with our enterprise-grade, LLM-assisted harmonization platform Polly that standardizes fragmented, multi-modal clinical data into an AI-ready foundation. By automating the alignment of diverse real-world data to shared analytical frameworks, Polly bridges the gap between phenotypic RWE and multi-omics data up to 3x faster and helps empower teams transition directly from raw data to mechanistically grounded, biologically stratified patient phenotypes.

The Problem: Why Patient Stratification Fails

To uncover these biological insights, teams must systematically integrate clinical EHR outcomes with molecular data (like TCGA) into a unified analytical framework. This relies on establishing linkage anchors that are common data points that allow researchers to connect a patient's real-world clinical record with their multi-omics profile.

However, turning messy, fragmented clinical data into these linkage anchors is slow and error-prone. Standardization into a common framework is often a months-long bottleneck.

Clinical pipelines break across three layers:

• Tabular & Semi-Structured Data: Proprietary source schemas vary wildly across hospitals, making alignment difficult.
• EMR Digitized Notes: Physician reasoning, subtle adverse event signals, and patient narratives are locked in semi-structured text.
• Unstructured Free Text: Literature, trial protocols, and regulatory filings remain disconnected from the core record.

Manual extraction and mapping of multi-modal data is painstaking. Without automation, linking clinical and molecular data at scale is nearly impossible.

Our Approach: OMOP + LLM-Assisted Harmonization

We standardize fragmented EHR data into the OMOP Common Data Model (CDM), creating a consistent and analysis-ready foundation across diverse healthcare systems. This enables clinical data to be reliably integrated with molecular datasets for downstream research.

Polly accelerates this process using LLM-powered automation - mapping raw data to standardized formats, performing quality checks, and aligning clinical terminology. The result is faster, more reliable harmonization with full transparency, allowing teams to move from raw data to insight up to 3x faster.

1. Intelligent Data Mapping

Powered by advanced Large Language Models, Polly maps raw, messy data to the correct OMOP tables automatically. Every mapping comes with a confidence score, so your team can see why each decision was made to ensure full transparency and trust.

2. Built-In Quality Control

Data quality is critical. Polly performs automated health checks before mapping, flagging missing values, duplicates, or formatting errors. By catching anomalies upfront, your downstream analysis rests on a solid, reliable foundation.

3. Vocabulary Standardization

Standard database queries fail when doctors use different terms for the same condition. Polly solves this using domain-specific AI models (pre-trained on millions of biomedical sentences) to understand semantic context. It easily recognizes that EHR abbreviations like "MI" correspond to "Myocardial Infarction" and accurately maps synonymous clinical concepts.

Transparent Confidence Scoring

Healthcare AI cannot operate as a black box. Polly maps each source field to OMOP tables automatically, but every decision is auditable. Each mapping is scored using a transparent Weighted Cumulative Score:

• Semantic Match (35%): Does the clinical meaning align with OMOP standards?
• Data Type & Sample Analysis (50%): Are the dates, integers, and units technically compatible?
• Healthcare Domain Knowledge (10%): Is it placed logically in the clinical workflow (e.g., routing lab results specifically to the MEASUREMENT table)?
• Data Quality Indicators (5%): Applies penalties for high null rates, duplicates, or inconsistent data.

3. Performance You Can Trust

To combat "LLM Overconfidence," the pipeline utilizes a penalty-based accuracy model. Incorrect mappings incur a heavy deduction proportional to their clinical impact. In evaluations against ground-truth mappings, the system achieved:

• Overall Accuracy: 79.6%
• Perfect Concordance: 76.8%

This ensures that even when the AI is uncertain, the most calibrated, accurate answers reliably rank at the top for fast human validation.

What This Enables for Clinicians and Bioinformaticians

Standardized EHR data transforms your evidence pipeline. Polly enables high-value downstream use cases, including:

• Pharmacovigilance & Signal Detection: Harmonize post-market surveillance across hospitals for faster, more reliable adverse event monitoring.
• Comparative Effectiveness Research: Conduct reproducible, head-to-head real-world outcome studies across multiple health systems.
• Trial Feasibility & Patient Identification: Quickly assess site viability and find eligible patients using standardized data.
• FDA RWE Submission-Ready Data Lineage: Each AI mapping comes with a confidence score and fully auditable trail, ensuring your data meets regulatory standards.

‍