A medical diagnosis is rarely the whole story. For decades, broad labels like "lung cancer" or "autoimmune disorder" dictated standard, one-size-fits-all treatments regardless of the unique biological mechanisms at play.

Research is now driving a shift toward precision medicine through disease subtyping. By looking past surface-level symptoms and defining the shared underlying biology of a condition, researchers are stratifying patients and slicing broad disease labels into highly specific, molecularly distinct subgroups.  

The ultimate goal for the industry is - to ensure that the targeted therapies they develop are the exact right treatments for a patient's unique biology, right from day one. For example,  Moving from a broad "breast cancer" diagnosis to identifying a HER2-positive subtype directly triggers the use of the drug trastuzumab.

We approach this by using Elucidata’s LLM-powered platform, Polly, to harmonize multi-omics profiles with real-world evidence (RWE) and help researchers bridge the gap between clinical phenotypes and molecular data to achieve the precise patient stratification needed for highly targeted therapies.

The Problem: The Invisible Barrier to Personalized Care

To discover clinically actionable subtypes, R&D teams must find the hidden links between a patient’s molecular profile (omics data) and Real-World Evidence (like Electronic Health Records and clinical notes). Unfortunately, integrating this data presents severe roadblocks-

• "ETL Purgatory" and Messy RWE: A patient’s medical reality is scattered across unstructured physician notes, disparate hospital codes, and incomplete lab reports. Data scientists end up spending 80% of their time manually cleaning data just to prepare a single dataset for machine learning.
• The Multi-Omics Integration : Modern research relies on high-throughput multi-omics, spanning genomics, transcriptomics, and single-cell analysis to uncover hidden disease drivers. However, standard clinical data models (like OMOP) are built purely for observational data and lack native support for these complex molecular layers. Forcing omics data into unstandardized hospital schemas introduces severe batch effects, requiring researchers to build custom workarounds from scratch.
• Evidence backed AI : The sheer volume of this data has made AI and computational models, such as unsupervised clustering to discover hidden subtypes or supervised models (like SVM and KNN) to categorize patients. However, as these models become more complex, they often act as black boxes. Even if an AI correctly identifies a novel disease subtype, clinicians and regulators need to understand why. Without Explainable AI and transparent, auditable confidence scoring, translating a computational discovery into trusted clinical practice is nearly impossible.

The Solution: LLM-Assisted Harmonization

To accelerate biomarker discovery and patient stratification at scale, our approach is to shift from manual ETL pipelines to LLM-assisted harmonization

• LLM-Powered Data Extraction (Polly Xtract): Instead of manually mapping vocabularies, Polly automatically convert localized source codes and unstructured physician notes into standard analysis-ready concepts, compressing the data preparation phase.
• True Multi-Modal Scale (Polly Norm): We extend the standard OMOP model by building custom Variant Tables with direct Foreign Key relationships to core clinical records. This provides the native support for genomics that single-source vendors cannot match.
• Transparent AI Mapping and Standardization: Polly uses advanced LLMs trained on millions of biomedical sentences for validated concept mapping. Every mapping comes with a confidence score based on semantic alignment and data quality ensuring an auditable, regulatory-ready trail results.

Real-World Impact: Uncovering Hidden Lung Cancer Phenotypes

A preclinical R&D team needed to de-risk their oncology programs by integrating Real-World Data (RWD) with Public Omics Data (TCGA) in under 4 months.

Using Polly platform, millions of RWD records were mapped to our extended OMOP model in just 4 weeks. By establishing linkage anchors  based on age bins and disease ICD codes, the team built a unified feature matrix and applied unsupervised clustering to the data.

The Outcome- The pipeline successfully uncovered 4 distinct, molecularly informed lung cancer phenotypes.

While these patients looked identical based on standard EHR billing codes, the integrated omics data revealed drastically different survival trajectories:

• EGFR-driven patients (early stage) had an Overall Survival (OS) of 44 months.
• KRAS/STK11-mutant patients had an OS of just 23 months.
• Advanced / high-burden patients had an OS of 12 months.

Patients in the worst-prognosis subgroup survived, on average, less than one-third as long as those in the best-prognosis group yet standard clinical coding treated them identically. Without LLM-assisted harmonization linking the molecular layer to real-world outcomes, these subtypes would have remained entirely hidden.

What This Means for Drug Development

When R&D teams can reliably stratify patients by molecular subtype early in the pipeline, the implications go far beyond a single study:

• Smarter Clinical Trials: Enrolling a molecularly similar group of patients cuts through the noise, requiring smaller, faster trials to prove a drug actually works.
• Lower Failure Rates: A massive number of late-stage clinical trials fail simply because they treat mixed patient groups as uniform. Subtyping solves this heterogeneity problem at the root.
• Clearer Regulatory Paths: Explainable, auditable AI workflows provide the exact, reproducible evidence regulators demand for biomarker-driven drug approvals.
• Companion Diagnostics: Once a molecular subtype is clearly defined, teams can confidently develop a diagnostic test right alongside the new therapy.

We are moving toward a future where a diagnosis is no longer just a label, but a precise molecular map. The combination of multi-omics research and AI-driven subtyping is already reshaping precision medicine; the only question is whether your team has the tools to do it fast enough to matter. For patients waiting on a treatment that actually matches their biology, every day counts.

Elucidata's Polly platform combines LLM-powered data harmonization with extended OMOP modeling and accelerates biomarker discovery and patient stratification. Get in touch with us to learn more and overcome the data bottlenecks in precision oncology and your complex disease research.

‍