Flatiron Health Clinico-Genomic Database

Flatiron Health's Clinico-Genomic Database (CGDB)

merges two oncology data layers: (1) curated EHR data from the Flatiron Network of ~280+ US community cancer clinics, where technology-assisted abstraction adds what raw EHR lacks — confirmed diagnoses, regimen definitions, lines of therapy, progression events, and mortality follow-up; and (2) Foundation Medicine comprehensive genomic profiling linked at the patient level, supplying mutation, copy-number, and fusion status plus tumor mutational burden and microsatellite instability.

Why it matters for RWE

Oncology exposes the weakness of generic sources: progression is never billed, regimens are hard to infer from claims J-codes alone, and biomarker status determines eligibility for modern therapy. Flatiron solves all three by construction — human-in-the-loop abstraction validates regimen starts/stops and progression, and CG linkage makes molecularly defined cohorts possible. Published comparisons show Flatiron populations resemble treated oncology populations at academic/community mix better than SEER-Medicare resembles all comers (who include untreated patients).

Operational characteristics

• Curated variables: line-of-therapy numbering, response/progression dates, ECOG-derived performance indicators, mortality via enhanced multi-source follow-up.
• Genomic layer: CGP reports (hundreds of genes), TMB, MSI, and selected PD-L1 results; testing availability is itself selection-conditioned (tested patients skew toward advanced/metastatic settings).
• Population: predominantly advanced/metastatic disease in community settings; under-represents early-stage and purely academic-managed patients.
• Regulatory standing: FDA has accepted Flatiron-based external controls in oncology submissions; the platform's curation processes are documented in published validation studies.

Common pitfalls

• Testing-conditional selection: genomic analyses describe tested patients only; test indication patterns create spectrum bias.
• Progression imprecision: abstracted progression improves on claims but remains scan-cycle-bound and reviewer-dependent.
• Line-of-therapy conventions: definitions (counting maintenance, peri-operative, rechallenge) must match study objectives explicitly.
• Community-clinic geography: coverage tracks the Flatiron Network footprint; not nationally representative by design.

Pros, cons, and trade-offs

• vs SEER-Medicare: clinical/genomic depth and all-age reach vs population-based incidence capture and elderly cost data.
• vs claims sources: progression/regimen fidelity vs complete utilization/cost capture.
• Trade-off: treated-cohort realism vs population denominators — Flatiron describes treated cancer patients superbly but cannot speak to undiagnosed/untreated populations.

When NOT to use

Incidence/prevalence estimation; cost-of-care requiring payer-paid amounts; early-stage/screening populations; questions about untested patients.

Decision diagram

flowchart LR
  FN[Flatiron Network clinics - 280+] --> TA[Tech-assisted abstraction\nlines - progression - death]
  FM[Foundation Medicine CGP reports] --> LK[Patient-level linkage]
  TA --> CGDB[Clinico-Genomic Database]
  LK --> CGDB
  CGDB --> O[Biomarker-defined oncology RWE\nexternal controls]

Worked example

Scenario

Describe overall survival by EGFR mutation class among first-line osimertinib-treated metastatic NSCLC patients.

Dataset

OS by EGFR mutation subclass in first-line osimertinib users.

Steps

Result

Ex20ins subgroup showed markedly inferior OS (13.1 vs 38.6 months); results robust to landmark and censoring-at-last-visit choices.

Trade-offs

Runnable example

\
import pandas as pd
from lifelines import KaplanMeierFitter

def first_line_cohort(regimens, dx, min_stage="IIIB"):
    r1 = regimens[regimens.line_number == 1].sort_values(["patient_id","start_date"])
    idx = r1.groupby("patient_id").first().reset_index()
    return idx.merge(dx[["patient_id","stage","histology"]], on="patient_id")

def os_curve(cohort, cutoff):
    kmf = KaplanMeierFitter()
    t = (cohort.death_date.fillna(cutoff) - cohort.start_date).dt.days.clip(lower=0)
    e = cohort.death_date.notna().astype(int)
    kmf.fit(t, e)
    return kmf.median_survival_time_, kmf.confidence_interval_

\
library(survival)

os_by_class <- function(cohort, cutoff) {
  cohort$time <- pmin(as.numeric(cohort$death_date - cohort$start_date, units = "days"),
                      as.numeric(cutoff - cohort$start_date, units = "days"))
  cohort$event <- as.integer(!is.na(cohort$death_date))
  survfit(Surv(time, event) ~ mutation_class, data = cohort) |>
    surv_median()
}

\
proc lifetest data=coh plots=survival(cb);
  time os_days * death(0);
  strata mutation_class / test=logrank adjust=sidak;
run;

/* Median OS with CI */
ods output quartiles=q;
proc lifetest data=coh;
  time os_days * death(0);
  strata mutation_class;
run;

Citations

1. [1]Snow TS, et al. Comparison of Population Characteristics in Real-World Clinical Oncology Databases in the US: Flatiron Health-Foundation Medicine Clinico-Genomic Databases, Flatiron Health Research Databases, and the National Cancer Database and SEER. medRxiv. 2023.
2. [2]Ma X, et al. Comparison of Population Characteristics in Real-World Clinical Oncology Databases in the US: Flatiron Health, National Cancer Database, and SEER. medRxiv. 2020.
3. [3]Flatiron Health. Flatiron Data and Research documentation.

1. [4]Dahlen AD, et al. Evaluating the generalizability of commercial healthcare claims data. American Journal of Epidemiology. 2025.