Skyline Collective · Real-world clinical data

1.79 million patients. Nine linked domains. A decade deep.

A de-identified EHR and claims dataset from a large integrated health system in the Southeastern U.S. — longitudinal, harmonized, and structured for therapeutic research and AI model development.

Nine domains · one de-identified patient key

Population

A full demographic cross-section, skewed where disease burden lives

Coverage spans every decade of life, with the deepest cohorts in the 50–80 range — where cardiometabolic, oncologic and neurodegenerative research recruits.

55.7%
female share of population
61 yrs
median age (as of 2026)
230K
patients aged 80+ — 12.9% of the population

Longitudinality

Patients you can follow for years, not visits

730,000 patients have more than a year between first and last encounter; 484,000 have three or more years — enough runway for outcomes, progression and treatment-pathway studies.

41%
of patients span ≥ 1 year of history
27%
span ≥ 3 years
16%
span ≥ 5 years

Clinical volume

58.7 million coded diagnosis records

Fully ICD-10-CM coded, running at 8–9 million records a year at peak, with over half a million active patients annually.

Condition coverage

Cohorts at power across nine disease domains

Thirty-six tracked chronic conditions and twenty oncology cohorts, defined on ICD-10-CM. From 573,000 hypertension patients to rare-disease cohorts in the hundreds — with the linked history behind every one.

Comorbidity structure

Disease doesn't come one at a time — neither does this data

604,000 patients carry two or more tracked conditions. The co-occurrence matrix below is the raw material for phenotyping, risk models and trial-eligibility engines.

Care settings

From clinic desk to ICU — the full care continuum

10.5 million encounters across ambulatory, emergency, inpatient and observation settings, with claims capturing place of service on the billing side.

Unstructured data

From clinical narrative to analysis-ready variables

Assessment-and-plan notes and imaging reports ship alongside the structured tables — and an LLM extraction pipeline renders them computable. Every report is read by two independent models, reconciled, and adjudicated by clinical reviewers. Watch it work on a de-identified echocardiogram report:

21
echo measurements targeted per imaging report
2 models
independent extractions per report, human-adjudicated
Redacted
names and dates stripped before annotation

Live extraction — echocardiogram report

Verbatim de-identified source text, dictation typos included — the model reads through them

IMAGING · TRANSTHORACIC ECHOCARDIOGRAM · DE-IDENTIFIED

Findings are consistent with bilateral pulmonary artery branch stenosis.
Small patent foramen ovale with left to right shunting of no hemodynamic significance
There is trivial bilateral peripheral pulmonic stenosis wwith a gradient of 10 mmHg in the RPA and 12 mmHg in the LPA
No PDA
aortic arch and pulmonary veonous returns was normal.
NOrmla ventricular function witha LVSF of 40% and an EF of 74%

─────────────────────────────────

Read by: PERSON On: DATE
Signed Electronically by: PERSON On: DATE

Structured output

Entity → value pairs, linked by relation, with source evidence

Data model

Nine linked domains, one patient key

Every record joins on a stable de-identified patient ID, harmonized across three specialty feeds: cardiology & oncology, neurology, and behavioral health.

Demographics

1.79 M PATIENTS

Age, sex, race, ethnicity. <0.01% missing DOB; <0.01% conflicting records after harmonization.

Diagnoses

58.7 M RECORDS

ICD-10-CM with description, code set and primary/secondary flag, tied to encounters.

Encounters

10.5 M VISITS

Typed visits — clinic, ED, urgent care, inpatient, observation, surgery — with facility and dates.

Claims

11.5 M LINES

Claim lines with CPT code, place of service, and encounter linkage.

Procedures

CPT-4 CODED

Procedure records with CPT code and description — imaging, surgery, therapy.

Laboratory

RESULTS + RANGES

Test code and name, result value, units, and reference ranges, per encounter.

Medications

RX ORDERS

Medication name, coded prescription, sig/instructions, quantity, refills, med-rec flags.

Vital signs

POINT OF CARE

Timestamped measurements with value and units, tied to encounters.

Clinical notes

NLP READY

Assessment & plan notes and imaging narratives, de-identified — with LLM-extracted measurements delivered as structured variables.

DE-IDENTIFIED

All records are de-identified at source; patient IDs are irreversible hashes. Every statistic on this page is an aggregate, and any cell counting fewer than 11 patients is suppressed.