Workstream 01

Data Platform — Healthcare Gold Layer

A semantic gold layer is the contract between your data and your AI. When that contract is precise, AI generates correct answers. When it's ambiguous, AI hallucinates.

Healthcare Gold Layer — Semantic Model Mockup

fact_patient_encounter FACT
encounter_keyPK
patient_keyFK
provider_keyFK
date_keyFK
length_of_stay_daysINT
total_charge_usdDECIMAL
total_cost_usdDECIMAL
readmission_flagBOOLEAN
discharge_status_cdVARCHAR
dim_patient DIM
patient_keyPK
patient_idVARCHAR
date_of_birthDATE
age_at_encounterINT (derived)
gender_cdVARCHAR
payer_type_cdVARCHAR
zip_codeVARCHAR
chronic_condition_flagsARRAY
risk_scoreDECIMAL
dim_provider DIM
provider_keyPK
provider_npiVARCHAR
provider_nameVARCHAR
specialty_cdVARCHAR
department_nmVARCHAR
facility_nmVARCHAR
is_attending_flagBOOLEAN
agg_monthly_clinical_kpi AGG
year_monthPK
facility_nmPK
total_encountersINT
avg_length_of_stayDECIMAL
readmission_rate_pctDECIMAL
cost_per_encounterDECIMAL
net_revenue_usdDECIMAL
patient_satisfaction_scoreDECIMAL
dim_diagnosis DIM
diagnosis_keyPK
icd10_codeVARCHAR
diagnosis_descVARCHAR
diagnosis_categoryVARCHAR
chronic_flagBOOLEAN
cms_drg_cdVARCHAR
semantic_metric_catalog SEMANTIC
metric_nmVARCHAR
canonical_definitionTEXT
source_tableVARCHAR
calculation_logicTEXT
grainVARCHAR
approved_byVARCHAR
last_validated_dtDATE

Why the semantic_metric_catalog matters: Without it, "readmission rate" means three different things across three departments, and your AI will average them and report a number that's wrong for everyone. The catalog is the single source of truth that makes AI answers trustworthy.

The core problem: When AI queries a data platform without a semantic layer, it guesses at metric definitions — and those guesses look confident even when they're dangerously wrong. Below are real-world examples of how the same question produces radically different results depending on whether a semantic gold layer exists.

✗ Example 1 — Readmission Rate Hallucination
AI Prompt
"What is our 30-day readmission rate for Q3?"
Data available (no semantic layer)
encounter table: readmit_flag (1/0) billing table: re_admit_within30 (Y/N) clinical table: readmission_30d (bool) -- No canonical definition exists -- AI picks encounter table, divides -- total readmits / total encounters -- including same-day procedures
AI Answer
"Your 30-day readmission rate for Q3 is 18.4% — significantly above the CMS benchmark of 15.2%. This represents ~2,840 readmissions across 15,435 encounters."
Reality
Actual rate: 11.2%. Same-day procedures were included — the AI didn't know CMS methodology excludes them. The CFO just declared a crisis that doesn't exist.
What went wrong: Three tables, three column names, no canonical definition. The AI picked one without knowing CMS methodology excludes same-day procedures. Result: 64% higher than reality — a board-level false alarm.
✗ Example 2 — Cost Per Encounter Ambiguity
AI Prompt
"Compare cost per encounter across our three facilities for cardiology."
What AI finds
facility_a: uses "direct_cost" facility_b: uses "total_charge" facility_c: uses "net_cost_after_adj" -- AI averages all three -- comparing apples to oranges -- to grapefruit
AI Answer
"Facility A: $8,200 · Facility B: $14,100 · Facility C: $9,800. Facility B is 72% more expensive — recommend operational review."
Reality
Facility B reports gross charges. On a net-cost basis, Facility B is actually the lowest cost at $7,900. AI just recommended penalizing your best performer.
What went wrong: No canonical definition of "cost per encounter." Each facility defined it differently. The AI confidently compared incompatible numbers.
✓ Example 1 — Readmission Rate with Semantic Layer
AI Prompt
"What is our 30-day readmission rate for Q3?"
Semantic catalog entry
metric: readmission_rate_30d definition: "Inpatient encounters where patient readmitted within 30 days, EXCLUDING same-day procedures (LOS=0), per CMS §412.4" source: fact_patient_encounter calc: SUM(readmission_flag) / COUNT(encounter_key WHERE length_of_stay_days > 0) grain: monthly / facility
AI Answer
"Your 30-day readmission rate for Q3 (CMS methodology, excluding same-day procedures) is 11.2% — below the CMS benchmark of 15.2%. This covers 14,107 eligible encounters."
Why it's trustworthy
✓ One definition · ✓ CMS-compliant exclusions · ✓ Correct denominator · ✓ Auditable to source table
What changed: The semantic catalog defined the metric precisely, including the CMS exclusion rule. No guessing, no variation across departments, fully auditable.
✓ Example 2 — Cost Per Encounter with Semantic Layer
AI Prompt
"Compare cost per encounter across our three facilities for cardiology."
Semantic catalog entry
metric: cost_per_encounter definition: "Net cost after contractual adjustments and write-offs, per completed inpatient encounter" source: fact_patient_encounter calc: SUM(total_cost_usd) / COUNT(DISTINCT encounter_key) -- maps ALL facility cost columns -- to the same net-cost standard
AI Answer
"Net cost per cardiology encounter: Facility A: $8,200 · Facility B: $7,900 · Facility C: $9,100. Facility B is your most cost-efficient site at 4% below system average."
Business outcome
Correct decision: leadership invests in Facility B expansion instead of ordering a review. Estimated value of correct decision: $2.3M capex directed correctly.
What changed: The semantic layer mapped all three facility cost columns to a single net-cost standard. The AI answered the real question, not the ambiguous one.

Reduce Reporting Cost

One gold layer = one answer. No more analyst hours explaining why Finance and Clinical have different numbers.

💰 30–50% reduction in BI support cost

Accelerate AI Time-to-Value

AI models prompted on a semantic gold layer produce trustworthy outputs immediately — no 6-month data cleaning project first.

💰 3× faster AI deployment

Avoid Costly Misdiagnoses

In healthcare, a wrong readmission rate drives wrong clinical decisions, wrong CMS submissions, and wrong quality program investments.

💰 Prevent $1M+ compliance risk

Enable Self-Service Analytics

When every metric has a canonical definition, any department can query confidently without an analyst intermediary.

💰 5× analytics reach per FTE
Calculate Your Full ROI →