Conversational Analytics for Healthcare: Why 'Natural Language' Means Something Different in a Hospital

Generic natural-language-to-SQL tools fail in healthcare because the vocabulary is encoded. “A1C” means HbA1c LOINC code 4548-4. “Diabetes” means a hierarchy of ICD-10 codes (E10.x, E11.x, E13.x, plus secondary codes). “Readmissions” carry a specific 30-day window with planned-readmission exclusions. A natural-language layer that doesn't understand the encoding produces wrong answers fluently.

What healthcare conversational analytics has to handle

1. Code system mapping. The user asks “diabetic patients.” The system has to translate to ICD-10 (E10-E13.9), include relevant secondary codes, exclude codes for “screening for diabetes” that aren't actual diagnoses.
2. Measure logic. “What's our MIPS score?” isn't a single number — it's a composite of category scores with specific weighting. The system has to know which measures the practice selected, what their denominators look like, and what exclusions apply.
3. Temporal logic. “Last quarter” can mean rolling 90 days or the last calendar quarter — different in healthcare contexts. “30-day readmissions” means within 30 days of discharge, not within the last 30 days.
4. Population logic. “Our patients” needs attribution rules — for an ACO, this is plurality-based; for an FQHC, it's patients with a qualifying visit in the look-back window.
5. Privacy preservation. The system must answer questions about cohorts without exposing individual PHI to anyone not authorized to see it.

What goes wrong with generic NL-to-SQL

Generic models trained on Stack Overflow and BI documentation produce SQL that joins the right tables but applies the wrong logic. Three failure modes:

• The query counts encounters where it should count unique patients.
• The query uses the most-recent-row pattern when measure logic requires a specific time-window aggregation.
• The query treats codes as exact-match strings when the measure requires hierarchical traversal.

The output is fluent SQL that returns a number. The number is wrong in subtle, hard-to-spot ways. A quality director who relies on it makes decisions based on incorrect data; trust erodes within a quarter.

What real healthcare conversational analytics requires

Three architectural elements:

1. A healthcare-specific semantic layer. Code systems, measure definitions, temporal patterns, and population logic encoded centrally — not inferred from a generic LLM.
2. Cited answers. Every output traces back to source rows. Users can audit every claim.
3. Tenant isolation. Each customer's data and conversations stay within their dedicated tenant. PHI does not flow to a shared model.

What Vizier built

Vizier's conversational analytics uses a healthcare-trained semantic layer specifically for the encoded vocabulary. Questions like “what's our 30-day all-cause readmission rate for CHF?” resolve to the correct ICD-10 codes, the correct 30-day window, the correct exclusions for planned readmissions, against your tenant's data. Every answer cites the rows.

See the three questions to ask every AI analytics vendor for the diligence framework that surfaces whether a vendor has built any of this.