A Build-Time AI, Run-Time Determinism Engine for Healthcare Code Classification

Challenge

Almost every healthcare analytics decision (where to target revenue cycle effort, how to risk-adjust a population, which service lines drive cost and complexity) rests on how diagnosis, procedure, and DRG codes are grouped and scored. The reference tables underneath those decisions are usually opaque, hand-maintained, and impossible to audit, so a downstream number can change for reasons no one can trace. We needed grouping and complexity logic that would produce the same answer every time, hold up under scrutiny from a coder or an actuary, and stay version-controlled so any output could be tied back to an exact set of rules.

Approach

This is our signature pattern. We use an LLM where it is genuinely good (reading source material and resolving language) and we never let it make a decision at run time. At build time the model assigns complexity tiers and the roughly 10% of category mappings that are genuinely ambiguous. Those assignments are then frozen into committed reference tables and a rule-based classifier, so production is pure deterministic lookup with no model call in the path.

• AI is confined to build-time extraction and language work; the model proposes complexity tiers and the ambiguous category assignments, and humans review before anything is committed.
• The frozen output is plain reference data: committed CSV and Parquet tables built from AHRQ and CMS source data plus YAML taxonomy configs, so a given code maps the same way on every run.
• The classifier runs as four explicit layers: Layer 1 code-type identification, Layer 2 deterministic rules, Layer 3 the AI-assisted category (already frozen, not a live call), and Layer 4 data-quality flags.
• Decisions ship as deterministic code: unified four-level taxonomies linked by 19 service lines, a 5-tier clinical complexity and patient-acuity score, SQL generators, and a Python scorer.
• Everything is loaded to PostgreSQL and Athena and gated by an 8-tier validation suite, so a build that drifts is caught before it lands.
• Because the tables are versioned, any production number can be traced to the exact taxonomy version and rules that produced it.

Impact

• Coverage spans 75,238 ICD-10-CM diagnosis codes (21 domains, 124 groups, 506 categories), about 100,000 procedure codes across CPT, ICD-10-PCS, and HCPCS, and all 771 MS-DRG codes with CMS weights, including 46 trauma DRGs flagged.
• Run-time classification involves no LLM calls, so the same input yields the same output every time and results are reproducible and auditable rather than probabilistic.
• The 8-tier validation suite and committed CSV/Parquet artifacts give a defensible trail: each result maps to a specific taxonomy version and rule set.
• The engine is reused across multiple healthcare clients rather than rebuilt per engagement, so revenue cycle, value-based care, and analytics work all draw on the same vetted reference layer.
• In one 15-organization validation the engine produced an average complexity index of 2.65. This is a single multi-organization study, not a universal benchmark, and the figure describes that sample.
• The taxonomy and build scripts are published as an external open reference package; the AI assignment scripts remain proprietary, so the deterministic layer is inspectable while the build-time tooling stays in-house.

Capabilities demonstrated

• Build-time AI used to freeze deterministic, versioned, auditable reference tables and a rule-based classifier (no run-time model calls).
• Unified four-level diagnosis, procedure, and DRG taxonomies linked by 19 service lines.
• A 5-tier clinical complexity and patient-acuity scoring model delivered as a Python scorer and SQL generators.
• Reference data engineering from AHRQ and CMS sources plus YAML taxonomy configs, loaded to PostgreSQL and Athena.
• An 8-tier validation suite that catches drift before a build is committed.
• A clear split between an open, inspectable taxonomy layer and proprietary build-time AI tooling.

Anonymized by design: client names stay off the narrative per our reference policy. Figures that are modeled, small-sample, or targets are identified as such above.