Suraj Vuduta
Mid-cycle · Inpatient CDI · Real de-identified ICU data

Site-Tailored CDI Optimizer

The same lab values should not produce the same query at every hospital. This engine reads de-identified ICU data from an academic medical center and a community hospital network, finds conditions with clinical evidence but missing or vague documentation, and routes each one by that site's policy and its physicians' learned agreement patterns. A model trained on one site barely beats a coin flip at the other.

Site policy

configured per client

By condition

documented when evidence is present
ConditionEligibleDocumentedQueriesHeld

Query worklist

EncounterOpportunityEvidenceAgree

Query

Why the model has to be local

physician agreement, AUC

Policy swap

eligible cases under each policy

Clinical validation and education

Pipeline

  1. Adapters map each site's raw tables (MIMIC-III and eICU have different schemas, units and code systems) to one shared encounter model.
  2. Condition rules flag clinical evidence for 10 CC/MCC conditions, such as Sepsis-3 or SIRS criteria, KDIGO creatinine rise, SpO2 with ventilatory support, BNP with IV diuretics, and hemoglobin drop with transfusion.
  3. Documentation check compares the evidence to what the site actually documented (billed codes at Site A, the clinician problem list at Site B) and labels it absent or nonspecific.
  4. Agreement model: a logistic model per site learns how often its physicians document the condition when evidence is present. Low-likelihood queries are held back, and conditions a site almost never documents go to education, not one-off queries.
  5. Ranking by impact (MCC over CC) × agreement likelihood, with a bump when an MCC would lift a low severity of illness, plus a DRG tier-shift estimate where grouper data exists.

Generative AI layer

An optional step drafts each query with Claude from the cited indicators only, following the ACDIS/AHIMA compliant-query practice: no leading, every reasonable option plus "Other" and "Unable to determine", and no mention of reimbursement. A grounding check blocks any draft that cites a number not present in the evidence, and note-review quotes must match the note verbatim. The queries on this page come from the rule-based template, so no API key is needed to view them.

Data and limits