NLP Development in Atlanta, GA

Healthcare / Revenue Cycle Management

A healthcare network in Atlanta (3 hospitals, 52 outpatient clinics, 1,400 providers, 2.4M annual encounters, $3.8B annual revenue) was losing revenue through incomplete clinical coding — physicians documenting conditions in clinical notes that coders did not capture because the notes were too long, the documentation was implicit rather than explicit, or the coder's time pressure caused them to miss secondary diagnoses. Coding problems: (1) HCC capture gap — Hierarchical Condition Category coding determining risk-adjusted Medicare payments. Clinical documentation: physicians frequently documenting chronic conditions in progress notes (e.g., "diabetes well controlled on metformin, A1c 6.8") but coders not always capturing every documented condition because notes averaging 800-1,200 words contained 8-14 codeable conditions buried in narrative text. HCC capture rate: estimated at 74% — meaning 26% of documented chronic conditions were not coded, directly reducing risk-adjusted payment, (2) coding specificity — ICD-10 requiring high specificity (e.g., Type 2 diabetes with diabetic chronic kidney disease — E11.22 versus the less specific E11.65). Clinical documentation: often using language that supported more specific codes than the coders assigned. Specificity gap: an estimated 18% of coded conditions assigned to less specific codes than the documentation supported — each specificity level reduction potentially reducing the risk-adjusted payment weight, (3) coder productivity — coders processing 18-22 encounters per hour for outpatient and 3-4 per hour for inpatient — the volume requiring rapid review that made thoroughness difficult. Coder turnover: 24% annually, with new coders requiring 6-8 months to reach full productivity. The productivity: creating a tension between throughput and completeness that favoured speed over revenue capture, (4) query volume — coders identifying documentation that was ambiguous or incomplete and sending physician queries to clarify — 4,200 queries per month. Physician response: averaging 4.8 days, with 12% of queries never answered. Each unanswered query: potentially representing a missed coding opportunity. The volume: consuming coder time to write and follow up on queries while consuming physician time to respond, and (5) retrospective review limitations — post-coding audits reviewing a 5% sample of coded encounters. Sample: too small to identify systematic gaps. Audit findings: consistently identifying missed codes and specificity opportunities but unable to quantify the full revenue impact across 2.4M annual encounters.

We developed a clinical coding NLP system over 12 weeks. Clinical text processing engine (weeks 1-5). Clinical NLP pipeline: custom NLP pipeline processing clinical notes — entity recognition identifying conditions, medications, procedures, and laboratory results. Contextual analysis: determining whether each entity was present, absent, historical, or attributed to family — the contextual understanding that distinguishes "patient has diabetes" from "father had diabetes." The pipeline: processing clinical notes in under 3 seconds per note, enabling real-time coding assistance. Medical terminology normalisation: clinical text mapped to standardised medical ontologies — SNOMED CT for clinical concepts and ICD-10 for coding. The normalisation: translating physician language ("sugar is high," "blood pressure running up") into standardised medical terminology that mapped to specific codes. Specificity extraction: NLP identifying documentation elements that supported higher-specificity codes — laterality (left/right), type specification (Type 1/Type 2), complication associations (diabetes with renal manifestation), and severity indicators. The extraction: finding the specificity details that human coders missed under time pressure. HCC coding assistance (weeks 3-8). Real-time HCC identification: as coders reviewed encounters, the NLP system highlighted conditions in the clinical note that mapped to HCC categories — each highlighted condition showing the suggested HCC code, the supporting text from the note, and a confidence score. The identification: bringing potential HCC codes to the coder's attention rather than requiring them to find every condition in lengthy notes. Gap identification: the system comparing NLP-identified conditions against the coder's assigned codes — flagging conditions found in the text but not yet coded. The identification: a second pair of eyes that never lost focus or felt time pressure. Specificity recommendations: when the NLP identified documentation supporting a more specific code than the coder had assigned, it recommended the higher-specificity code with the supporting text highlighted. The recommendations: improving coding specificity without requiring the coder to re-read the entire note. Historical carry-forward: the system tracking conditions documented in prior encounters — flagging chronic conditions that were documented in previous visits but absent from the current note's coding. The carry-forward: identifying conditions that patients continued to have but that the current visit's note did not explicitly re-document. Query automation (weeks 5-10). Automated query generation: when the NLP identified documentation that was ambiguous or incomplete, it generated a structured physician query — identifying the specific documentation gap, the clinical context, and the coding impact. The generation: queries that were more specific and clinically relevant than manually written queries. Query prioritisation: queries ranked by revenue impact — high-value queries (conditions with significant risk adjustment weight) prioritised over low-value clarifications. The prioritisation: ensuring physician query time was spent on the highest-value opportunities. In-basket integration: queries delivered to physicians through the EMR's in-basket — the familiar communication channel rather than a separate system. Physician response: one-click response options for common clarifications rather than free-text requirements. The integration: increasing physician response rates by reducing response burden. Retrospective analysis (weeks 7-12). Full-volume retrospective review: NLP system analysing all encounters (not a 5% sample) — identifying systematic coding gaps across provider, specialty, and condition categories. The review: quantifying the total revenue impact of coding gaps rather than estimating from small samples. Provider-specific feedback: each provider receiving personalised documentation feedback — conditions they consistently documented but that were under-coded, and documentation patterns that created coding ambiguity. The feedback: targeted improvement rather than generic documentation training. Revenue opportunity quantification: the system calculating the revenue impact of each coding gap — HCC weight impact, specificity improvement value, and total revenue opportunity by provider and condition category. The quantification: enabling the revenue cycle team to prioritise improvement efforts by expected return.

The NLP coding system deployed across the network within 14 weeks. HCC capture improvement: HCC capture rate: from 74% to 91% — the NLP system identifying conditions in clinical text that human coders missed under time pressure. Additional HCC captures: approximately 168,000 additional condition codes captured annually across 2.4M encounters. Risk adjustment revenue impact: $14.2M annually in improved risk-adjusted payments — each captured HCC contributing to more accurate risk scoring and payment. Coding specificity: specificity improvement: 18% of encounters receiving higher-specificity codes — the NLP identifying laterality, type, complication, and severity details that supported more specific ICD-10 codes. Revenue impact: $3.8M annually from improved code specificity. Coder productivity: coder productivity: improved 22% — from 20 encounters/hour (average) to 24.4 encounters/hour. Productivity mechanism: NLP highlighting relevant conditions and suggesting codes eliminated the need to read every line of lengthy clinical notes. Coder satisfaction: improved from 3.1/5 to 4.2/5 — the system reducing the tedium of searching for conditions in long notes. Physician queries: query volume: from 4,200/month to 2,800/month — the NLP extracting information from clinical text that previously required physician clarification. Physician response rate: from 88% to 96% — one-click responses reducing the response burden. Query revenue yield: improved 34% — prioritised queries targeting higher-value opportunities. Retrospective analysis: coding gaps identified across all 2.4M encounters (vs. 5% sample previously). Systematic patterns identified: 23 provider-specific documentation patterns and 8 specialty-wide coding gaps that the small-sample audit had not detected. Financial results: HCC capture revenue improvement: $14.2M/year. Coding specificity improvement: $3.8M/year. Coder productivity improvement: $1.2M/year. Query efficiency improvement: $340K/year. Total annual impact: $19.54M/year. NLP system development: $460,000. Annual operations: $84,000/year. First-year ROI: 4,148%.