RiskQ360by Next-Era
White Paper · June 2026
Technical White Paper

RiskQ360 — AI Chart Intelligence for Risk Adjustment & Quality See the whole chart. Capture every payable HCC.

Evidence-grounded · PDF-native · human-in-the-loop
RiskQ360 is an AI-powered chart intelligence platform for healthcare risk adjustment (HCC/RAF) and quality measurement (HEDIS). It reads medical records directly from PDFs — with no EHR integration — and converts them into coded, cited, audit-defensible intelligence. A dual ML + LLM pipeline pairs a BioClinicalBERT HCC predictor with TF-IDF ICD-10 retrieval and GPT-4o verification, so every payable code is backed by a MEAT breakdown and page-level evidence. Coders confirm with one click; compliance teams export RADV-ready packages; revenue and quality leaders see captured-vs-potential RAF and care-gap closure on one surface. The model never presents an unsupported code as payable.
01

Executive summary


Health plans lose risk-adjustment revenue and miss quality measures not because the evidence isn't in the chart — but because no one can read every chart, prove every code, and defend it on audit.

Risk adjustment and quality both come down to the same problem: finding clinically supported conditions and care events buried in long, messy clinical documents, then proving them with evidence a coder and an auditor will both accept. Today that work is manual, slow, and inconsistent — and the tools that automate it typically demand deep EHR integration and still produce codes a human can't trace. RiskQ360 closes that gap. It reads charts straight from PDFs, runs a dual ML + LLM pipeline to surface every legitimate HCC at the right ICD-10 specificity, gates each candidate through deterministic negation logic and MEAT verification, and presents the result as a payable pack with page-level citations. Risk adjustment, HEDIS quality, clinical suspicion signals, and revenue intelligence run on one platform, configurable per plan and line of business.

02

The problem


A single chart can run forty pages of progress notes, scanned faxes, labs, and discharge summaries. The conditions that drive RAF and the evidence that closes HEDIS gaps are in there — but they are scattered, inconsistently worded, and often negated, historical, or family-history mentions that must not be coded. Human coders are accurate but slow and expensive; they cannot review every chart, every year. Single-model AI tools are fast but hallucinate codes, miss specificity, and produce nothing a coder can audit line-by-line. And most platforms require an EHR integration project before they can see a single chart.

40+
pages in a typical chart — far more than a coder can review at volume
7,903
ICD-10-CM codes a tool must choose the right specificity from
0
EHR integrations RiskQ360 needs to start reading charts

The cost shows up as un-captured RAF, open care gaps, and audit exposure when codes can't be defended. The opposite failure — over-coding unsupported conditions — is just as dangerous under RADV. What's missing is a system that reads everything, codes precisely, proves every code, and keeps a human in control.

03

What RiskQ360 is


RiskQ360 is an AI chart intelligence platform that ingests medical records from PDFs (with per-page quality scoring and GPT-4o Vision OCR fallback), runs five LLM extraction pipelines plus a dual ML + LLM coding engine, and produces a Payable HCC Pack, a HEDIS Quality Pack, and an Audit Pack for every chart. The ML route (BioClinicalBERT) predicts HCC categories from raw text; TF-IDF retrieval finds the precise ICD-10s; ConText/NegEx gating and GPT-4o MEAT verification confirm support. Every payable code carries an ICD, an HCC, a RAF weight, a MEAT breakdown, and a page-level evidence span. Coders Accept / Reject / Modify; everything is timestamped and audit-logged.

Read Extract Predict Verify Score Confirm
It is not a black-box coder bolted onto an EHR feed. It is an evidence-grounded reasoning platform that reads every chart, codes at the right specificity, and proves every code a human is asked to confirm.
04

How it works


A single LLM asked to "code this chart" will hallucinate codes and miss specificity in long documents. RiskQ360 instead runs a deterministic pipeline in which two independent engines find candidates and a verification layer decides what is payable — nothing is coded on a single model's say-so.

ML — BioClinicalBERT

A multi-label HCC predictor reads the full chart text and surfaces candidate HCC categories — catching risk patterns the LLM misses across 40 pages, working on raw text rather than pre-extracted lines.

Retrieval — TF-IDF + ICD

For each predicted HCC, TF-IDF searches the 7,903-code ICD-10-CM catalog, constrained to the V28-mapped subset, matching clinical language to code descriptions to find the right specificity (E11.65 vs E11.9).

Verify — NegEx + GPT-4o

ConText/NegEx deterministically gates out negated, historical, and family-history mentions; GPT-4o then verifies clinical support and MEAT evidence and assigns a confidence score.

The dual approach is the edge. The LLM extraction route is good at obvious diagnoses but can invent codes; the ML + retrieval route catches patterns the LLM misses and lands the correct ICD-10 specificity that LLMs routinely get wrong. RiskQ360 merges both candidate sets, de-duplicates, then runs every survivor through negation gating and MEAT verification. The result captures more legitimate conditions, at higher code specificity, with fewer false positives.

Evidence-grounding & the six-status negation model. Every code links back to the exact page and character span that supports it. Mentions are classified into six polarities — active, negated, resolved, historical, family_history, uncertain — and only active conditions are payable (historical only when recaptured with current-year management evidence). Speculation is never presented as fact.

05

The processing pipeline


RiskQ360 decomposes chart coding into eight stages, run with parallel pipelines, so each stage contributes grounded evidence the next can build on.

01
PDF ingestion

Per-page text extraction with quality scoring; GPT-4o Vision OCR fallback for low-quality or scanned pages.

02
LLM extraction ×5

Demographics, clinical sentences (22 categories), risk diagnoses, HEDIS evidence, and encounters — in parallel.

03
ML HCC prediction

BioClinicalBERT predicts HCC categories from the full chart text, independent of the LLM extraction.

04
ICD-10 retrieval

TF-IDF retrieves precise ICD-10s for each predicted HCC, constrained to the V28-mapped subset.

05
Negation & context gating

Deterministic ConText/NegEx rules assign one of six polarities and filter non-payable mentions.

06
LLM + MEAT verification

GPT-4o verifies clinical support and extracts Monitor / Evaluate / Assess / Treat evidence per code.

07
HCC mapping & RAF

V28 ICD→HCC mapping with hierarchy suppression and constraining rules; RAF computed per encounter and member-year.

08
Output & review

Payable HCC Pack, HEDIS Quality Pack, and Audit Pack — surfaced for one-click Accept / Reject / Modify.

06

The knowledge-worker cockpit


RiskQ360 is a working surface for coders, auditors, analysts, and leaders — not a chat window. The core interaction is Accept / Reject / Add / Delete / Modify, every action timestamped and cited:

HCC Pack viewer · payable HCCs, supported ICDs, RAF weights, hierarchy MEAT evidence · Monitor / Evaluate / Assess / Treat per code Page-level PDF viewer · evidence spans traced to source text Full-text PDF search · find any term across the chart Confidence scoring · ML, LLM, and combined per code HEDIS measure matrix · eligibility, numerator, and open gaps Clinical suspicion signals · undercoded conditions surfaced from labs, vitals, meds RAF dashboard · captured vs. potential, top HCC categories Review workflow · Accept / Reject / Modify with audit trail Coder analytics · performance and acceptance over time Audit pack export · RADV-ready package per member
07

Inputs & integrations


RiskQ360 is PDF-native and meets charts where they already live — no EHR integration project to start. CMS-HCC V28 is the reference anchor; LLM providers are pluggable.

Chart sources
Storage
LLM providers
Reference data
ML engine
Persistence
Export
08

Use cases


Flagship · Risk adjustment coding

Pre-coded charts, coder-confirmed, audit-defensible

RiskQ360 pre-processes every chart and presents a Payable HCC Pack: each HCC with its supported ICD-10s, RAF weight, hierarchy explanation, MEAT breakdown, and page-level evidence. A certified coder reviews and Accepts / Rejects / Modifies — turning a blank-page coding task into a confirm-and-correct workflow at far higher throughput, with the audit trail built in.

  • Prospective — review before submission to capture supported HCCs and avoid over-coding.
  • Retrospective — re-read prior-year charts to recapture conditions and prevent RAF leakage.
  • Specificity lift — TF-IDF retrieval lands E11.65 where an LLM would default to E11.9.
Compliance & audit

RADV-ready documentation on demand

Every code is traceable to the page and span that supports it, with a MEAT checklist and confidence scores. Compliance teams export a complete audit package per member, validate vendor-submitted codes, and respond to RADV requests with documentation already assembled.

Quality & gap closure

Chart-based HEDIS evaluation across 100+ measures

RiskQ360 extracts BP readings, A1C values, screenings, immunizations, and medications, then evaluates eligibility, numerator compliance, and open gaps for measures such as COL, CCS, BCS, CBP, and TRC — configurable per plan and measurement year.

Clinical & revenue intelligence

Suspicion signals and captured-vs-potential RAF

The clinical intelligence engine cross-references labs, vitals, medications, and diagnoses to flag undercoded conditions — uncontrolled diabetes, CKD-stage discrepancies, drug-condition mismatches. Population dashboards rank opportunities by RAF-dollar impact and track annual recapture so nothing payable is left behind.

09

Architecture & technology


RiskQ360 runs as a containerized platform that deploys inside the customer environment — on-prem, private cloud, or folder-based — so PHI never has to leave.

Frontend

React 19 + Mantine 8 + TypeScript on Vite — a knowledge-worker cockpit with HCC, HEDIS, audit, and PDF-viewer surfaces and six visual themes.

Backend & orchestration

FastAPI on Python 3.12 with SQLAlchemy 2.0 and Pydantic v2, orchestrating ingestion, extraction, ML, and decisioning with feature-flag control per run.

ML & retrieval

BioClinicalBERT multi-label HCC predictor, TF-IDF vectorizer over the ICD-10-CM catalog, and a deterministic ConText/NegEx negation layer.

Data & reference

PostgreSQL 16 with an assertion-centric, 35+ table schema, seeded with CMS-HCC V28 mappings, coefficients, hierarchy rules, and HEDIS specs.

Configurable without code. LLM provider and model, prompt templates, chunk sizes, confidence and similarity thresholds, measurement year, and which capabilities run (risk-only, HEDIS-only, ML on/off, OCR on/off) are all controlled by feature flags — globally or per chart run.

10

How we compare


Established risk-adjustment vendors correlate claims and require EHR integration. RiskQ360's differentiators are PDF-native deployment, a dual ML + LLM pipeline, and a clinical suspicion-signal engine — delivered inside the customer's environment.

CapabilityRiskQ360Typical incumbents
PDF-native, no EHR integration✓ YesIntegration project
Dual ML + LLM coding pipeline✓ BioClinicalBERT + GPT-4oRules / single model
Clinical suspicion-signal engine✓ YesLimited / none
Page-level evidence & MEAT per code✓ YesPartial
Runs inside customer environment✓ On-prem / cloud / folderVendor cloud
Deployment speed✓ Fast — no EHR setupMonths

Positioned against Cotiviti, Optum, Inovalon, Reveleer, and Navina, RiskQ360 trades vendor-cloud breadth for evidence transparency, deployment speed, and a coding pipeline a coder can audit line by line.

11

Outcomes & evaluation


RiskQ360 is built to be measured. Every capability maps to an operational metric, evaluated at note level and member-year level with RAF-weighted precision and recall.

MetricWhat it measuresTarget direction
Review throughputCharts a coder confirms per hour vs. manual coding▲ Higher
HCC recallShare of supported HCCs the pipeline surfaces▲ Higher
Coding precisionShare of suggested codes that are clinically supported▲ Higher
RAF captureCaptured vs. potential RAF per member-year▲ Higher
ICD specificityCorrect sub-code selection vs. unspecified default▲ Higher
MEAT completenessPayable codes with documented MEAT evidence▲ Higher
Coder acceptance rateShare of suggestions accepted without override▲ Higher
12

Security & governance


RiskQ360 is designed for regulated, PHI-bearing environments. Human control and full auditability are architectural, not optional.

Human-in-the-loop confirmation

The AI pre-codes; a certified coder Accepts / Rejects / Modifies. No code is submitted without a human decision on record.

Page-level evidence citations

No ungrounded codes: each ICD, HCC, and HEDIS evidence item links to the exact page and span that supports it.

Full audit trail

Every action — who, what, when, before/after — is recorded for RADV response and internal review.

HIPAA / PHI handling

Deploys inside the customer environment so PHI never leaves; sensitive fields are access-controlled and governed.

MEAT & six-status negation

Deterministic clinical logic ensures only active, MEAT-supported conditions are coded payable.

Configurable governance

Feature flags, confidence thresholds, and prompt templates are controlled per plan and audit-logged on change.

About Next-Era

Next-Era builds evidence-grounded intelligence for the teams that run critical operations. RiskQ360 is our chart intelligence platform: it reads every chart, codes at the right specificity with a dual ML + LLM pipeline, proves every code with MEAT and page-level evidence, and keeps a human in control — so health plans capture what they're owed, close the gaps that matter, and defend every code on audit.

Request a demo
next-era.com
RiskQ360 · AI Chart Intelligence for Risk Adjustment & Quality — Technical White Paper
June 2026 · © Next-Era