Building Scalable LLM Evaluation Systems for Healthcare Prior Authorization

anterior · Healthcare · 2023 · source Agentic: no · promotional bias: med · source-backed · extracted with prompt 0.5.5 (schema v2)

Generated from extraction.json — do not edit by hand. Every claim cites verbatim source evidence; stated means the quote directly asserts the claim, implied means it supports it indirectly.

Problem

Domain: Healthcare prior authorization quality assurance

Anterior needed to evaluate prior-authorization AI outputs in real time for mission-critical healthcare decisions, while human-review sampling did not scale with production volume and new edge cases appeared only after deployment.

Constraints:

• Prior-authorization errors can have medical, regulatory, and customer-trust consequences.
• Rare mistakes become operationally significant at more than 100,000 cases per day.
• Human review at fixed sampling ratios scales linearly with case volume and clinician headcount.
• Offline datasets lag behind production because edge cases appear first in customer traffic.
• The system needs to know both which cases to review and how it performed on unreviewed cases.
• Human reviewers need efficient tooling and context to produce high-quality critiques.

Architecture

Anterior built a real-time reference-free evaluation architecture around its prior-authorization LLM pipeline. Medical records, guidelines, and questions flow into the primary pipeline, which emits an approval or escalation decision. A separate reference-free evaluator, potentially LLM-as-judge or logic-based confidence estimation, scores confidence before human labels exist and predicts the correct output. Confidence combines with contextual factors such as procedure cost, bias risk, and historical error rates to prioritize cases for human review. Clinical reviewers use the Scalp dashboard to inspect records and guidelines, label incorrect outputs, and write critiques; critiques generate ground truths for offline datasets and improve both the primary pipeline and the evaluator.

Diagram confidence: explicit (topology described by the source).

flowchart LR
  medical-records-and-guidelines[Medical records and guidelines] --> primary-llm-pipeline[Primary LLM pipeline]
  medical-records-and-guidelines --> scalp-review-dashboard[Scalp review dashboard]
  primary-llm-pipeline --> reference-free-evaluator[Reference free evaluator]
  primary-llm-pipeline --> customer-response[Customer response]
  reference-free-evaluator --> confidence-grading[Confidence grading]
  reference-free-evaluator --> predicted-correct-output[Predicted correct output]
  confidence-grading --> dynamic-review-prioritizer[Dynamic review prioritizer]
  confidence-grading --> production-gate[Production gate]
  predicted-correct-output --> alignment-monitor[Alignment monitor]
  dynamic-review-prioritizer --> scalp-review-dashboard
  scalp-review-dashboard --> ground-truth-generator[Ground truth generator]
  scalp-review-dashboard --> alignment-monitor
  ground-truth-generator --> offline-evaluation-datasets[Offline evaluation datasets]
  offline-evaluation-datasets --> primary-llm-pipeline
  offline-evaluation-datasets --> reference-free-evaluator
  alignment-monitor --> reference-free-evaluator
  production-gate --> customer-response
  production-gate --> fallback-model-pipeline[Fallback model pipeline]
  production-gate --> scalp-review-dashboard
  production-gate --> customer-review-dashboard[Customer review dashboard]
  fallback-model-pipeline --> customer-response

Components

Data flow

A medical decision request enters the primary LLM pipeline with records and guidelines, producing either approval or escalation for review. Before human labels exist, the reference-free evaluator scores the output, estimates confidence, and predicts the correct answer. The system uses those signals to estimate performance across all cases, compare evaluator outputs with human-reviewed cases, and prioritize the highest-risk cases for clinical review. Scalp reviewers see the medical record, guideline, question, and required context, label incorrect outputs, and write critiques. Those critiques generate ground truths for segmented offline datasets, which improve both the pipeline and the validator. In production, high-confidence cases can return directly to customers while low-confidence cases route to a more expensive model, internal clinician, or customer dashboard.

Why this architecture

Anterior uses reference-free real-time evaluation because offline datasets and fixed human-review percentages cannot keep up with heterogeneous medical records at production scale. The evaluator lets the system estimate performance before review labels exist, focus scarce clinicians on the riskiest cases, and validate the validator against cases that do receive human review. The production gate then turns evaluation from retrospective reporting into an active safety mechanism.

Models used: Primary LLM pipeline for prior-authorization approval/escalation decisions · LLM-as-judge reference-free evaluator · Logic-based confidence estimation methods

Scale

• Anterior serves insurance providers covering 50 million American lives.
  “we've scaled to now serve insurance providers covering 50 million American lives”transcript, stated
• The talk summarizes roughly 18 months of learning.
  “I'll share what we've learned in the last 18 months”transcript, stated
• A rare one-in-1,000 or one-in-10,000 mistake becomes significant above 100,000 cases per day.
  “this kind of mistake might happen every thousand cases or even every 10,000 cases but if you're processing more than 100 ,000 cases every day then that's a lot of mistakes”transcript, stated
• At 1,000 decisions per day, reviewing half requires about 500 reviews and five clinicians.
  “we're making around a th000 medical decisions per day well we want to know how we're doing so let's say we'll review half of those cases”transcript, stated

  “reviewing half of those means 500 human reviews per day and if every clinician on our team can do about 100 reviews per day that means we need five clinicians”transcript, stated
• At 10,000 decisions per day, the same review ratio would require 5,000 reviews and 50 clinicians.
  “we start doing 10,000 medical decisions a day to maintain the same percentage we would now have to do 5,000 human reviews every day”transcript, stated

  “maintaining the same ratio we now need 50 clinicians and that's bigger than our entire companies at the moment”transcript, stated
• Even a 5% review rate at 100,000 decisions per day still means 5,000 reviews and 50 clinicians.
  “let's say we now grow to 100,000 medical decisions per day which is still a very conservative number again we're back at 5,000 human reviews and 50 clinicians”transcript, stated
• Anterior reports reviewing tens of thousands of cases with fewer than 10 clinical experts.
  “we're able to review tens of thousands of cases with a review team of less than 10 clinical experts”transcript, stated
• The talk contrasts Anterior with a competitor that reportedly hired over 800 nurses.
  “one of our biggest competitors has hired over 800 nurses to perform reviews”transcript, stated
• Anterior reports nearly 96% F1 on prior authorization in a recent study.
  “performance at prior authorization with an F1 score of nearly 96% in a recent study”transcript, stated

Alternatives considered (roads not taken)

Review a fixed large percentage of all AI outputs

Rejected because: Human review needs grow linearly with case volume and quickly exceed available clinical staff.

Rely only on offline evaluation datasets

Rejected because: Offline datasets are downstream of delivered customer traffic and may include new edge cases only after damage has occurred.

Treat reference-free evaluation as a standalone score

Rejected because: The evaluator itself must be validated against human reviews and improved over time.

Return every primary pipeline output directly

Rejected because: Low-confidence decisions need further action through a stronger model path, internal review, or customer review.

Optimize review quantity without custom reviewer tooling

Rejected because: Reviewer quality and efficient context access matter more than raw review volume.

Patterns

llm-as-judge (evaluation)

Anterior feeds prior-authorization outputs into a separate evaluator with a scoring rubric to estimate correctness before human labels exist.

confidence-thresholding (routing)

The reference-free evaluator produces confidence grades that convert into predicted outputs and determine whether cases need review or further action.

human-in-the-loop-gate (safety)

Low-confidence or high-risk cases are routed to internal clinical reviewers or customer review dashboards instead of being returned automatically.

eval-driven-development (quality)

Human critiques generate ground truths, offline datasets track segmented performance, and production reference-free evals improve both the product and evaluator.

observability-tracing (operations)

The system estimates performance in real time across all processed cases and compares evaluator predictions against human-reviewed cases.

model-routing (routing)

Reference-free evaluation can route uncertain cases to a more expensive LLM pipeline before customer delivery.

Tradeoffs

Use reference-free evaluation for live traffic

Chose Evaluate before human review labels exist over Wait for offline labeled datasets.

Rationale: Real-time estimates let Anterior respond before edge cases propagate through customer workflows.

Prioritize review by risk

Chose Confidence grades plus procedure cost, bias risk, and historical error rates over Uniform random or fixed-percentage review.

Rationale: The same clinician capacity is more valuable when focused on cases with the highest expected risk of error.

Route uncertain cases to further action

Chose Customer return only for confident responses; model, internal, or customer review for uncertain ones over Always returning the primary LLM output.

Rationale: The gate preserves efficiency for high-confidence cases while adding protection where the evaluator sees risk.

Invest in expert clinical reviewers and custom tools

Chose Small expert team with Scalp dashboard over Ever-expanding reviewer headcount.

Rationale: High-quality reviewers with efficient tooling can produce critiques and ground truths that improve the whole system.

Use review data to improve the evaluator

Chose Build, audit, and improve the auditing system over Use review data only as a retrospective accuracy report.

Rationale: The feedback loop compounds: reference-free evals surface cases, human review validates them, and the evaluation system improves.

Gotchas (what bit them)

Rare medical nuance errors become major at production volume.

Root cause: An answer that sounds reasonable can still alter the medical meaning, such as confusing suspicious findings with a confirmed diagnosis.

Resolution: Use clinical review, reference-free evaluation, and prioritization to catch subtle high-risk cases.

Human-review percentages collapse economically as volume grows.

Root cause: Clinician review capacity scales linearly with case count.

Resolution: Use confidence and contextual risk to choose which cases receive scarce human review.

Offline evals can miss live edge cases until after customers are affected.

Root cause: Medical-record input space is highly heterogeneous and offline datasets are built downstream of customer delivery.

Resolution: Evaluate on live production data with reference-free methods and respond immediately.

The evaluator can become another unvalidated source of error.

Root cause: Reference-free scores are predictions and need calibration against human-reviewed cases.

Resolution: Compute AI-human alignment and validate the validator as part of the loop.

Reviewer throughput depends heavily on interface quality.

Root cause: Clinical reviewers need records, guidelines, questions, and required context available without friction.

Resolution: Build custom review tooling that surfaces context and captures critiques efficiently.

Evaluation

Anterior evaluates its production prior-authorization system through real-time reference-free confidence grading, predicted correct outputs, human-review alignment checks, segmented offline datasets built from reviewer critiques, and production routing outcomes.

Metrics: Confidence grade from high confidence correct to likely wrong · Predicted correct output · Estimated performance across all cases · AI-to-human alignment · Human-to-human alignment comparison · F1 score on prior authorization · Customer SLA response time

Outcomes

• Anterior reports serving insurance providers covering 50 million American lives.
• A review team of fewer than 10 clinical experts reviews tens of thousands of cases.
• AI-human review alignment became comparable to human-human reviewer alignment after several iterations.
• The system can identify and respond to errors quickly enough to maintain customer SLA expectations.
• A recent study reported nearly 96% F1 on prior authorization.

The reference-free evaluation architecture is explicitly motivated by scaling from MVP to 100,000+ daily medical decisions without needing 50 clinicians for even a 5% review rate.