Case study · OncoAuth

Cutting prior-auth burden in oncology imaging

A decision support platform co-built during the BiteLabs Digital Health, AI & Innovation Fellowship that turns payer policy documents into structured, queryable rules, so a primary care or oncology team can know within seconds whether a CT or PET scan request will clear.

Problem

Prior authorization for advanced imaging is the single largest drag on oncology workflow. Payer policies are buried in 40-page PDFs, change quarterly, and vary by plan. Clinicians and their staff spend hours per case re-reading documents they read last month, and patients wait days for a yes/no that should take minutes.

92%
Oncologists report PA-caused treatment delays
80.7%
Denied claims overturned on appeal
11.5%
Providers who ever file an appeal
39/wk
PA requests per physician

In 2024, Medicare Advantage processed 52.8 million prior authorization requests and denied 4.1 million. The system is not denying care because it is clinically inappropriate. It is denying care through attrition, because only 11.5% of providers ever appeal. Sources: ASTRO 2024; KFF/CMS January 2026; AMA 2024.

The denial rate is not a measure of clinical appropriateness. It is a measure of how many providers have time to fight back.

The insight

Inside that system is a more specific failure. Cigna, through its clinical review arm EviCore, denies PET scans for Stage IIB breast cancer. The American College of Radiology rates that same scan a 9 out of 9 ("Usually Appropriate") for exactly that patient. Same patient. Same scan. Guideline says yes. Payer says no.

That is not an administrative disagreement. It is a clinical contradiction. In our first extraction of just eight payer policy PDFs, we found 14 of these contradictions across eight insurers. No existing tool had ever identified them systematically, and no product compared payer policy against ACR or NCCN guidelines.

Example: Stage IIB breast cancer PET scan
Cigna / EviCore
Denied
Requires Stage III or higher for PET coverage.
vs
ACR Appropriateness Criteria
9 / 9
Rates PET as "Usually Appropriate" for Stage IIB+ breast cancer. Peer-reviewed clinical guideline.

The same logic applies across cancer types. Below is a prostate case: the payer requires prior authorization for PSMA PET, but ACR rates the scan 9/9 and the patient's PI-RADS 4 score indicates high probability of clinically significant cancer. OncoAuth flags the contradiction automatically:

OncoAuth contradiction alert showing a clinical contradiction between a payer's PSMA PET denial and ACR Appropriateness Rating of 9/9, with PI-RADS 4 High score and option to view full contradiction analysis
Live product: Contradiction Alert with ACR rating and PI-RADS integration. OncoAuth v0.6.

Approach

We modeled payer policies as a structured knowledge graph rather than free text. A retrieval layer translates clinician input (cancer type, stage, prior workup, suspected metastasis) into the canonical clinical concepts referenced by each plan, then walks the graph to a determinative result with a citation back to the source PDF.

The architecture is deterministic lookup, not generative AI. Every output is traceable to a published ACR, NCCN, or payer document. The Claude API is used only to extract and structure data from payer PDFs, not to make clinical decisions. No hallucination risk.

The architecture is modular. Each ACR scoring system (PI-RADS for prostate, BI-RADS for breast, Lung-RADS for lung, LI-RADS for liver, TI-RADS for thyroid) follows the same integration pattern, so adding a cancer type is a data task, not an engineering task.

The workflow is four steps: select the cancer type and stage, enter the insurance plan, choose the imaging test, and receive an instant color-coded result.

Green Pre-approved. No prior authorization needed.
Yellow PA required. Here is exactly what to document.
Orange Contradiction detected between the payer's policy and published guidelines.
Red Not covered under this plan. Here is the covered alternative.
OncoAuth v0.6 · Coverage Check
1
Cancer Type & Stage
Prostate, breast, lung, colorectal, melanoma
2
Insurance Plan
8 carriers: Medicare, Cigna, UHC, Aetna, Humana, BCBS, Kaiser, Cohere
3
Imaging Test
PET (FDG/PSMA), MRI, CT, bone scan
4
Instant Result
GREEN, YELLOW, ORANGE, or RED with source citation

Key features

Contradiction detection

Cross-references payer criteria against ACR and NCCN guidelines. Flags when denials conflict with clinical evidence. 14 contradictions found across 8 insurers.

Biomarker contradiction

New in v0.6: catches cases where a payer accepts NCCN imaging but denies NCCN biomarkers (Stockholm3, 4Kscore, PHI, SelectMDx, ExoDx) for the same patient.

PI-RADS v2.1 integration

ACR prostate imaging scores mapped to coverage status. PI-RADS 4 or 5 with a denied scan triggers an automatic contradiction flag. No competitor does this.

Delay Harm Score

Proprietary 0 to 100 scale quantifying patient risk from PA delays, built on NCI SEER survival data and stage migration probability. DHS 100 = critical urgency.

The Delay Harm Score in action. For a PI-RADS 4 prostate case with a 10-day expected delay, OncoAuth calculates a DHS of 100 (critical), pulls SEER survival data, and generates an appeal letter pre-loaded with the clinical evidence:

OncoAuth Delay Harm Score showing DHS 100 out of 100 Critical for a prostate cancer case, with stage migration risk, survival impact, cost delta, time window, natural history data, citations, and buttons to generate an appeal letter or request peer-to-peer review
Live product: Delay Harm Projection with SEER-derived survival data and one-click appeal generation. OncoAuth v0.6.

Why now

Four regulatory and technical forces converged in 2026 that made this tool possible:

What I learned and what's next

I owned the product end to end: policy-to-guideline matching schema, four-color classification system, Delay Harm Score methodology, closed beta with practicing oncologists, regulatory positioning under the Cures Act, and the pitch at BiteLabs Demo Day at Jones Day, Miami. First line of code to live beta in eight weeks.

The biggest lesson: the hard problem in prior authorization is not automation. It is that payer policies actively contradict published clinical guidelines, and no tool was surfacing those contradictions before the denial arrived. That insight shaped every design decision: deterministic over generative, citations over predictions, provider-side over payer-side. Next is validating against real-world PA outcomes in community oncology and expanding the contradiction dataset to breast, lung, and colorectal imaging.