The load-bearing element of Optum's new dental patent is not the image recognition. It is a geometry test. Three neural networks each draw a box around the teeth they believe they see, and the system trusts its own answer only when those boxes land on top of one another. US 12,718,300 B2 issued August 25, 2026 with 20 claims (USPTO Official Gazette, Vol. 1549 No. 4). What issued automates a judgment a dental consultant now makes by eye, one radiograph at a time.

Key Takeaways

  • A commonality confidence score of one is the perfect case. The claim computes the ratio of common overlap between three bounding boxes to the total area of the overlay shape, so any disagreement between the procedure, material and tooth-surface models mechanically depresses it.
  • Three separate flag paths issue in the same patent: the submitted code not matching the AI-derived code, the procedure not being visibly detectable at all, or a periodontal chart read by OCR not supporting the code billed.
  • 4mm of pocket depth is what the ADA calls the most commonly recognized indicator in the literature for scaling and root planing. The chart-reading path turns that clinical benchmark into a machine-checkable condition on a claim line.
  • $8.4 million in rebates was ordered from six Massachusetts dental carriers in August 2026 for missing the state's 83% dental loss ratio floor, which caps what code-deviation savings are worth in a regulated state.

Patent Details

FieldValue
PatentUS 12,718,300 B2
TitleMethods, systems, and computer program products for flagging dental claims for further scrutiny based on processing of dental clinical images and periodontal charts using multiple artificial intelligence (AI) models
AssigneeOptum, Inc., Minnetonka, MN
InventorsQian Diao, Feili Yu, Chenda Deng, Zelong Zhang, Letitia Murr, Bo Han
Application18/400,276
FiledDecember 29, 2023
GrantedAugust 25, 2026 (Official Gazette Vol. 1549 No. 4)
Pre-grant publicationUS 2025/0217893 A1, July 3, 2025
Claims20
CPCG06Q 40/08, with G06T 7/0012 and G06T 2207/30036

The Geometry Test Inside Claim 1

Claim 1 carries an unusual amount of structure, and the structure is where the actuarial content sits. A clinical image goes through three models trained on three different labelling tasks: the first returns a predicted procedure code and a bounding box, the second a material type and its own box, the third the tooth numbers and surfaces involved with a third box. Material is what separates an amalgam restoration from a composite or a porcelain crown, and those are different codes at different allowables.

The three boxes are then overlaid, and the patent computes the ratio of the area of common overlap to the total surface area of the overlay shape. Where the boxes coincide the ratio is one. Where the procedure model is looking at tooth 30 and the surface model is looking at tooth 3, the ratio collapses toward zero and the derived code is not trustworthy.

That is a cheap answer to a real problem in ensembles of specialist vision models. Three independently confident predictions can each be right about a different tooth, and averaging their softmax scores would conceal it entirely. Spatial agreement catches it and requires no additional labelled data. Claim 1 then ends on the operative step, "flagging the dental claim when the submitted dental procedure code does not match any of the one or more dental procedure codes based on processing the clinical image" (US 12,718,300 B2, August 2026).

The comparison worth drawing is with the rest of this summer's claims-integrity cohort. Verisk's forgery detector examines the document a claim arrives on, TD Bank's model scores the network of people attached to it, and USAA's converts inspection density into a severity code. All three interrogate the claim's surroundings. Optum re-derives the code itself, which is narrower and more aggressive, because the output is not a suspicion score handed to an investigator but a competing answer to the question the provider already answered on the form. It also lands in CPC G06Q 40/08, the insurance class where subject-matter eligibility has been least forgiving.

The Threshold Nobody Specified

Nowhere in the issued claims is a numeric confidence threshold given, and that omission is the whole economic story. The score is continuous and the flag is binary. Something converts one into the other, and whoever sets that cutoff sets the flag rate, the review workload, the recovery volume and, eventually, the paid claims figure a pricing actuary trends forward.

The chart-reading path is where that bites hardest, because it automates a request payers already make by letter. The ADA notes that claims for more than two quadrants of scaling and root planing in a single visit will usually trigger a request for additional information such as a full-mouth periodontal charting. Optum runs OCR over that chart, binds each pocket measurement to the correct tooth by layout rules, and tests it against the ADA's own benchmark of 4mm. The distinction between a D1110 prophylaxis and a D4341 quadrant of scaling and root planing is the canonical dental upcoding vector, and it turns on exactly those printed numbers.

Every deviation the system catches moves a claim line from a higher allowable to a lower one, so paid dental claims per member per month fall by whatever recovery rate the model achieves. A rate filed on post-deployment experience therefore carries an embedded assumption about model performance that appears nowhere in the assumption exhibit.

The flag rate will drift when radiograph hardware changes or the CDT code set is revised, at which point the experience period stops being comparable to the projection period and nothing in the filing says so. Against $189 billion of national dental spending in 2024 (ADA Health Policy Institute), a one-point shift in that recovery rate is a material number.

The 83% Ceiling

The mechanism then runs into a constraint that has nothing to do with model quality. Dental carries no federal minimum loss ratio, but states have been building one. North Dakota enacted a 75% floor in April 2025, and Massachusetts voters approved an 83% floor by 71.6% at the 2022 ballot, effective January 2024.

Massachusetts enforced it for the first time this month. The Division of Insurance identified six carriers below the standard for 2025 and ordered $8.4 million in rebates on August 10, 2026, from Blue Cross and Blue Shield of Massachusetts, Guardian Life, Harvard Pilgrim, Reliance Standard, Standard Insurance and Starmount Life (ADA News, August 2026).

A code-deviation engine removes dollars from the numerator of that ratio. Every downcoded line is a benefit not paid, and in a state with a floor the saving does not stay with the carrier. It pushes the reported ratio toward the threshold and converts into a rebate obligation a year later.

The exposure is asymmetric in the other direction too. DPPO plans carry 89% of commercial dental enrollment, with 46% of commercial benefits self-insured (NADP 2025 Dental Benefits Report, December 2025), so the plan sponsor takes the claim savings while the carrier takes the abrasion of a bad flag. Guardian, a named payer deployment for Overjet, the incumbent vendor in this space, is on the Massachusetts rebate list. The tool that works hardest in a DLR state is the tool most likely to trigger the refund.

Further Reading

Sources