MassMutual's Newly Granted Patent Turns Medical Claims Into a Mortality Score

US Patent 12,682,400, granted to Massachusetts Mutual Life Insurance Company on July 14, 2026, encodes an applicant's medical claim history as 70 GloVe embedding coordinates and feeds them into a random forest that outputs a relative mortality risk score, trained on roughly 40,000 historical underwriting applications carrying about 21,000 distinct diagnostic codes (USPTO, July 2026). The claim language, not the grant date, is what carriers renting a mortality-scoring model should read closely.

Inside the Claim: GloVe Embeddings and a Random Forest

The patent's independent claims describe a specific pipeline rather than a general appeal to machine learning. For each historical applicant record, the system feeds medical claims codes data, ICD diagnosis codes and CPT procedure codes drawn from a third-party claims source, into a word embedding model trained with GloVe (Global Vectors for Word Representation), the unsupervised algorithm built for natural-language corpora that MassMutual repurposes here to treat a person's chronological medical history as a document. The embedding model "encodes relationships amongst vector offsets representing the medical claims codes data" and reduces that relationship space to coordinates in two or three dimensions, then selects a minimum and maximum value for each dimension across every code in the applicant's history (US Patent 12,682,400 B1, claim 1).

In the specification's illustrative embodiment, GloVe maps codes into a 35-dimension Euclidean space. MassMutual then aggregates from the code level to the applicant level by taking the minimum and maximum of each of those 35 dimensions across every diagnosis and procedure in a person's claims history, which yields the 70 summarized embedding coordinates per applicant that the random forest actually consumes. The regression step is separate from the embedding step: a random forest trained on those 70 coordinates plus age, fit against underwriting decision data for the roughly 40,000 historical applicants whose complete claims and underwriting outcomes MassMutual could match, drawn from a production dataset covering about 21,000 unique diagnostic codes and 8,000 unique procedural codes. The disclosed model form is close to austere: relative mortality is modeled as a function of age plus the 70 aggregated embedding features, minimum and maximum, dimension by dimension.

Claim 1 also requires the system to generate an explanation tied to "the medical claims codes associated with the minimum values and the maximum values that resulted in a classification," a direct answer to the black-box objection regulators and reinsurers raise about claims-derived scoring. The patent claims interpretability as part of the invention rather than as a dashboard bolted on afterward, tying each classification back to the specific codes and embedding extremes that drove it. That detail also does useful legal work: a claim anchored in a traceable technical mechanism, not a generic prediction output, sits closer to the side of the current patent-eligibility line that has survived recent Federal Circuit scrutiny of machine-learning claims, a divide actuary.info covered in detail in its analysis of the 2025-2026 Section 101 reset for insurance AI patents.

Where the Score Replaces a Fluid Draw

MassMutual already runs a fluidless underwriting program, MassExpress, that qualifies some applicants for coverage without labs or a paramedical exam, based on underwriting rules, algorithms, and a client's digital footprint (MassMutual, Underwriting Practices). A Medical Claims Risk Score engine is built for exactly that decision point. Instead of waiting on an Attending Physician Statement, the multi-week pull of a doctor's records that has historically been the fallback for medically complex applicants, the system scores the applicant directly off claims data the carrier can obtain in days rather than weeks, and routes the file into an automated risk class, a request for an APS, or a full manual review.

The industry context makes the timing legible. Munich Re Life US's fourth biennial accelerated underwriting survey found the average acceleration rate with no human underwriter review running at 11% as of late 2024, with participating carriers projecting that figure could reach 49% of total life insurance business by 2030 (Munich Re, 2024). NAIC's most recent life insurance AI/ML survey found 58% of 161 responding life insurers use, plan to use, or are exploring AI or ML models in their operations, trailing the 88% reported by auto insurers and 70% by home insurers surveyed in the same series (NAIC, December 2023). A granted, explanation-generating mortality score sitting at the fluidless decision point is built for a market moving from an 11% acceleration rate toward that 49% ceiling, not for the narrower slice of applications that already clear on age and face amount alone.

A Claims-Derived Score and a Priced Mortality Basis

The reconciliation question the patent does not answer is what the random forest is actually trained to predict. The specification is not fully consistent on this point: the abstract describes training the model "to predict relative mortality risk for underwriting applicants," while the detailed description elsewhere describes the same random forest as trained to predict "underwriter-assigned risk rating." Those are different targets. A model trained against actual mortality outcomes is calibrated to what applicants died of and when. A model trained against historical underwriter risk-class assignments is calibrated to what underwriters, working from the same claims data and the company's existing manual, decided to call an applicant. The patent's training set, roughly 40,000 historical applications carrying "complete underwriting information such as final underwriting path and risk class," reads like the second case: the label is the underwriter's classification, not a matured mortality-experience study.

That distinction matters to a pricing actuary the moment the score leaves the underwriting desk and touches an assumption. A carrier's priced mortality basis, built from VM-20 experience studies and company-specific investigations, is calibrated to observed claim experience accumulated over years of exposure. A claims-derived score trained to reproduce underwriter judgment inherits whatever the underwriting manual already encoded, correctly or not, and offers no independent check on it; used as a triage gate, it can accelerate applications faster without ever being validated against the thing the pricing basis actually measures. SOA's accelerated underwriting research has flagged this general gap for several years, and actuary.info's own coverage of accelerated underwriting mortality slippage found average slippage running 15%, with individual programs ranging from 5% to more than 30% against priced assumptions, largely because carriers frequently lack the three to five years of seasoned claims experience needed to validate an AUW program credibly (Munich Re Life US, cited in actuary.info coverage). A patent claim can protect a pipeline. It cannot certify that the pipeline's output still means what the pricing actuary assumed it meant a policy year later.

Ownership of that reconciliation is an organizational question the filing leaves open by design; patents describe mechanisms, not governance. In practice it falls to whichever function owns model risk, sometimes a chief actuary's model governance committee, sometimes a separate model risk function reporting outside underwriting, to periodically revalidate the Medical Claims Risk Score output against emerging mortality experience and flag drift back to the pricing basis before it compounds across a growing block of MassExpress-issued business.

Four Years, Five Patents: The Family Behind This Filing

US 12,682,400 did not arrive alone. It is the newest member of a predictive-modeling patent family MassMutual has been building since at least 2020, several of which reuse overlapping language around a "fluidless mortality module" paired with a smoking-propensity model and a prescription-fills model, evidence of a continuation strategy that keeps extending claim coverage around the same underlying architecture rather than a single opportunistic filing.

Patent	Granted	What it covers
US 11,694,775 B1	July 4, 2023	Underwriting based on predictive modeling with excluded mortality risk factors
US 11,710,564 B1	July 25, 2023	Fluidless mortality, smoking-propensity, and prescription-fills modules trained on historical clinical data
US 11,983,777 B1	May 14, 2024	Underwriting estimator returning an immediate risk-class estimate with feature-attribution explanations
US 12,205,690 B1	February 25, 2025	Excluded risk factor predictive modeling, filed March 8, 2022
US 12,682,400 B1	July 14, 2026	Medical Claims Risk Score: GloVe embeddings plus random forest mortality scoring, filed August 26, 2022

The commercial context sharpens the IP-moat math. MassMutual has operated a mortality-scoring licensing business since at least 2018, when its LifeScore Labs subsidiary partnered with Swiss Re to bring the LifeScore360 algorithm to other carriers. MassMutual's then chief data scientist, Sears Merritt, said at the time that the company's "commitment to investing in data science and technology to develop risk scoring solutions with the potential to set a transparent, industry standard benefiting consumers and life insurance carriers alike is validated through LifeScore Labs' partnership with Swiss Re" (LifeHealthPro, March 2018). A newly granted, five-patent-deep family covering the mechanics of claims-based mortality scoring does two things for that business at once: it gives MassMutual's own LifeScore Labs product line a defensible claim set to point to, and it raises freedom-to-operate risk for any vendor selling a competing claims-to-embedding-to-random-forest pipeline into the same market. Reinsurer-supplied underwriting engines, RGA's AURA decision platform and Munich Re Automation Solutions' ALLFINANZ suite among them, are the obvious point of friction. A carrier licensing a third-party mortality-scoring module now has a live question its vendor contract should already answer: whether that module's architecture maps closely enough onto MassMutual's claim language to matter, and whose indemnification clause covers it if it does.

The Disparate-Impact Question Behind Claim-Code Density

The same claims-code density that makes the model interpretable is also what makes it a plausible proxy-discrimination vector. The number and specificity of ICD and CPT codes in a person's history is a function of how often that person saw a clinician and what was billed, not a direct measure of underlying health. That is the same mechanism a 2019 study in Science by Obermeyer, Powers, Vogeli, and Mullainathan found driving racial bias in a widely used healthcare risk-prediction algorithm: because the algorithm predicted health care cost rather than illness, and because historically less is spent caring for Black patients at a given level of need, the tool systematically underestimated their risk. The researchers calculated that correcting the bias would raise the share of Black patients flagged for extra care from 17.7% to 46.5% at the same level of clinical need (Obermeyer et al., Science, October 2019). A claims-derived mortality score built on code density inherits a structurally similar risk: an applicant with less healthcare access generates a sparser, lower-dimensional claims signature that a random forest can read as lower risk, precisely backward from the truth for an underinsured population.

Regulators have already written testing obligations for exactly this failure mode. New York's Circular Letter No. 7 (2024) requires insurers using AI or external consumer data in underwriting to demonstrate the tools do not use, explicitly or as a proxy, any protected class status, and to correct any unfair or unlawfully discriminatory outcome the testing finds (NY DFS, July 2024). NAIC's Model Bulletin on the Use of Artificial Intelligence Systems by Insurers, adopted December 4, 2023 and in force in 23 states plus the District of Columbia by late 2025, requires a documented AI governance program that covers underwriting models specifically (NAIC, 2023-2025). The American Academy of Actuaries has separately warned that predictive models inheriting bias from historical data or flawed assumptions can produce inequitable access and reputational risk even when no protected-class variable ever enters the model directly (American Academy of Actuaries, risk brief). None of that testing obligation is satisfied by the patent's explanation feature. Showing which embedding coordinates drove a classification tells a reviewer how the model reached its answer; it does not show whether the underlying claims-code density correlates with a protected characteristic through unequal access to care.

What the Filing Signals for MassMutual and Its Competitors

Read against MassMutual's four-year prosecution timeline and its five-patent family, the Medical Claims Risk Score grant looks less like a single defensive filing and more like the codification of a roadmap the company has been executing since at least 2020: fluidless mortality modules, excluded-risk-factor modeling, explanation-generating estimators, and now a specific, patented method for turning claims history into an interpretable mortality score. Every carrier currently deciding whether to build its own accelerated-underwriting mortality model or license one from a reinsurer has a sharper version of that question in front of it now. Building in-house means clearing the same GloVe-embedding-to-random-forest territory MassMutual has spent four years patenting around, with the freedom-to-operate diligence that requires. Licensing means inheriting whatever training-target ambiguity, disparate-impact exposure, and pricing-basis reconciliation burden the vendor's model carries, without the leverage to inspect claim 1 of the patent it might be infringing. Pricing and model-risk actuaries evaluating either path have a concrete checklist worth running now: confirm what a claims-derived score is actually trained to predict, test it for the proxy effects claims-code density can carry, and set a revalidation cadence against emerging mortality experience before the model's first seasoned policy year arrives, not after.