Kythera Labs Data Asset Passes Nine-Block Validation

The Kythera Labs all-payer real-world data asset met every pre-registered threshold across nine data-quality dimensions in an open-source validation study published in BMC Medical Research Methodology by researcher Onur Baser. Despite the expanding use of administrative claims data for regulatory, health-economics, and payer decision-making, no pre-registered, open-source validation of the Kythera Labs asset had been published previously. The findings support using the repository as an initial starting point for question-specific fitness-for-purpose assessments in regulatory, health economics and outcomes research, and payer research settings.

The Kythera Labs asset is one of the largest all-payer, tokenized administrative claims real-world data collections in the United States, containing 12.5 billion medical claims and 12.9 billion pharmacy claims contributed by more than 500 payer organizations and refreshed on a weekly basis. In the study, all-payer denotes adjudicated claims from commercial, Medicare, and Medicaid payers rather than a single plan or benefit type. Tokenized denotes that direct patient identifiers are irreversibly hashed into encrypted Datavant tokens at the data originator behind the Health Insurance Portability and Accountability Act boundary. The asset is distributed through Columbia Data Analytics, which licenses and delivers the data under a Data Use Agreement without re-adjudicating the source claims. Across the cumulative observation window from 2015 to 2025, the source universe contains 315,817,351 unique patient tokens. The author notes that this source universe does not represent a point-in-time census and makes no claim of contemporaneous coverage of the insured United States population.

To evaluate the asset, the study pre-specified a nine-block validation framework addressing internal consistency, concurrent validity framed as external consistency, construct validity, representativeness, missingness, record linkage, inter-coder reliability, temporal stability, and predictive validity. Operationalized through seventeen pre-registered equations, the framework anchors its nine analytic blocks to an explicit appraisal of data provenance and intake quality assurance, documenting the data-generation process rather than relying solely on derived statistics. The investigator applied the framework to an analytic cohort of 38,947,221 patients who had at least two longitudinal observations and at least one tokenized linkage element between January 1, 2015, and July 31, 2025. Record-linkage performance was estimated in a blinded, clinician-adjudicated probability sample of 11,882 candidate pairs with pre-specified subgroup stratification, alongside a purposive sample of 2,400 pairs representing four difficult matching scenarios. Predictive models were evaluated under random, temporal, and geographic validation schemes, and all analysis code was released open-source under an MIT licence.

Internal consistency proved strong for multi-item claims-derived constructs, yielding a Charlson comorbidity composite alpha of 0.87 and an Elixhauser comorbidity composite alpha of 0.89. Because internal consistency is not a sufficient reliability criterion for count-type comorbidity indices, the analysis added twelve-month test-retest reliability, which yielded intraclass correlations from 0.72 to 0.85, as well as known-group validity and discrimination for twelve-month mortality with c-statistics between 0.617 and 0.793. External consistency comparisons against the National Health and Nutrition Examination Survey and the Medical Expenditure Panel Survey demonstrated a mean absolute percentage error of 2.8 percent for diabetes prevalence and no more than 4.7 percent across six chronic conditions. For construct validity, a two-factor social-determinants confirmatory factor model achieved a Comparative Fit Index of 0.971, a Tucker-Lewis Index of 0.958, a Root Mean Square Error of Approximation of 0.038, and a Standardized Root Mean Squared Residual of 0.034, outperforming a one-factor alternative model.

Representativeness comparisons against the 2023 American Community Survey yielded standardized mean differences of no more than 0.061 across all fourteen demographic and geographic strata evaluated. In record-linkage assessments, Fellegi-Sunter tokenized linkage achieved a 96.2 percent match rate with an estimated false-match rate of 0.003. Against blinded clinician adjudication, linkage sensitivity was 96.7 percent, with a 95 percent confidence interval of 96.3 to 97.1 percent, and specificity was 99.5 percent, with a 95 percent confidence interval of 99.3 to 99.7 percent. Sensitivity ranged from 94.8 percent for Medicaid to 97.6 percent for Medicare across subgroups, declining to between 91.9 percent and 94.0 percent in the difficult matching scenarios.

Inter-coder reliability on International Classification of Diseases, Tenth Revision, Clinical Modification diagnoses yielded an overall Cohen’s kappa of 0.84, with condition-specific kappa values spanning 0.71 to 0.92 and positive predictive values ranging from 0.82 to 0.96. In predictive validity assessments, the area under the receiver operating characteristic curve reached 0.834 for 90-day mortality and 0.791 for 30-day readmission in a held-out random split, compared with 0.821 and 0.779 in a temporal holdout. Across leave-one-region-out folds, the values ranged from 0.812 to 0.845 for 90-day mortality and from 0.768 to 0.803 for 30-day readmission. The author emphasizes that fitness-for-purpose is defined relative to a specific research question, exposure, outcome, and decision context, meaning the findings do not certify the asset as universally fit for regulatory use. Additionally, several blocks evaluate derived statistics rather than individual clinical events. The nine-block framework remains reusable by sponsors, regulators, and academic investigators to qualify similar all-payer real-world data assets.

Subject of Research: Validation of the Kythera Labs all-payer real-world data asset using a nine-block framework

Article Title: Validation of the Kythera Labs all-payer real-world data asset: a provenance-anchored nine-block framework for tokenized evidence generation

Article References: Baser, O. (2026). Validation of the Kythera Labs all-payer real-world data asset: a provenance-anchored nine-block framework for tokenized evidence generation. BMC Medical Research Methodology. https://doi.org/10.1186/s12874-026-03070-1

Image Credits: AI Generated

DOI: 10.1186/s12874-026-03070-1

Keywords: real-world data, data validation, administrative claims, Kythera Labs, health informatics, predictive modeling, record linkage, data quality, Validation, Kythera, Labs, all-payer