About

U.S. government data from the NIH All of Us Research Program, the CDC, CMS, and the Census Bureau trains nine open models that share one scorer. This page has the data sources, the conditions, the models, the papers, and the team.

Data

U.S. government data sources: the NIH All of Us Research Program (health records, surveys, Fitbit), the CDC (surveys, Lyme cases), CMS (Medicare costs), and the U.S. Census Bureau (county population).

All of Us is studied inside its secure Researcher Workbench, and only disclosure-checked aggregate results leave it. The CDC, Census Bureau, and CMS data are public files.

Data sources

Datasets

Standards and code systems

  • HL7 FHIR R4The record format the apps and the scorer read.
  • SMART App LaunchOpening the clinician app from inside an EHR.
  • CDS Hooks 2.0The service that returns a card to the EHR.
  • USCDIThe data classes read: problems, medications, laboratory tests, procedures, encounters, and demographics.
  • ICD-10-CMDiagnosis codes, for model labels and chart findings.
  • LOINCLaboratory tests.
  • RxNormMedications.
  • CPTProcedures and tests.
  • SNOMED CTDiagnoses already recorded in FHIR charts, such as Long COVID.

How often each condition is coded

Show the numbers

How the numbers were made

All of Us lets researchers take only group counts out of its research enclave: counts from 1 to 20 are suppressed, and every other count is rounded up to the nearest multiple of 10. The models in this demo were fit outside the enclave from those released counts, so no patient-level data left it.

All of Us results on this site follow the program's Data and Statistics Dissemination Policy, and each public output is reported to the program before release.

Acknowledgment

We gratefully acknowledge All of Us participants for their contributions, without whom this research would not have been possible. We also thank the National Institutes of Health's All of Us Research Program for making available the participant data examined in this study. The All of Us Research Program is supported by the National Institutes of Health, Office of the Director: Regional Medical Centers: 1 OT2 OD026549; 1 OT2 OD026554; 1 OT2 OD026557; 1 OT2 OD026556; 1 OT2 OD026550; 1 OT2 OD 026552; 1 OT2 OD026553; 1 OT2 OD026548; 1 OT2 OD026551; 1 OT2 OD026555; IAA #: AOD 16037; Federally Qualified Health Centers: HHSN 263201600085U; Data and Research Center: 5 U2C OD023196; Biobank: 1 U24 OD023121; The Participant Center: U24 OD023176; Participant Technology Systems Center: 1 U24 OD023163; Communications and Engagement: 3 OT2 OD023205; 3 OT2 OD023206; and Community Partners: 1 OT2 OD025277; 3 OT2 OD025315; 1 OT2 OD025337; 1 OT2 OD025276. In addition, the All of Us Research Program would not be possible without the partnership of its participants.

Nine invisible illnesses

For each condition we measured, in the NIH All of Us Research Program, how often people in care for a diagnosis they report have its code in their record. For each of the nine we also built a model that finds people like them whose diagnosis is missing from the record.

What we built for each condition. Have a code: share of people in care who report the diagnosis (5 or more years of records; lupus, all lengths of record) with its code; for Long COVID, adults who ever had it, against survey prevalence. Model: area under the ROC curve in a held-out validation split, among people who report the diagnosis and were never coded (Long COVID: a later code).
ConditionHave a codeModelClinicianPatient questionsAtlas
Infection-associated chronic illness
Long COVIDabout 1 in 90.67CardRECOVER, DePaul, PCFSAtlas
ME/CFS24%0.74CardDePaul, IOM checkAtlas
Lyme disease39%0.65CardExposure questionsAtlas
Autoimmune
Lupus69%0.82Cardnot yetAtlas
Chronic pain
Fibromyalgia63%0.73Card2016 criteria checkAtlas
Migraine67%0.69Cardnot yetAtlas
Women's health
Endometriosis41%0.61Cardnot yetAtlas
Mental health
Depression68%0.76CardPHQ-9Atlas
Anxiety68%0.79CardGAD-7Atlas
For comparison
Hypertension, type 2 diabetes85%, 87%not applicablenot applicablenot applicablenot applicable

AI models

Eight of the nine models are trained on patients' own reports that a clinician diagnosed them, instead of on diagnosis codes, so they learn to find people whose diagnosis was never coded. The Long COVID model is trained on a later Long COVID code. Each model is fit outside the All of Us Researcher Workbench from released aggregate counts, and one shared scorer runs all nine in the clinician app, the CDS Hooks service, and the patient app.

Training on patients' reports finds the uncoded

Show the numbers

Key figures

More results from the papers. Each figure links to the paper it comes from, where the full caption and methods are.

Panel A: AUROC with 95% intervals for four models predicting a later Long COVID code at day 90, from about 0.57 to 0.68 in cross-validation and 0.62 to 0.67 at held-out health systems. Panel B: among patients without a later Long COVID code, rate ratios comparing high and low scorers for later codes of Long COVID symptoms and of unrelated conditions such as fracture and cataract, on a log scale.

Early chart data modestly predict a later code

Models using the first 90 days after COVID-19 modestly predicted a later Long COVID code, in cross-validation and at 18 held-out sites. After adjusting for use of care, high scorers got new codes for unrelated conditions about as often as for Long COVID symptoms (rate ratios 1.2 to 2.1 and 1.8 to 2.1), so the score tracks documentation and use of care in general. From the Long COVID undercoding paper

Panel A: share coded relative to hypertension and type 2 diabetes, by race and ethnicity, for seven conditions on a log scale; chronic fatigue is near 0.2 to 0.3 and endometriosis, relative to the female anchors, near 0.4 to 0.6 in every group shown. Panel B: share coded by years of records before the survey; hypertension and type 2 diabetes rise from about 0.75 to about 0.87, while chronic fatigue rises only from about 0.2 to 0.25.

The gap holds across groups and years of records

The low coding of contested conditions appears in every racial and ethnic group shown, and it closes only slowly as patients accumulate more years of records. From the self-reported diagnosis coding paper

Horizontal dot plot of the 20 largest weights in the exported Long COVID model, with 95% intervals. The largest are a personal history of COVID-19 coded in days 31 to 90 (about 1.3) and COVID-19 still coded at a follow-up visit in days 31 to 90 (about 1.2), followed by a chest CT, shortness of breath coded by day 90, three or more and two or more symptom types coded in days 31 to 90, and fatigue or malaise coded by day 90. Others include a D-dimer test, new chest pain, a new inhaler, a history of sleep apnea, and a C-reactive protein test.

What moves the Long COVID score

The chart findings that raise the exported model's score most, led by a personal history of COVID-19 code and two or more symptom types coded in days 31 to 90. Each weight is what a finding adds to the score when it is present. From the Long COVID undercoding paper

Papers

Four papers by Amir Asiaee and Kaveh Aryan. Each is a preprint in preparation; the medRxiv link will appear here when it is posted.

  1. Long COVID undercoding and early prediction of its documentation: survey comparison and an exportable model in the All of Us Research Program

    Amir Asiaee and Kaveh Aryan

    How often Long COVID is missing from the record and for whom, and whether early chart data can flag patients before the code appears, with a model that can leave the research enclave.

    Preprint in preparation; medRxiv link here when posted

  2. Electronic Health Record Coding of Self-Reported Diagnoses in the All of Us Research Program

    Amir Asiaee and Kaveh Aryan

    How often the record carries a diagnosis that patients say a clinician gave them, for ten conditions, including chronic fatigue, fibromyalgia, endometriosis, and Lyme disease.

    Preprint in preparation; medRxiv link here when posted

  3. Self-reported diagnoses as training labels for finding uncoded chronic fatigue, fibromyalgia, and endometriosis in the All of Us Research Program

    Amir Asiaee and Kaveh Aryan

    Whether models trained on patients' reports of a diagnosis, instead of on diagnosis codes, rank higher the people who report a diagnosis but have no code for it. The chronic fatigue (ME/CFS) and fibromyalgia cards in this demo use these models.

    Preprint in preparation; medRxiv link here when posted

  4. Design and Technical Evaluation of a Standards-Based System That Delivers Enclave-Derived Case-Finding Models to Patients and Clinicians

    Amir Asiaee and Kaveh Aryan

    The system behind this demo and its testing on synthetic records: versioned model files, one scorer shared by the patient app, the clinician app, and a CDS Hooks service, and the Atlas of aggregate results.

    Preprint in preparation; medRxiv link here when posted

Team

Built for the HHS/NIH TOPx Tech Sprint, Invisible Illness track, October 2026.

Contact: [email protected]