zeroknowledgediscovery/zebra_adrd: ZeBRA-ADRD: Zero-Burden Risk Assessment for Early Alzheimer's Disease Screening

This release provides the reference implementation and public API interface for ZeBRA-ADRD, a machine-learning–based algorithm for early screening of Alzheimer's disease and related dementias (ADRD) using only routinely collected electronic health record (EHR) data. ZeBRA-ADRD enables passive, bloodwork-free risk assessment and is designed for scalable population-level screening and presymptomatic trial enrichment.

The algorithm models longitudinal comorbidity patterns derived from diagnoses, prescriptions, and procedures, without requiring laboratory tests, imaging, genetic data, or questionnaires. In large retrospective evaluations across multiple independent cohorts, ZeBRA-ADRD demonstrates strong and stable discrimination for predicting incident ADRD up to ten years before first recorded diagnosis, with consistent performance across age, sex, race, and baseline risk strata. The framework includes a noise-corrected attribution mechanism (Λ-OR) that enables interpretable population-level risk analysis aligned with the clinical literature.

This repository accompanies a public demo and a documented API endpoint that allow programmatic access to ZeBRA-ADRD for research and validation use.

Single Patient Example

curl -X POST -H "Content-Type: application/json" -d '[{"patient_id": "000012", "sex": "M", "age": 89, "birth_date": "01-01-1921", "fips": "35644", "DX_record": [{"date": "01-05-2012", "code": "G35"}, {"date": "02-02-2012", "code": "H35.359"}, {"date": "03-29-2012", "code": "G35"}, {"date": "04-05-2012", "code": "R94.09"}, {"date": "04-05-2012", "code": "G35"}, {"date": "06-21-2012", "code": "G35"}], "RX_record": [], "PROC_record": [{"date": "03-29-2012", "code": "72170"}]}]' "https://us-central1-pkcsaas-01.cloudfunctions.net/zebra-predict-adrd2026?api_key=63d87942b4a2d61985a377db05a35e73"

Multiple patient data in file

curl -s -X POST -H "Content-Type: application/json" -d @sample_input.json "https://us-central1-pkcsaas-01.cloudfunctions.net/zebra-predict-adrd2026?api_key=63d87942b4a2d61985a377db05a35e73"

Input Format

Input is provided as a JSON file containing a list of patient records. Each patient record is represented as a JSON object with the fields below.

• patient_id - numeric value, can be deidentified ID.
• birth_date - Format: "MM-DD-YYYY". Does not have to be exact date of birth, but has to be approximate enough to estimate patient's age.
• sex (required) - "M" and "F" values are accepted.
• DX_record - A list of dictionaries, each containing a date of diagnosis in "MM-DD-YYYY" format (date), and a diagnostic code in ICD-10 or ICD-9 format (code).
• RX_record - A list of dictionaries, each containing a date of prescription in "MM-DD-YYYY" format (date), and a prescription code in NDC (National Drug Code) format (code).
• PROC_record - A list of dictionaries, each containing a date of procedure in "MM-DD-YYYY" format (date), and a procedural code in CPT, HCPCS, or ICD format (code).

Example Input (valid JSON)

[
    { 
        'patient_id': 33443873802,
        'birth_date': '01-01-2006',
        'sex': 'F',
        'DX_record': [
            {'date': '07-31-2006', 'code': 'Z38.00'},
            {'date': '08-07-2006', 'code': 'Z00.129'},
            {'date': '08-07-2006', 'code': 'P59.9'},
            {'date': '08-29-2016', 'code': 'J01.90'}
        ],
        'RX_record': [
            {'date': '10-29-2011', 'code': '61168010101'},
            {'date': '05-16-2015', 'code': '00071439703'},
            {'date': '08-08-2015', 'code': '00005501523'},
        ],
        'PROC_record': [
            {'date': '02-05-2007', 'code': '90723'},
            {'date': '11-05-2007', 'code': 'J1100'},
            {'date': '11-05-2007', 'code': '99214'},
        ]
    }
]

Example output for one patient:

[
    {
        'patient_id': 33443873802,
        'predicted_risk': 0.00353191883909529,
        'error_code': '',
        'probability': 0.9696941114273572
    }
]

Input Data Requirements

• Each record must span at least 52 weeks between the earliest and latest diagnosis dates in DX_record. Records with shorter diagnostic histories are not assessed in order to ensure sufficient longitudinal context for feature construction.