Patient and public involvement

A Community Advisory Board of 11 persons from the three sites including diversity in race and ethnicity as represented in the AI-READI sample contributes to the development of the protocol.

Design

AI-READI is a cross-sectional study whose target enrolment is 4000 people aged 40 and older, triple-balanced by self-reported race/ethnicity, T2DM presence and severity, and biological sex (figure 1). Building balanced data sets is critical for the development of unbiased machine learning models, so rather than targeting the demographic distribution of the US population, the study is triple-balanced by recruiting the following populations in equal proportions: four race/ethnic groups (Asian, black, white, Hispanic), four categories of T2DM (no diabetes, pre-diabetes/lifestyle-controlled diabetes, diabetes treated with oral medications or non-insulin-injectable medications, and insulin-controlled diabetes) and biological males and females. Participants are recruited across three data collection sites in different geographical locations: Birmingham, Alabama (University of Alabama at Birmingham (UAB)), San Diego, California (University of California, San Diego (UCSD)), and Seattle, Washington (University of Washington (UW)). All study groups are recruited from each site to capture diversity, but the proportion of each group will vary depending on the demographic prevalence of each group in the site’s geographical area. Participants are required to speak, read and understand English. Pregnancy and type 1 diabetes are exclusionary for participation.

• Download figure
• Open in new tab
• Download powerpoint

Figure 1

This project is generating an accessible, shared data set that includes a diverse set of health and behavioural domains and is harmonised across all variables to be artificial intelligence and machine learning ready. Descriptions of variables are available in tables 1–4. BMI, body mass index; OCTA, optical coherence tomography angiography.

Source population

The study base is all patients aged ≥40 years of age who had a medical encounter within each health system site (UAB, UCSD, UW) between 2020 and 2025. Patients with T2DM and pre-diabetes are identified by screening electronic health records for ICD-10 diagnosis codes R73.09 and E11.X, respectively. Patients without diabetes will not have encounters with these ICD codes.

Recruitment

Enrolment began on 18 July 2023 and will continue until 30 November 2026. Participants are recruited in waves to facilitate the efficient sampling of the study base. The composition and size of each wave are influenced by the observed participation characteristics of race/ethnicity, gender, severity of T2DM and site, according to health records. As recruitment progresses, the composition of participants is monitored in terms of diversity and inclusion and will be adjusted by under- and oversampling groups as needed. Recruitment in waves also allows sufficient time for coordinators to respond to persons who are interested and follow-up on mailings with expediency. For each recruitment wave, a contact pool is identified by screening electronic health records at each site. Individuals in each pool are mailed a hardcopy invitation letter and sent an invitation email, both personalised to direct them into our online REDCap recruitment interface using links, access codes and QR codes. Once in the REDCap interface, individuals can read an overview of the research programme, expectations for participation and answers to Frequently Asked Questions. They are also given the opportunity to download the informed consent document, request a call back from study staff, complete a screening survey for qualification and enroll by signing our electronic consent. Once enrolled, they have access to study questionnaires (see below). Those who do not have access to the internet may call research staff for alternative methods of enrolment.

Protocol

The protocol is performed as a single visit lasting between 2.5 and 4 hours, depending on the participant. Participants are volunteers; therefore, there is selection bias known as volunteer bias which may limit the generalisability of the results among those who are not represented in the study population. For most participants, informed consent is performed remotely by computer, tablet or smart phone through a link in the hardcopy invitation letter or the invitation email which directed them into our REDCap database. Once enrolled, participants are presented with questionnaires for online completion. If participants did not select the option to consent electronically, it is performed in-person at the start of the visit, followed by the completion of all questionnaires. All research personnel at the three data sites who recruited, enrolled and collected data underwent training by the data manager at UAB and a detailed Manual of Procedures (MOP) for reference. Successful completion of a certification process is required for all coordinators, which involved mandatory institutional compliance training for human subject research. In addition, prior to enrolling and testing participants, all coordinators are required to successfully administer all protocol elements to at least three volunteer practice subjects, meeting the standards outlined in the MOP. Data managers at each site oversee the process. Preliminary pilot enrolment occurred between 18 July 2023 and 30 November 2023 in order to secure sufficient familiarity for all coordinators. The formal data collection process began on 1 December 2023. Preliminary data from the pilot enrollment period were included in the first released data set released in May 2024. A subsequent version of the data set that includes all data collected in the first year of the study (up to 31 July 2024) is planned for release in November 2024. Information on how data can be accessed is available at https://fairhub.io/

Figure 1 presents information on all the data domains collected in AI-READI. Questionnaires addressed several content domains as listed in table 1. Height, weight and waist and hip circumference are measured, followed by calculation of the waist-hip ratio and body mass index. Systolic blood pressure, diastolic blood pressure and heart rate were measured twice separated by 2 min with an automatic oscillometric, medically approved device. A 12-lead ECG is performed (Philips Pagewriter TC30 Cardiograph, Amsterdam, The Netherlands) while the participant sits in a reclining chair or lies supine; the position is recorded (0°, 30°, 60° or 90°) relative to the supine position. Peripheral neuropathy is performed using the monofilament test11 assessing touch perception on both feet with shoes and socks removed. During testing, the participants’ eyes are closed while responding yes/no whether they feel the 10 g filament in three locations on each foot (10 times per location). A general cognitive screener is carried out using the Montreal Cognitive Assessment12 (MoCA), administered electronically on an iPad using the MoCA Duo Application (MoCA Cognition, Quebec, Canada). The total possible score is 30, with higher numbers representing better performance.

Visual acuity and contrast sensitivity are assessed under both photopic (daylight) conditions and mesopic (dim light) conditions. The right and left eyes are tested separately. Photopic letter visual acuity is measured with the Electronic Visual Acuity tester13 (M&S Technology, Niles, Illinois) using the instrument’s routine protocol. Mesopic acuity is measured by the participant viewing the display through a 2.0 neutral density (ND) filter which reduces the light level of the test to a mesopic level.14 Visual acuity is measured while the participant views the display under best-corrected conditions, expressed as the logarithm of the minimum angle of resolution. Lower numbers are better resolution. Photopic letter contrast sensitivity is measured with the Mars chart (Mars Perceptrix, Chappaqua, New York)15 using the routine procedure. Mesopic contrast sensitivity is assessed by viewing through the 2.0 ND filter. Contrast sensitivity is expressed as log sensitivity, with higher numbers meaning better sensitivity. Autorefraction data for each eye are obtained (spherical and cylindrical components expressed as diopter with cylindrical axis) using the autorefractor’s standard protocol (Topcon KR 800, Topcon Healthcare, Oakland, New Jersey).

Blood (non-fasting) and urine samples are collected from participants during the visit. Whole blood, blood derivatives and urine are used for clinical lab testing, summarised in table 2. Clinical Laboratory Improvement Amendments-certified laboratories local to each data collection site perform complete blood count (CBC) testing on fresh whole blood samples; all other lab tests are performed by the University of Washington Nutrition and Obesity Research Center (NORC) testing facility using stored, frozen samples. Additionally, blood derivatives are sent to the biorepository at UAB’s Center for Clinical and Translational Science (CCTS) and will be available to researchers for future ancillary studies (see table 3 for a summary and detailed discussion of the biorepository).

The retinal imaging protocol is designed to capture imaging from multiple devices. Both eyes are imaged on each participant. Table 4 details the scans and data formats used for each retinal imaging device. The imaging devices were chosen due to their different, yet partially overlapping, modalities. The specific scan types were chosen to provide widespread coverage of the retina with the overarching goal of detecting pixel-level differences in pathology. All retinal images are collected under dilated conditions, except for the Optomed (see table 4). All participant images are exported from imaging devices in their raw format and some images required conversion to a DICOM standard format prior to upload to an external AI-READI data storage site (discussed below).

At the end of their in-person visit, participants are sent home with three data monitoring devices as listed in table 4: (1) home environmental sensor, (2) Garmin fitness tracker and (3) Dexcom Continuous Glucose Monitor (CGM). Coordinators provide instructions on how to use these monitoring devices continuously for 10 days. Participants are instructed to return the devices to study staff by overnight mail (mailing materials and fees provided by the site) or in-person after 10 days. The environmental sensor and the Garmin fitness tracker were specifically chosen with participant privacy as a primary concern. These devices do not capture video or audio, are not equipped with GPS location monitoring and are not synced with participant-owned devices. Participants are masked to the results from the Dexcom CGM during the monitoring period. Participants return the monitoring devices with a data sheet indicating which wrist the fitness tracker was worn on, their dominant hand and where the environmental sensor was placed in the home.

The following results are returned to participants either as they leave the visit or later by HIPAA-compliant, encrypted email. Results from heart rate, blood pressure and visual acuity assessments are provided to the participant on an exam card before they leave the visit, along with instructions in lay language for interpreting the data and any follow-up recommendations. Following the 10-day at-home monitoring period, a Dexcom report providing average blood glucose levels (hourly, daily and overall) and an overall Glucose Management Indicator is provided by encrypted email. At the time of the yearly data release (see below), laboratory test results are sent with information on normative values by Health Insurance Portabillity and Accountability-compliant, encrypted email.

Coordinators alert participants to several incidental findings before they leave a visit. Participants are recommended to go for emergency care for the following reasons: individuals with systolic blood pressure readings of over 180 mm Hg or less than 100 mm Hg (with any symptoms of haemodynamic instability), diastolic blood pressure readings of greater than 120 mm Hg or less than 60 mm Hg (along with any symptoms of haemodynamic instability) or heart rate readings of greater than 100 bpm or less than 60 bpm (if not known to be usual for them and with any symptoms of haemodynamic instability). Retinal imaging technicians are trained to detect certain conditions that are potentially life or vision threatening (retinal detachment, tumour and optic disc oedema). If one of these conditions is suspected, immediate follow-up with the participant is performed with a referral to the emergency department (disc oedema) or referral to an ophthalmologist for immediate care (retinal detachment, tumour).

Data management

REDCap is used for collecting patient-reported data from questionnaires and the following clinical data from the visit: medications assessment, vitals, visual acuity, contrast sensitivity, monofilament test results and CBC test results. Data from the following devices are exported in their raw format to local storage: ECG (.xml), MoCA Duo application (total score, section subscores, Memory Index Score and task completion times; .csv), retinal imaging (table 4) and at-home monitoring devices (environmental sensors, CGMs, fitness trackers; Table 3). Blood and urine testing by the NORC lab is provided in .csv format. All data are mapped to applicable data standard formats, such as the Observational Medical Outcomes Partnership Common Data Model for clinical data and the DICOM format for retinal imaging. All data are uploaded at regular intervals to the AI-READI-specific data management platform called FAIRhub (https://fairhub.io/), a Microsoft Azure cloud-based platform developed for this project. For some devices, the data are transformed from its proprietary state into a standard model format prior to upload to FAIRhub (tables 3 and 4). The Data Manager at each data site is responsible for the quality control of all data before uploading to FAIRhub. All data are stored and shared using FAIRhub as ‘AI-ready’ which enables immediate AI/ML research without reformatting and preprocessing. More details about the AI-readiness of the data set are provided in the data set documentation (https://docs.aireadi.org/).

Data availability

A more detailed explanation of FAIRhub was provided previously1 and at https://aireadi.org. Deidentified data are released approximately yearly in two data sets: a controlled access set and publicly accessible set with fewer requirements for access. The controlled access data set contains all data available in the publicly accessible set, plus more sensitive data that are only available for use by scientists whose institutions have completed legal and privacy use agreements with the AI-READI research programme. The publicly accessible data set from pilot data collection was made available in May 2024. All data collected through 31 July 2024 were made available in November 2024.16 Information regarding data set availability and access may be found at https://aireadi.org/goals/data-sharing. The requirements for controlled access data sets are being currently developed by the Data Access Committee.

Biospecimen processing and repository

Blood (53 mL) is collected for clinical lab assays and biobanking purposes, and a urine specimen is also collected during the study encounter (table 2). The collection materials, sample type, specimen handling and location details are summarised in table 3. Local processing for plasma, serum and buffy coats is performed using standardised operating procedures, ensuring consistent handling of these biospecimens for subsequent biobanking and clinical lab analyses. One EDTA vacutainer is dedicated for delivery to a local clinical testing lab for CBC analysis. Whole blood, plasma, serum and urine collected for central clinical lab assays are batch-shipped to the UW NORC on a regular basis (table 3). Likewise, plasma, serum and buffy coats collected at UW and UCSD are batch-shipped to the UAB CCTS for integration into a central study-wide biobank. Genomic DNA extractions from stored buffy coats are performed at UAB CCTS. Because of the complexity in isolating and cryopreserving peripheral blood mononuclear cells (PBMCs), centralised processing of the CPT tubes occurs at the UAB CCTS. While this strategy does introduce an additional time variable for isolation and cryopreservation of PBMC from participants recruited at UW and UCSD, the ability to pre-centrifuge these tubes locally followed by overnight delivery to UAB allows for consistency in initial vacutainer handling. Finally, the PAXgene RNA vacutainers are stable for up to 72 hours after collection, allowing for overnight shipment of vacutainers from UW and UCSD to the UAB CCTS. Biobanked samples will eventually be available to scientists according to procedures and policies that are in development.

AI-READI has several strengths which make it highly suitable for AI/ML studies on T2DM. The targeted sample size of 4000 persons is the largest publicly accessible data set currently available containing many multidomain variables relevant to T2DM. Furthermore, after 1 year of enrolment, the sites are currently on schedule for reaching a final sample of 4000 persons within 3 years of data collection. The sample will be approximately balanced with respect to Asians, blacks, Hispanics and whites, a demographic improvement over many previous epidemiological and clinical trials on T2DM. A meta-epidemiological review of T2DM studies published from 2000 to 2020 indicated that samples lacked racial and ethnic diversity,17 thus making AI-READI an important step forward in addressing these demographic inequities in research. The study design process enacted ethical and equitable data collection and management practices, as well as data sharing with adherence to the FAIR principles.

Challenges will also be encountered. It is a well-documented observation in health research that recruiting black men into studies is more difficult than recruiting other populations,18 yet this population represents a disproportionate percentage of burden in T2DM.19 Factors such as medical mistrust and fear of safety stem from historical events such as the Tuskegee syphilis study.20 Researchers have identified strategies to encourage interest among black men to participate in research such as tailoring printed materials and using a personalised, participatory approach to recruitment.21 These trends for under-recruitment and for disease burden have also been reported for Hispanic22 and Asian persons.23 If sample balancing on race/ethnicity emerges as a challenge in AI-READI, we will implement new recruitment strategies to overcome the imbalance. Requiring that eligible participants speak, read and understand English may have created challenges in recruiting Hispanic and/or Asian persons. A limitation of our biorepository is that there are a finite number of samples to share with scientists interested in using them in research. Procedures for reviewing and prioritising written requests will be developed before the biorepository is complete.

In summary, AI-READI is an NIH Common Fund Bridge2AI-supported data collection effort to generate a 4000-person data set with T2DM that is suitable for AI/ML. It is triple-balanced with respect to race/ethnicity, biological sex and T2DM severity (including those who do not have T2DM). It is multimodal with over 10 variable domains that can be associated with adverse conditions in T2DM, yet it is hypothesis agnostic. AI-READI emphasises making data ‘AI-ready’ and establishing standards, best practices and guidelines for collection, preparation and sharing of the data. This approach can be viewed as an example for future efforts to develop other health data sets targeted for AI/ML. AI-READI opens the door for novel insights into understanding T2DM salutogenesis.