Prospective validation of an AI algorithm to identify adult-onset type 1 diabetes misclassification: protocol for a non-interventional multicentre study

Abstract

Introduction

Adult-onset type 1 diabetes (T1D) is often misclassified as type 2 diabetes (T2D), resulting in delayed treatment, missed opportunities for referrals to specialists and increased risk of complications including diabetic ketoacidosis. An electronic medical record (EMR)-based algorithm—originally trained on a large national EMR dataset to identify likely misclassified adult-onset T1D cases—was tested and retrained on a health information exchange (HIE) dataset from HealthShare Exchange (HSX). Promising results were achieved on historical data, particularly when using the retrained algorithm. However, its prospective validation is essential to more reliably assess its clinical utility and real-world precision in flagging high-risk patients for clinician review.

Methods and analysis

This is a prospective, multicentre, non-interventional cohort study in two HSX-member healthcare organisations (HCOs) in southeastern Pennsylvania. At the onset of the study, all adult T2D patients are scored by the algorithm analysing HIE data on relevant predictors found in the 24-month lookback period. Patients meeting a prespecified score threshold estimated in retrospective testing to yield 10% recall will be presented to designated endocrinology or primary care providers for structured chart review, attribution confirmation and guideline-concordant follow-up (including autoantibody testing where appropriate). The primary endpoint is positive predictive value for confirmed adult-onset T1D among flagged patients. Secondary endpoints characterise operational cascade metrics (attribution, provider recommendation, test ordering/results and diagnosis updates) along with 95% CIs. Exploratory endpoints will assess provider adoption, interpretability and workflow integration via structured provider interviews.

Ethics and dissemination

This study was reviewed and approved by Advarra Institutional Review Board (protocol Pro00075945). The Institutional Review Board waived patient informed consent and granted a full waiver of HIPAA authorisation for patient records, while providers were required to provide written informed consent. HSX data were accessed and shared under its member-defined use cases. Findings will be disseminated via peer-reviewed publications and conference presentations. Reporting will follow Strengthening the Reporting of Observational Studies in Epidemiology guidance for cohort studies.

STRENGTHS AND LIMITATIONS OF THIS STUDY.

• To our knowledge, this is the first prospective study to evaluate an artificial intelligence (AI) decision-support tool for identifying adult-onset type 1 diabetes (T1D) misclassification across multiple healthcare organisations.

• Algorithm outputs are assessed through structured chart reviews and standard-of-care follow-up processes, rather than solely study-driven interventions.

• The AI algorithm relies on health information exchange data, which enables deployment at-scale but may also vary in completeness across sites.

• Follow-up testing is dependent on clinician judgement and patient engagement, which may lead to incomplete capture of the study outcomes.

• Algorithm performance is evaluated at a single pre-specified risk threshold, which may be suboptimal for certain healthcare organisations and limits assessment of the algorithm’s potential across alternative decision thresholds.

Introduction

Type 1 diabetes (T1D) is often viewed as a childhood-onset disease, but a substantial proportion of cases actually present in adulthood with more than half of all new T1D cases occurring in adults.¹ In the USA, about 37% of adults with T1D report being diagnosed after age 30,² which is consistent with international findings that roughly 40% of T1D cases emerging after young adulthood are initially classified as type 2 diabetes (T2D).¹ The risk of misclassification increases with older age at onset, in part because adult-onset autoimmune diabetes can have a more indolent presentation (sometimes termed latent autoimmune diabetes in adults, or LADA) and may not immediately require insulin¹ as these patients are often assumed to have T2D based on age alone.³ T1D also represents a relatively small fraction (~5%) of total diabetes cases in adults,³ which can obscure the true prevalence of adult-onset T1D.

Misclassifying adult-onset T1D as T2D can have serious implications. Patients may be managed with oral hypoglycaemic agents or other T2D therapies that are insufficient for autoimmune beta-cell failure, delaying necessary insulin treatments and unintended risks such as worsening hyperglycaemia and, in some cases, diabetic ketoacidosis (DKA). T1D misclassification is associated with an estimated 18% increased risk of DKA.⁴ Patients who experience DKA at diagnosis have higher initial HbA1c levels, loss of residual β-cell function and protracted recovery.⁴ Recurrent DKA further increases the risk of cerebral oedema, cognitive impairment and organ damage.⁴ Each hospitalisation for DKA in the USA costs approximately US$29 981.⁵ An estimated US$1 in US$4 of direct medical spending on T1D is related to DKA management, and in those with repeated DKA, this may rise to US$1 in every US$2.⁵ Prolonged mismanagement of T1D can also lead to preventable microvascular and macrovascular complications,³ with retinopathy, nephropathy and cardiovascular events, potentially bringing additional long-term costs.

Primary care providers (PCPs) are the front-line for diagnosis, treatment and long-term management of T2D. Primary care is the dominant entry point for initial T2D diagnosis, creating risk for delayed identification of T1D and referrals to Endocrinology for insulin initiation and complex disease management. Misclassification of T1D as T2D can also be attributed to limitations in diagnostic heuristics. There is often over-reliance on a patient’s age and phenotype (assuming any older or overweight adult must have T2D). Confirmatory tests such as C-peptide levels are not routinely performed for adults with new-onset diabetes.³ A patient may suffer multiple DKA episodes over several years before antibody and C-peptide tests are finally checked.³ Until recently, there was no robust evidence-based guideline on identifying adult-onset T1D. Historically, clinical definitions of diabetes subtypes were developed around paediatric cases or obvious phenotypes.⁶ Many clinical features traditionally thought to distinguish T1D (leanness, acute presentation or ketosis) have overlap with T2D or may not be present in all T1D adults.⁶ Even when adult-onset T1D is suspected (eg, when a patient fails oral therapy), confirmation is sometimes delayed or not pursued due to cost or insurance coverage, or the misconception that the results would not change management. Recent consensus recommendations by the American Diabetes Association and European Association for the Study of Diabetes have started to advocate for routine islet autoantibody testing in adults with new-onset diabetes when T1D is even remotely suspected.⁷ Implementation of these recommendations, however, is still nascent.

With the availability of longitudinal national EMR data and health information exchange (HIE) data, and in collaboration with Breakthrough T1D (formerly known as the Juvenile Diabetes Research Foundation), IQVIA has developed an artificial intelligence (AI) algorithm to identify adult-onset T1D patients misclassified as T2D. The algorithm, originally trained on a large national EMR dataset,⁸ was retrained and retrospectively validated on HIE data captured through HealthShare Exchange (HSX), a regional HIE covering healthcare organisations (HCOs) in southeastern Pennsylvania.⁹ This allowed the algorithm to learn from HSX-specific dynamics and care practices. While retrospective validation has shown promising results in identifying misclassified cases, we have not assessed the algorithm’s clinical utility, real-world precision or usability by frontline providers.

During this study, we aim to prospectively evaluate the Adult T1D Detection algorithm for detecting misclassified cases among patients initially diagnosed with T2D. We will assess real-world impact of the algorithm on health outcomes when implemented at the point of care and evaluate its utility in supporting clinicians to consider prioritising patients for T1D assessment and potential reclassification. Providers will review algorithm-flagged cases using their own clinical judgement and, if appropriate, initiate confirmatory testing. We will also gather qualitative insights on implementation feasibility, provider trust in algorithmic support and data completeness.

Objectives

The primary objective of this study is to prospectively evaluate the precision—defined as the positive predictive value (PPV)—of the algorithm when integrated into real-world clinical practice. Specifically, the study aims to measure the proportion of patients flagged by the algorithm who are ultimately confirmed to have a true diagnosis of T1D, based on structured chart review and follow-up testing where clinically appropriate.

Additionally, the study will assess several operational and clinical process metrics, including the proportion of patients attributed to managing providers, presence of existing T1D diagnoses not captured in HIE data, provider recommendations for follow-up, and availability and results of autoantibody testing during the follow-up period. We will also gather qualitative feedback from designated providers on the interpretability, usability and workflow integration of the algorithm.

By addressing these objectives, it will be possible not only to demonstrate whether the algorithm can reliably identify misclassified individuals at the point of care but also if clinicians can effectively act on these insights within routine workflows, and if anything can be done to further improve the algorithm’s clinical workflow integration in the future. Over time, this AI-driven approach has the potential to shift diabetes management upstream, enabling earlier and more accurate diagnosis, more appropriate treatment initiation and reduced risk of preventable complications.

Methods and analysis

Study design

This is a prospective, multicentre, non-interventional study designed to evaluate the performance of the Adult T1D Detection algorithm⁹ when implemented as an additional input to standard clinical practice in a real-world setting in the USA. The study does not require any deviation from routine standard of care and is considered to pose no more than minimal risk to patients. Study oversight was performed by Advarra Institutional Review Board. The study began in January 2025 and initial protocol approval was provided on 7 May 2024, lasting through 7 May 2025. Two 1-year continuing reviews were then approved by Advarra Institutional Review Board on 28 April 2025 and 15 April 2026, lasting through 15 April 2027, to allow for a study extension to complete data analysis.

Dataset

The study will be conducted across two US HCOs located in southeastern Pennsylvania, both of which are members of HSX. The study will leverage HIE data routinely collected at the point of care in HSX-member HCOs and transmitted in real time to HSX via HL7 messages. Relevant data elements include diagnoses, treatments, laboratory results and procedures that are all used as predictors by the algorithm. A data dictionary and full list of algorithm inputs are provided in online supplementary table 1 and online supplementary list 1, respectively. Although data sharing agreements exist between HSX and its member HCOs, the completeness of data across different types of records received by HSX can vary across participating HCOs. For example, one HCO may share complete laboratory results feeds, while another may provide none. These inherent limitations of HIE data underscore why model retraining was particularly important on this dataset, as well as the need for prospective validation to assess the impact of such data gaps on clinical utilisation of the algorithm.

Study time period

The active study period at each health system will span 6 months from the initial deployment of the algorithm to the final collection of provider feedback. HIE data from the 24 months preceding the algorithm deployment will be extracted to serve as the lookback period for generating algorithm inputs. The algorithm will be deployed once at the start of the study; this deployment date will serve as the index date. Outcomes will be reviewed and recorded by designated providers within each HCO at two time points: 3 months and 6 months post-deployment. See figure 1 for an overview of the study timeline.

Figure 1. Overview of the prospective study time period in each of the two participating HCOs. HCOs, healthcare organisations; HSX; HealthShare Exchange.

High-risk patient identification and provider attribution

The algorithm will analyse the latest available data aggregated by HSX and assign a risk score to patients for being misclassified adult-onset T1D. Patients with a risk score above the 10% recall threshold—as estimated during retrospective validation⁹—will be included in the study for further evaluation. This threshold corresponds to the score cut-off at which the algorithm was estimated to identify approximately 10% of adults who truly had T1D misclassified as type 2. As described by Brusini et al,⁹ this threshold was found to provide a good balance between PPV (estimated to be 18.3% on retrospective HSX data) and screening capacity at the two HCOs participating in the present prospective study, and therefore was deemed appropriate to use for this prospective validation.

For each patient, the algorithm will also produce a set of key predictors (eg, most recent body mass index, most recent A1c result, evidence of insulin use, etc.) to support interpretability and transparency.

Each patient will be attributed to an endocrinologist or PCP who actively manages their care within the HCO, using the following hierarchical logic:

1. Most recent endocrinologist within the 12 months prior to the index date.

2. Most frequently visited PCP within the 12 months prior to the index date.

3. Most recent endocrinologist within the 24 months prior to the index date.

4. Most frequently visited PCP within the 24 months prior to the index date.

Patient-level outputs, including demographic information, key predictors and provider attribution, will be sent securely to HSX via secure file transfer protocol (SFTP) for reidentification. HSX will then distribute the reidentified lists to designated providers at each HCO in Excel format via secure email.

Assessment of high-risk patients

Designated providers will be given 4 weeks to complete a structured chart review and validate each flagged patient against information available in their EMR. This includes checking the accuracy of the managing provider’s attribution, reviewing relevant medical history and making an independent assessment of whether further follow-up or diagnostic testing (eg, pancreatic autoantibody testing) is clinically appropriate, in accordance with standard clinical guidelines.

The follow-up process will be based on the patient’s existing clinical relationships:

• If the patient is already under the care of an endocrinologist, the designated provider will direct clinical staff to arrange appropriate follow-up, such as autoantibody testing or alternative diagnostic assessments.

• If the patient is not under the care of an endocrinologist but has an assigned PCP, the designated provider will notify the PCP of the algorithm output and recommend follow-up via existing EMR workflows.

• If the patient lacks a visible relationship with either an endocrinologist or PCP in HSX data, the designated provider will confirm this through chart review. If no suitable attribution is identified, the patient will not undergo further evaluation.

Following the chart review, providers will record their assessment and actions in the patient list Excel file, including the following:

• Whether follow-up for T1D screening was recommended.

• Rationale for not recommending follow-up (if applicable).

• Whether a T1D diagnosis was already present in the chart.

• Whether follow-up with an endocrinologist or PCP was advised.

Test results will be reviewed by the ordering provider, who will also be responsible for updating the patient’s EMR problem list or encounter diagnoses as appropriate. At the 3-month and 6-month follow-up points, designated providers will review the same cohort of patients and document outcomes. These updated Excel files will be securely returned to HSX via email, deidentified by HSX and transferred to IQVIA via SFTP for analysis.

Primary and secondary endpoints will be assessed upon study completion, based on aggregated findings from these provider-completed files.

Sample selection

Study population and sample size justification

A formal sample size calculation is not required for this prospective, non-interventional study. The objective is to estimate the PPV of the algorithm when deployed in routine clinical practice, not to test a prespecified hypothesis. As such, the focus is on feasibility and estimation of precision rather than statistical power.

Based on retrospective data from HSX, it is anticipated that the algorithm will flag approximately 150–230 patients per HCO using a 10% recall threshold. This threshold reflects a pragmatic balance between clinical manageability and algorithm sensitivity. The resulting total sample—approximately 300–460 patients across two HCOs—is aligned with the operational capacity of designated providers at each site to conduct structured chart reviews and follow-ups over the 6-month study period. All patients will already be under the care of providers within the participating HCOs, and therefore, no explicit patient consent or enrolment will be required.

In addition to the primary analysis, two designated providers per HCO (four in total) will be enrolled to participate in structured interviews. These interviews will capture qualitative feedback to inform exploratory endpoints related to algorithm interpretability, usability and integration into clinical workflow. The number of participating providers is expected to yield sufficient insight into clinician experience and contextual factors influencing adoption.

Inclusion and exclusion criteria

Patients must meet all the following conditions during the 24-month lookback period to be included in the cohort evaluated by the algorithm:

• At least one diagnosis of T2D.

• Age ≥18 years as of the index date (date of algorithm deployment).

Patients will be excluded from algorithm scoring if they meet any of the following conditions within the same 24-month lookback period:

• Prior diagnosis of T1D or evidence of a positive pancreatic autoantibody test.

• Any of the following clinical conditions that may impact pancreatic function or mimic autoimmune diabetes: acromegaly, chronic pancreatitis, Cushing’s disease, cystic fibrosis, gestational diabetes, maturity-onset diabetes of the young, history of pancreatectomy, use of pancreatic enzyme replacement therapy (eg, Creon).

After applying these criteria, the algorithm will be run on the eligible population to generate a ranked risk score for each patient based on the likelihood of misclassification. A consecutive series of all eligible patients above a prespecified score threshold (10% recall) will be selected for further chart review and clinical assessment by the designated providers.

Two designated providers per HCO will be enrolled to evaluate the patients flagged by the algorithm. Inclusion criteria for providers are as follows:

• Clinical specialty in endocrinology or primary care.

• Active clinical practice at the participating HCO.

• Willingness to participate in the study and provide structured chart reviews and feedback.

No formal exclusion criteria apply to providers beyond those implied by the above.

Interventions to be measured

The intervention under evaluation is the algorithm-generated list of flagged patients, shared with designated providers in each HCO. The algorithm itself is not a prescriptive tool and does not dictate diagnostic or therapeutic actions. It serves as a decision-support input, identifying patients who may benefit from additional review and confirmatory testing based on standard clinical guidelines. No study-specific treatments will be administered, and no investigational medication will be involved.

Outcomes

Outcome measures will be evaluated for the primary endpoints based on data collected via the patient list Excel file completed by designated providers during the follow-up period. All outcomes will be assessed at the patient level and reflect real-world provider responses to the algorithm’s output (see table 1).

Table 1. Outcomes and associated endpoints definitions.

Outcome	Endpoint	Definition
Presence of patient attribution to endocrinologist or PCP	Count/percent of patients at the selected level of recall attributed to an endocrinologist or PCP	Of the patients identified above the 10% recall cut-off, proportion attributed to an endocrinologist or PCP based on attribution hierarchy
Presence of previously confirmed T1D diagnosis not visible in HSX data	Count/percent of patients at the selected level of recall with a confirmed T1D diagnosis previously recorded in the EMR but not captured in HSX data	Among attributed patients, those with a T1D diagnosis recorded in the local EMR during chart review but not reflected in HSX data
Presence of recommendation for follow-up by the designated provider	Count/percent of patients at the selected level of recall suspected for T1D and recommended for follow-up by the provider	Among attributed patients, those for whom the provider recommended additional diagnostic evaluation for possible T1D
Presence of autoantibody test order during follow-up period	Count/percent of patients at the selected level of recall for whom autoantibody testing was ordered within the follow-up period	Among attributed patients, those with at least one order for GAD, IAA, IA-2A or ZnT8 autoantibody testing following chart review
Presence of autoantibody test result during follow-up period	Count/percent of patients at the selected level of recall with an autoantibody test result available during the follow-up period	Among attributed patients, those with autoantibody test results available in the local EMR within the 6-month follow-up period
Presence of T1D diagnosis or 2+ positive autoantibody results during follow-up	Count/percent of patients at the selected level of recall who received a T1D diagnosis or had ≥2 positive autoantibody test results during follow-up	Among attributed patients, those either newly diagnosed with T1D or showing at least two positive autoantibody results during the follow-up period
Presence of confirmed T1D, whether from chart history or follow-up	Count/percent of patients at the selected level of recall who are true T1D patients	Among attributed patients, those with either a previously unrecorded T1D diagnosis found during chart review or who met the criteria for new T1D confirmation during follow-up (≥2 positive autoantibody results or newly documented T1D diagnosis)

EMR, electronic medical record; GAD, glutamic acid decarboxylase; HSX, HealthShare Exchange; IAA, insulin autoantibody; PCP, primary care provider; T1D, type 1 diabetes; ZnT8, zinc transporter 8.

Autoantibody results will be assessed using thresholds drawn from test-specific standards, with a positive result defined as follows:

• Glutamic acid decarboxylase 65: text result of ‘positive’ or numerical value >0.02 IU/mL.

• Insulin autoantibody: text result of ‘positive’ or numerical value >0.02 IU/mL.

• Insulinoma-associated antigen-2 autoantibody: text result of ‘positive’ or numerical value >0.02 IU/mL.

• Zinc transporter 8: text result of ‘positive’ or value >15 IU/mL.

These thresholds are based on standard laboratory reference ranges and validated interpretation criteria. A positive diagnosis based on autoantibody testing requires at least two different positive markers consistent with established diagnostic guidelines for adult-onset autoimmune diabetes.

Data analysis plan

Quantitative data analysis

Descriptive statistics will be used to summarise the study population and all outcome measures. Analyses will be conducted separately for each participating HCO and then aggregated across sites to characterise the full cohort.

Continuous variables (eg, age) will be summarised using means and SD, or medians and ranges where appropriate. Categorical variables (eg, provider attribution, test ordering, confirmed diagnosis) will be reported as counts and percentages. Subgroup analyses may be presented where relevant (eg, by provider type or presence of endocrinology relationship), although no formal stratification is planned.

For each primary outcome, the proportion of patients meeting the relevant endpoint will be calculated, and 95% CIs will be estimated using binomial or Poisson distribution methods, depending on the data structure. No formal hypothesis testing is planned, as the study is exploratory in nature and intended to estimate the real-world PPV of the algorithm in practice.

Qualitative data analysis

Designated providers will participate in two structured interviews to gather qualitative insights into their experience acting on the algorithm’s outputs. These interviews will support exploratory objectives, including assessment of the algorithm’s clinical interpretability, usability and integration into routine care.

• The first interview (approximately 30 min) will occur 2 weeks after the providers receive the patient list. It will focus on initial impressions, time requirements for chart review and integration into existing clinical workflows.

• The second interview (approximately 60 min) will occur 2–3 months after the start of the study. It will explore in greater depth the utility of algorithmic recommendations, data completeness, barriers to follow-up and provider trust in model predictions.

Interview topics will include, but are not limited to the following:

1. Time spent reviewing EMRs for flagged patients.

2. Patterns in model outputs associated with presence or absence of endocrinology relationships.

3. Accuracy of structured diagnosis data (problem list and encounter diagnoses), and how this varies by provider type.

4. Presence or absence of appropriate diagnostic follow-up for flagged patients.

5. Data elements available in local EMRs but not captured by HSX, and their perceived impact on risk stratification.

6. Common features of patients who were correctly identified as likely T1D.

7. Common features of patients incorrectly identified by the algorithm.

8. Instances of provider disagreement with algorithm output and associated clinical reasoning.

9. Alignment of disagreements with personal clinical judgement, clinical guidelines or limitations in structured/unstructured data.

All interviews will be audio-recorded and detailed notes will be captured. These will be stored securely and reviewed by the study team to identify recurring themes. Responses will inform the interpretation of study findings and highlight areas for future improvement in algorithm deployment.

Handling of missing data

Missing data will be minimised through structured chart review workflows and defined provider responsibilities at each follow-up point. Key mitigation strategies include the following:

• Requiring providers to confirm algorithm inputs and attribution details at initial review.

• Instructing providers to document outcomes (including test results and diagnosis updates) during both follow-up assessments.

• Flagging any patient-level data discrepancies or absence of follow-up for resolution where possible.

Due to known limitations in HIE data completeness across sites, providers will be asked to identify any discrepancies between HSX and local EMR data (eg, historical diagnoses not visible to the algorithm). While outcome data are expected to be complete for most patients, any missingness will be documented and summarised descriptively; imputation methods will not be used.

Anticipated limitations

This study is subject to several limitations that may affect the interpretation and generalisability of its findings:

• The study population and participating providers are engaged with healthcare services solely at the two participating HCOs and represent a specific geographic footprint of southeastern Pennsylvania. Individuals with limited or no healthcare access, or those outside of these two HSX-affiliated institutions, will not be captured. As such, findings may not be generalisable to a broader population, so external validation of the predictive model beyond this study will be necessary to assess factors impacting scalability across a variety of regions and practice types.

• As the algorithm is intended to serve as a clinical decision-support tool rather than a directive, designated providers will retain full discretion in determining whether follow-up testing or referrals are warranted. Consequently, some patients flagged by the algorithm may not undergo further diagnostic workup, which may limit the ability to evaluate true PPV and influence downstream outcomes measurement.

• Variation in the types and completeness of clinical data shared with HSX across HCOs may affect both algorithm performance and outcome ascertainment. For example, some sites may have limited laboratory or encounter data availability. To mitigate this, the algorithm was trained using HSX data to account for site-specific data patterns. Providers will also be instructed to supplement missing HSX information with chart review data from local EMRs.

• It is possible that patients flagged by the algorithm may already have a confirmed T1D diagnosis documented in the local EMR that is not visible in HSX data. To address this, designated providers will be asked to review each patient’s chart and flag any previously unrecognised diagnoses before proceeding with further evaluation.

• A subset of patients may not be attributable to either a PCP or endocrinologist within the HCO based on available data, thereby limiting the ability to follow-up or confirm outcomes. While providers will review charts to complete or validate attribution, this may not be possible in all cases.

• Due to the limitations of routinely collected HIE data, outcome ascertainment relies on structured chart review and provider-completed Excel files. This approach ensures accurate outcome capture but does not allow for comparisons with historical controls or broader baseline rates.

• Patients flagged by the algorithm may not complete follow-up for several reasons, including lack of provider capacity to initiate testing, reliance on referred providers who may not act, or patient refusal or non-attendance. These real-world barriers may lead to underestimation of the true yield of the algorithm.

• The study only includes patients identified above the 10% recall threshold. The performance of the algorithm for patients below this threshold (ie, lower-risk group) will not be assessed, which may limit insights into its full discriminative capacity.

• Differences in how individual providers interpret and act on algorithm outputs may lead to variability in observed outcomes. These contextual factors will be explored in qualitative interviews but may introduce site-level or provider-level variation not fully accounted for in the quantitative analysis.

Patient and public involvement

No plans exist for involving patients and members of the public in the design, conduct, reporting or dissemination plans of this study.

Ethics and dissemination

This study will be conducted in accordance with the ethical principles set out in the Declaration of Helsinki and its amendments, Good Pharmacoepidemiology Practices and applicable US regulatory requirements for non-interventional studies, including the Common Rule (45 CFR 46) and HIPAA. HSX data will be accessed and shared in accordance with its member-approved data use cases.

The study poses no more than minimal risk to participants and version 1.0 was reviewed and approved by Advarra Institutional Review Board (IRB#00000971) on 7 May 2024 (protocol number Pro00075945). Version 1.1 was reviewed and approved on 2 April 2026 with a modification that enabled study extension to support additional follow-up data gathering at one of the two HCOs. The IRB waived patient informed consent under 45 CFR 46.116(f), based on the following criteria:

• The research involves secondary use of routinely collected, partially deidentified EMR data.

• The algorithm is used solely as a clinical decision-support tool; it does not prescribe actions nor over-ride provider judgement.

• Patients will not be denied or diverted from standard care. Follow-up testing, when ordered, will follow usual clinical workflows.

• Requiring individual patient consent would introduce practical and scientific limitations, including selection bias and undue burden on providers and patients particularly where patients are no longer engaged with the health system.

Full waiver of HIPAA authorisation for patient records was also granted under 45 CFR 164.512(i)(2).

Only designated healthcare providers participating in chart review and interviews are required to give informed consent to participate. They will receive a written information sheet outlining their role in the study and consent will be documented prior to participation.

Patient confidentiality will be strictly protected throughout the study. HSX will assign a unique study identifier to each patient and remove all identifiable information—except date elements required for algorithm performance—prior to data transfer. All data transmitted to IQVIA will be handled as if fully identifiable and managed within a secure Azure-based environment with two-factor authentication, encrypted file transfers and access restricted by internet protocol whitelisting.

Safety monitoring is not applicable to this study, as it is non-interventional and does not involve administration of medications or medical procedures. No adverse event reporting is planned.

Findings from the study will be disseminated through peer-reviewed publications and conference presentations. All outputs will adhere to the Strengthening the Reporting of Observational Studies in Epidemiology guidelines for cohort studies. A final study report will be generated at study completion, summarising cohort characteristics, endpoint findings and qualitative insights. No patient-level data will be published. All reported results will be aggregated and anonymised to prevent reidentification.