CGM/AID Eligibility Case-Finder for Diabetes Clinics

Concept

A run-in-a-day audit for endocrinology and primary care practices. The tool joins four data sources the clinic already holds — problem lists/diagnosis codes, medication and insulin fill history, device/supply billing records, and clinician notes mined with natural language processing — and applies the eligibility logic of the ADA's current Standards of Care chapter on Diabetes Technology [0], using the diagnosis/classification chapter [1] to define the diabetes population and the pharmacologic-treatment chapter [2] to identify insulin regimens, including non-intensive (basal-only) ones. Patients who meet device criteria but show no active CGM or AID prescription of any modality are surfaced as a clinician-reviewed shortlist, each row showing the qualifying attribute, where it was found (code, fill, or note), and a suggested next step (order, referral, or prior-auth packet). The same run produces a practice-level benchmark — the share of eligible insulin-treated patients actually on CGM/AID — that feeds quality reporting and value-based contracts. The wedge stays deliberately narrow: it does not diagnose, titrate, or manage devices; it only closes the prescription gap for patients already inside the clinic's records.

Why now

The 2026 Standards of Care chapter 7 is the ADA's current, evidence-graded recommendation set for diabetes technology, maintained by an interprofessional expert committee and refreshed annually or more frequently as warranted [0] — device-eligibility criteria are now versioned and citable enough to encode in software, yet change often enough that static lists drift. The companion 2026 chapters on diagnosis/classification [1] and pharmacologic treatment [2] supply the patient attributes needed to operationalize eligibility, and the Standards are explicitly framed as providing "tools to evaluate quality of care" [0], so an eligibility-and-uptake audit sits within the document's intended use. New evidence strengthens the premise and reshapes the method:

• The adoption gap is documented, not hypothetical. A review of CGM technology notes that although real-time CGM improves glycemic control in insulin-treated type 1 and type 2 diabetes, rates of adoption have been low [4] — precisely the prescription gap this tool exists to close.
• Structured codes alone would miss most eligible patients. In an analogous records-based cohort-finding task (identifying LDCT lung-cancer screening candidates), structured ICD diagnosis tables identified only 12.5% of ever-smokers, quantified none of their smoking histories, and contained no smoking information at all for 87.5% of patients; NLP over clinician notes recovered the cohort, validated by manual review of 10% of charts [3]. By analogy, a codes-only query for insulin use, diabetes type, or existing device status would systematically undercount. The tool therefore treats notes as a first-class data source and keeps clinician review as the backstop. Caveat: that evidence comes from a different domain (a critical-care research database, smoking history), so each deployment should run the same kind of manual chart-review validation on a patient sample before trusting the shortlist, and rows carry a "confirm at next visit" flag for patients whose device use may simply be undocumented (e.g., purchased out of pocket).
• Basal-only type 2 diabetes belongs in the eligibility net. RCT evidence summarized for the primary care setting supports CGM effectiveness in type 2 diabetes on non-intensive (basal) insulin, alongside real-world outcome comparisons [5]. The tool's insulin-regimen logic therefore includes basal-only patients, which widens the closable gap most in primary care, where such patients are predominantly managed. The ADA chapter's graded recommendations remain the authoritative criteria; [5] corroborates the primary-care need rather than replacing them.
• Device logic must be modality-agnostic. The first long-term implantable CGM appears safe and accurate, with mean absolute relative difference similar to available CGMs and A1c change possibly comparable to rtCGM — though glycemic-efficacy evidence is still maturing and the review itself calls for further clinical trial investigation [4]. The tool treats any CGM modality (real-time, intermittently scanned, or implantable) as satisfying "covered," flags only patients with none, and lists modality options with patient-selection considerations rather than endorsing one.

Open questions this iteration researched

• False negatives in EHR data — partially answered: structured data alone under-captures clinical attributes severely [3]; note-mining plus clinician review mitigates, but device use documented nowhere in the record remains a residual source, hence the confirm-at-visit flag.
• Insulin for diabetes vs other conditions — notes routinely carry indications (post-transplant, steroid-induced hyperglycemia) that codes omit [3]; NLP over notes is the proposed discriminator, pending local validation.
• Primary care vs endocrinology uptake delta — primary care is supported as a high-gap setting for basal-insulin type 2 diabetes [5]; the benchmark run measures the local delta rather than assuming it.
• Clinician workload, liability when an alerted clinician fails to act, and value-based contracts rewarding closure of the technology gap — remain open and unchanged; none of the new evidence addresses them.