Skip to content

RESEARCH · PRE-PRODUCT

Clinical Diabetes AI

A hybrid, multi-agent pipeline that coordinates a six-step analysis of the recommended treatment plan in any patient with T2DM — an expert-driven recommendation engine. Pre-product research, internal validation completed; prospective validation underway and manuscript scheduled.

Clinical Diabetes AI · Patient Data step 2 of 2 (Greek UI). Left panel lists current drug-class chips with insulin and combination products flagged in orange; the selected medication 'Metformin 1000mg twice daily' sits below. Right panel exposes the guardrail surfaces — diabetes-related complications, comorbidities, contraindications (with a warning icon) and observations — for the supervising physician to set before the recommendation engine runs.

Clinical Diabetes AI · physician input view with the guardrail surfaces (contraindications, comorbidities, complications) exposed before the recommendation engine runs. No PII; sample medication values shown.

Ethics: 424 GMH IRB · Αριθμ. Πρωτ. 47266 · Research-stage prototype · not for clinical use.

Protocol summary

The pipeline is registered on the Open Science Framework (OSF, DOI 10.17605/OSF.IO/9W34Z) and protocol-approved at the 424 GMH — see Ethics & governance, opposite.

Methods

Inputs. A structured T2DM patient record: demographics, duration, current regimen, relevant labs (HbA1c, eGFR), comorbidities, contraindications, and patient's preferences. No imaging, no continuous-monitor streams, no external EHR pull.

Guideline basis. The recommendation engine is grounded against the most recent ADA Standards of Care and the EASD position statements on glucose-lowering treatment in Type 2 diabetes.

Pipeline. Six coordinated agents/steps, with a shared state:

  1. Triager — filters cases by complexity.
  2. Archetyper — classifies clinical phenotype.
  3. Pinpointer — weights treatment goals.
  4. Guardian — validates against contraindications.
  5. DoseOptimiser — checks titration headroom.
  6. Verdicter — fires a Scoring Engine that executes the Rule-Based DSL to generate candidate medications via a custom Parsimonious Greedy Selection Algorithm, and synthesises the final narrative.

What's next

The algorithm was frozen and prospective-validation onboarding began in Q2 2026. The validation cohort runs through Q4 2026, and a public protocol + interim-results paper is targeted for Q1 2027. See Governance for the regulatory framing.