Project and Research Question¶
Motivation¶
Automated insulin-delivery research combines software, physiology, noisy sensor data and safety-critical decisions. A controller can improve an average metric while still producing dangerous edge cases. A polished graph can also hide unrealistic rates of change, missing data, excessive interventions or a favourable choice of seed.
IINTS-AF was built to make those problems inspectable. It treats a simulation run as a complete research object: protocol, inputs, model, candidate decision, safety decision, raw outputs, validation and interpretation.
Primary research question¶
Can an open-source simulation workbench make risky or unrealistic diabetes-algorithm behaviour visible and reproducible before an algorithm is considered for real-world testing?
Hypothesis¶
An algorithm that reacts to incomplete glucose context can produce questionable actions in stress or failure scenarios. A separate deterministic safety layer, combined with reproducible scenarios and evidence artifacts, should make at least part of this behaviour detectable, blockable and reviewable.
Experimental structure¶
flowchart LR
Q["Research question"] --> B["Normal baseline"]
Q --> S["Physiology or meal stress"]
Q --> R["Risk or corrupted-input case"]
B --> C["Candidate and baseline comparison"]
S --> C
R --> C
C --> V["Deterministic validation and safety evidence"]
V --> E["CSV, metrics, reports and limitations"]
Main contribution¶
The project contribution is the research workflow around algorithms, not a claim that one controller solves diabetes:
- Define a patient, scenario, seed and algorithm.
- Run a mechanistic or transparent virtual-patient model.
- Keep the candidate algorithm separate from safety authority.
- Record candidate and delivered actions.
- Validate data and physiological plausibility.
- Compare baselines over a locked matrix.
- Export raw evidence and human-readable reports.
- Allow AI to explain the evidence without changing it.
Why this is different from one notebook¶
| Notebook-only risk | IINTS-AF response |
|---|---|
| Hidden state and manual cells | Versioned CLI/SDK workflows and fixed seeds |
| One attractive run | Scenario, patient, algorithm and seed matrices |
| Black-box output | Candidate action, safety action and intervention reason |
| Metrics without raw evidence | CSV traces, manifests, reports and validation files |
| AI-generated interpretation | Deterministic metrics first; LLM explanation second |
| Unclear data provenance | Source manifests, hashes, contracts and MDMP reports |
| Demo code disconnected from research | The demo runs the packaged SDK workflow |
Scope of the work¶
The SDK covers four connected research areas.
1. Simulation¶
Virtual patients expose latent glucose, insulin action, meal absorption and selected research extensions. A sensor model converts latent glucose into a CGM-like observation.
2. Algorithm and safety experiments¶
Baselines, experimental controllers and forecasting models can be tested, but the deterministic supervisor remains a separate authority in simulated action paths.
3. Data and AI research¶
The SDK can prepare diabetes datasets, derive context features, train glucose forecasting models and compare models with both numerical and physiological checks.
4. Evidence and communication¶
Runs can produce clinical-style reports, AGP-style assets, posters, audit artifacts and a structured evidence bundle for independent review.
Claim ladder¶
The project uses a claim ladder to prevent results from being overstated.
| Level | Claim that may be made |
|---|---|
| Implemented | The code path exists and is covered by tests or inspectable output |
| Verified in software | Expected behaviour was observed under specified test conditions |
| Calibrated in a dataset | Parameters or metrics were compared with a documented dataset |
| Externally validated | Performance was tested on an independent population or reference simulator |
| Clinically validated | Requires clinical and regulatory evidence outside the present project |
Most IINTS-AF claims are at the first two levels. Some data and realism workflows reach the calibration level. The project does not claim clinical validation.
Success criteria for EUCYS¶
The EUCYS presentation is successful when a reviewer can answer:
- What was tested?
- Which equations and parameters produced the virtual state?
- What did the candidate propose?
- Did the safety layer change it?
- Which output proves that?
- Can the run be reproduced?
- Which limitations prevent a clinical claim?