Diabetic: Multi-Layer Metabolic Intelligence Specification

This document defines the five data layers used by the digital-twin and CNN prediction path. They isolate Static Traits, Environmental Forcing, Human Agency, Systemic Error, and Subjective Truth.

---

🏗️ Layer 1: The Bio-Basal Vessel (Static & Predictable)

Focus: Personal baseline and real-time biometric telemetry.

Feature	Type	Frequency	Rationale
Blood Glucose (BG)	Continuous	5-min	Core signal for temporal pattern extraction.
Heart Rate (HR-BPM)	Continuous	1-min	Real-time physiological stress/arousal index.
Age	Predictable	Yearly	Shifts metabolic baseline and ISF/CSF parameters.
Weight / Height	Fixed/Manual	Monthly	Biological mass for volume-of-distribution math.
Ethnicity / Nationality	Static	Constant	Genetic/Cultural metabolic traits (e.g. higher T1D risk groups).
GENDER	STATIC	Constant	Genetics, determine hormone
---

🌊 Layer 2: The Adaptive Regimes (Uncontrollable Cycles)

Focus: External and internal forced oscillations.

Feature	Type	Cycle	Rationale
Climate (Temp/Hum)	Stochastic	Hourly	Affects subcutaneous insulin absorption rates.
Air Quality (AQI)	Stochastic	Hourly	Pollution-driven systemic inflammation / resistance.
Biphasic Hormones	Sinusoidal	24h & 28d	Male/Female hormonal resistance waves.
Seasonality	Slow Wave	Yearly	Broad shifts in insulin sensitivity (Summer vs Winter).
Regime Detection	Computed	Adaptive	Real-time resistance gain (Normal vs Sick/High Stress).

---

🎮 Layer 3: The Behavioral Engine (Controllable Agency)

Focus: User-driven inputs and lifestyle choices.

Feature	Type	Source	Rationale
Dietary Intake	Discrete	Manual/App	Carb counts, GI index, and nutritional density.
Drink/Hydration	Discrete	Manual	Impacts blood viscosity and glucose concentration.
Insulin Dosing	Discrete	App/Pump	Direct pharmacodynamic dropping events.
Sleep Quality	Periodic	Sensor	Circadian recovery state and cortisol clearing.
Exercise Intensity	Stochastic	HR-Relay	Aerobic vs Anaerobic glucose utilization.

---

🎯 Layer 4: The Meta-Correction Layer (Audit & Feedback)

Focus: Systemic self-awareness and forensic back-audit.

Feature	Type	Source	Rationale
Model Residuals	Analytic	Back-Audit	Tracks the error delta between validated forecasts and actual outcomes.
Sensor Integrity	Diagnostic	Signal Jitter	Detects compression artifacts or sensor end-of-life failures.
Metabolic Inertia	Longitudinal	History	Long-term HbA1c/Fructosamin drift (3-month baselines).
Confidence Score	Computed	Ensemble	A meta-index defining how much to "trust" the Layer 1-3 synthesis.

---

🔄 Layer 5: The Interaction & Calibration Layer (RLHF)

Focus: Subjective truth and user-labeled ground truth.

Feature	Type	Source	Rationale
User Flags	Subjective	UI / Manual	Marking "False Alarms" or "Missed Faints" to tune thresholds.
Symptom Mapping	Event	Survey	Linking L1-2 signal patterns to felt symptoms (fog, dizziness).
Threshold Bias	Control	App Settings	Customizing the aggressive vs. conservative alert sensitivity.

---

Canonical glucose units

All five layers exchange glucose internally as mmol/L. Provider adapters normalize at ingestion; model features, forecasts, thresholds, and audit state never switch units. PREFER_MMOL affects presentation formatting only and cannot alter a safety decision.

🧠 Model Ensemble Strategy

1. Digital Twin (Behavioral): Simulates meal, insulin, and basal effects.
2. 1D-CNN (Signal): Recognizes Layer 1 temporal signatures when validated weights and real cardiac telemetry are available.
3. Meta-Oracle (Layer 4): Estimates circadian basal drift.
4. Safety Gate (Layer 5): Applies conservative confidence and physiology constraints before an alert is considered.