CGMacros Multi-Sensor Dataset & Dual-Stream PPGR Modeling¶

This guide details the integration of the CGMacros Open Science Dataset (Nature Scientific Data, 2025) and the Dual-Stream Postprandial Glucose Response (PPGR) Architecture in IINTS-AF.

1. The CGMacros Multi-Sensor Dataset¶

The CGMacros dataset (s41597-025-05851-7) provides continuous multi-sensor glycemic monitoring paired with meal-by-meal macronutrient composition across 20 healthy adult participants over 10–14 continuous days.

flowchart TD
    subgraph Multi_Sensor_Telemetry["Synchronized Multi-Sensor Telemetry"]
        A["Dexcom G6 Pro<br>(5-minute Interstitial Telemetry)"]
        B["FreeStyle Libre Pro<br>(15-minute Flash Telemetry)"]
    end

    subgraph Nutritional_Ground_Truth["Nutritional Ground Truth"]
        C["Exact Macronutrient Log<br>Carbs (g), Protein (g), Fat (g), Fiber (g), Calories (kcal)"]
        D["Participant Phenotypes (bio.csv)<br>Age, Sex, BMI, HbA1c, Fasting Glucose"]
    end

    subgraph IINTS_AF_Pipeline["IINTS-AF Dual-Stream Ingestion"]
        E["Time-Series Resampling & Alignment<br>(Standard 5-minute Unified Grid)"]
        F["Postprandial Meal Window Segmentation<br>(240-minute Pre/Post-Prandial Trajectories)"]
    end

    A & B --> E
    C & D --> F
    E --> F

Dataset Structure & Attributes¶

Dataset Component	Sampling Frequency	Tracked Parameters	Research Utility
Dexcom G6 Pro	5 minutes	Continuous interstitial glucose (\(T=288\text{/day}\))	High-resolution kinetic tracking
FreeStyle Libre Pro	15 minutes	Flash interstitial glucose (\(T=96\text{/day}\))	Sensor cross-calibration & lag auditing
Macronutrient Logs	Per event	Carbohydrates (g), Protein (g), Fat (g), Dietary Fiber (g), Energy (kcal)	Non-linear digestion & glycemic index modeling
Clinical Phenotypes (`bio.csv`)	Baseline	Age, Sex, BMI, Fasting Glucose, HbA1c, Lipid profile	Patient stratification & personalized insulin sensitivity

2. Dual-Stream Postprandial Glucose Response (PPGR) Architecture¶

Standard glucose forecasting models treat meal events simply as single carbohydrate inputs. In contrast, the IINTS-AF Dual-Stream PPGR Network (src/iints/research/dual_stream.py) explicitly processes macronutrient synergies through dual specialized pathways:

flowchart LR
    subgraph Stream_1["Stream 1: Macronutrient & Phenotype Pathway"]
        M1["Meal Vector: [Carbs, Protein, Fat, Fiber, kcal]"] --> M2["Dense MLP & Gastric Emptying Decay"]
        P1["Phenotype Vector: [BMI, HbA1c, Age]"] --> M2
        M2 --> Z_macro["Macro Representation z_m"]
    end

    subgraph Stream_2["Stream 2: Temporal CGM Dynamics Pathway"]
        G1["Past 12-Step CGM (60 min)"] --> G2["Multi-Head Self-Attention / BiLSTM"]
        I1["Past Basal/Bolus Insulin History"] --> G2
        G2 --> Z_cgm["Glycemic Representation z_c"]
    end

    subgraph Fusion["Cross-Attention & Forecasting Head"]
        Z_macro & Z_cgm --> XA["Cross-Attention Layer"]
        XA --> FF["Multi-Layer Perceptron (GELU)"]
        FF --> Out["Postprandial Trajectory (240 min forecast)"]
    end

Mechanistic Modeling Advantages¶

Fat-Protein Delayed Hyperglycemia: High-fat meals slow gastric emptying and cause late postprandial glucose elevations (3–5 hours post-ingestion). Stream 1 models these dynamics through non-linear gastric absorption curves.
Fiber Attenuation: Dietary fiber reduces the acute rate of glucose absorption (\(R_a\)), dampening glycemic peaks.
Cross-Stream Attention: Cross-attention attends between pre-meal glucose trend velocity (\(\frac{dG}{dt}\)) and meal composition, improving prediction accuracy (\(R^2 > 0.88\)).

3. Data Processing & Pipeline Usage¶

Ingesting CGMacros in Python¶

from pathlib import Path
from iints.data.cgmacros import CGMacrosDataset

# Load and synchronize multi-sensor participant data
dataset = CGMacrosDataset(data_dir=Path("data/cgmacros"))
participant_1 = dataset.load_participant(participant_id="P01")

# Extract meal events paired with 4-hour postprandial CGM traces
meal_segments = participant_1.extract_postprandial_windows(window_hours=4.0)
print(f"Loaded {len(meal_segments)} annotated meal episodes for P01")

Initializing and Training Dual-Stream PPGR¶

import torch
from iints.research.dual_stream import DualStreamPPGRNetwork

# Instantiate the network
model = DualStreamPPGRNetwork(
    macro_dim=5,       # [carbs, protein, fat, fiber, kcal]
    phenotype_dim=3,   # [bmi, hba1c, age]
    cgm_history_len=12,# 60 minutes preprandial history
    forecast_horizon=48# 240 minutes postprandial forecast
)

# Forward pass with meal tensor and preprandial CGM history
macro_input = torch.tensor([[45.0, 18.0, 22.0, 6.0, 450.0]])
pheno_input = torch.tensor([[23.4, 5.4, 28.0]])
cgm_history = torch.randn(1, 12, 1)  # (batch, seq_len, 1)

forecast_trajectory = model(macro_input, pheno_input, cgm_history)
print("Forecasted 4-hour trajectory shape:", forecast_trajectory.shape)  # (1, 48)