Beyond Carbs: Tackling the Taboo

Independent Research by Anthony Kerr | AI4ALL Ignite Fellow
📥 Click to view High-Resolution PDF

---

Beyond Carbs: An ML-Driven Solution for Foods That Break Standard T1D Therapy

A Predictive "Safety Net" for Type 1 Diabetes Management

Author: Anthony Kerr
Program: AI4ALL Ignite Accelerator (Portfolio Project)
Status: Phase 1 Complete (Proof of Concept)

---

📌 Project Overview

Standard Type 1 Diabetes (T1D) dosing fails on complex meals because it myopically focuses on carbohydrates. As a young man living with T1D for the past decade, I've experienced this failure countless times. This project shifts the paradigm by training a machine learning "wrapper" to analyze the critical, overlooked impact of fat, protein, and fiber.

By accurately classifying "taboo" foods (like pizza, plantains, or rich pastas) as High-Risk before insulin delivery, this model serves as a predictive safety net for stable blood sugar where current treatments—whether manual injections or automated pumps—consistently fall short.

The Core Problem

• Standard Therapy: Relies on "Carb Counting," which assumes digestion is linear and predictable.
• The Reality: High-fat and high-protein meals delay gastric emptying, causing a "delayed spike" 3–5 hours post-meal—long after standard insulin has peaked.
• The Result: A dangerous glucose "rollercoaster" and immense mental burden ("Diabetes Distress") for patients.

---

🧠 The Solution: Machine Learning "Wrapper"

Think of standard Type 1 Diabetes therapy (insulin delivery) as a car (it moves you from point A to B). This model serves as the car's GPS Navigator, ensuring patients with T1D arrive safely in any condition (bad weather, high/low traffic, low fuel, etc.) by predicting potential roadblocks and suggesting the best route per situation (insulin delivery), before they leave.

It uses a Random Forest Classification algorithm to analyze the full nutritional context of a meal. It predicts whether a specific combination of macros will cause a delayed spike that standard math misses, outputting a simple, actionable signal:

• 🟢 Low Risk: Proceed with standard dosing.
• 🔴 High Risk: Intervention required (e.g., split bolus, extended delivery).

---

📊 Data Source & Attribution

This project utilizes real-world longitudinal data from the University of Manchester.

• Dataset: T1D-UOM – A Longitudinal Multimodal Dataset of Type 1 Diabetes (V1.0.1)
• Repository: ManchesterCSCoordinatedDiabetesStudy
• Scope: 16 individuals, data collected Oct 2023 – Aug 2024.
• Key Features Used:
  • Nutritional: Carbohydrates, Fat, Protein, Fiber.
  • Physiological: Continuous Glucose Monitor (CGM) readings, Insulin (Bolus/Basal).

Attribution: I gratefully acknowledge the creators of the T1D-UOM dataset for providing the granular nutritional data necessary to train this model.

Alsuhaymi, A., Bilal, A., Gasca Garcia, D., Kongdee, R., Lubasinski, N., Hood, T., Paul, N., & Harper, S. (2025). T1D-UOM – A Longitudinal Multimodal Dataset of Type 1 Diabetes (Version 0.1.0) [Computer software]. https://github.com/sharpic/ManchesterCSCoordinatedDiabetesStudy

---

⚙️ Methodology

1. Feature Engineering

I engineered specific features to capture the physiological complexity of digestion:

• Fat-to-Carb Ratio (FCR): A key indicator of the "Pizza Effect" (delayed absorption).
• Protein-to-Carb Ratio (PCR): Indicator of sustained glucose rise.
• Insulin On Board (IOB): Calculated active insulin to prevent stacking errors.
• Glucose Rate of Change (RoC): The dynamic trend of blood sugar pre-meal.

2. Model Architecture

• Type: Supervised Learning (Binary Classification)
• Algorithm: Random Forest Classifier
• Why Random Forest? It efficiently handles the non-linear, complex interactions between fat/protein and digestion significantly better than linear regression models. It is also robust against noise in self-reported food logs.

---

⚠️ Ethical Considerations & Limitations (Responsible AI)

As part of the AI4ALL Responsible AI framework, explicitly acknowledging bias is critical:

1. Evaluation Bias (Population): The model was trained on a small cohort of 16 individuals in Manchester, UK. The training data heavily favors Western diets.
2. Cultural Inclusivity vs. Reality: While the approach (using macronutrients) is culturally inclusive, the model may underperform on culturally significant non-Western foods (e.g., high-fiber Caribbean dishes like Jamaican bammy or roti) if those specific macro-combinations were absent from the UK training set.
3. Proof of Concept: This tool is currently a Proof of Concept. It is not a medical device and should not be used for active treatment decisions without retraining on a larger, more diverse dataset.

---

🚀 Usage

Prerequisites

• Python 3.x
• Jupyter Notebook
• Libraries: pandas, numpy, scikit-learn, matplotlib, seaborn

Installation

1. Clone the repo:

   git clone https://github.com/YourUsername/AI4ALL-T1D-ML-Wrapper.git

2. Install requirements:

   pip install -r requirements.txt

3. Run the analysis notebook:
   • Open T1D_ML_Wrapper_Analysis.ipynb in Jupyter.

---

🔮 Next Steps

• Phase 2 (March 2026): Retrain model on diverse, international data to mitigate Evaluation Bias.
• Phase 3 (May 2026): Develop a web-based MVP tool for user testing.
• Long-term: Integrate with open-source artificial pancreas systems (like AndroidAPS/Loop) as a plugin "safety layer."

---

This project was created for the AI4ALL Ignite Accelerator.