Abstract: Managing Type 2 Diabetes (T2D) requires interpreting noisy, non-linear patient-generated health data. This paper presents Attributable Components Analysis (ACA), a mathematical framework based on optimal transport theory, to model glycemic response. ACA significantly outperforms linear regression by capturing non-linear biological trends and maintaining robustness against outliers, providing a foundation for interpretable, personalized decision support.
Introduction
The rise of mobile health apps has enabled patients with chronic conditions like Type 2 Diabetes to generate vast amounts of self-monitoring data. However, translating these manual logs into actionable clinical insights remains a challenge. Patient-generated health data (PGHD) is frequently plagued by manual entry errors and outliers. Furthermore, the human body's response to nutrition is inherently non-linear, making simple linear models inadequate for precise personalized guidance.
The ACA Framework
To address these challenges, the authors propose Attributable Components Analysis (ACA). Rooted in optimal transport theory, ACA models the joint distribution of data rather than simply fitting a line to points. It decomposes the target variable—glycemic impact—into a sum of functions representing individual macronutrients (carbohydrates, fats, and proteins). This approach allows for "marginalization," where the complex multi-dimensional model can be projected onto 2D plots for human interpretation, showing how a single nutrient affects glucose while controlling for others.
Data and Experimental Setup
The study utilized real-world data from 16 T2D patients who logged meals and blood glucose levels over 4 to 12 weeks. To ensure data quality, a Registered Dietitian cross-verified meal logs against the USDA food composition database. The features analyzed included carbohydrates, fat, protein, fiber, and pre-meal blood glucose levels, with the target being "Glycemic Impact" (Post-meal BG minus Pre-meal BG).
Key Results
ACA demonstrated a significant performance leap over Multiple Linear Regression. For the full model, ACA achieved a Root Mean Squared Error (RMSE) of 4.36, nearly seven times lower than the linear baseline's 29.15.
Beyond raw accuracy, ACA's primary strengths lie in its qualitative performance:
- Non-linearity: ACA successfully identified the "bends" in glycemic response, such as the plateauing effect of high carbohydrate intake, which linear models are mathematically incapable of representing.
- Robustness to Human Error: When faced with erroneous data (e.g., an accidental entry of 50g of fiber instead of 5g), ACA remained stable. While linear regression produced nonsensical trends by "chasing" the outlier, ACA recognized the lack of data density in that region and effectively ignored the error.
- Expressive Uncertainty Estimates: ACA provided broader, more realistic confidence intervals. While linear regression was often overconfident (capturing only ~10% of actual data points), ACA's intervals captured 25–62% of data points, offering a more honest assessment of predictive uncertainty vital for clinical settings.
Discussion and Future Work
While ACA is more data-intensive than linear models—requiring approximately three times the data to reach peak performance—the gain in interpretability and accuracy is substantial. There is a noted tradeoff between the accuracy of the full model and the simplified, marginalized versions used for human decision support. Future research will focus on scaling this approach to larger populations and integrating causal inference to move from identifying associations to providing prescriptive dietary recommendations.
Conclusion
By bridging the gap between complex machine learning and interpretable linear models, ACA provides a robust framework for turning messy patient data into personalized medical insights. This work represents a critical step toward reliable, automated decision support for chronic disease management.
Summary updated