Enabling personalized decision support with patient-generated data and attributable components.

Area of Science:

• Computational biology
• Health informatics
• Machine learning applications in healthcare

Background:

• Health decision-making is complex and often exceeds human cognitive capacity.
• Patient-generated data and machine learning (ML) offer potential for personalized health insights.
• Not all ML-generated information is equally useful for practical health decisions.

Purpose of the Study:

• To develop and apply Attributable Components Analysis (ACA) to identify patterns between nutrition and blood glucose control using type 2 diabetes self-monitoring data.
• To compare ACA with traditional methods like linear regression for utility in decision support.
• To explore the trade-offs between model accuracy and interpretability in ML-driven health decision support.

Main Methods:

• Developed and applied Attributable Components Analysis (ACA), a novel method inspired by optimal transport theory.
• Utilized self-monitoring data from patients with type 2 diabetes, focusing on nutrition and blood glucose levels.
• Compared ACA's performance against linear regression for identifying associations.

Main Results:

• ACA identified non-linear relationships between nutrition and blood glucose control, surpassing linear regression.
• ACA demonstrated greater robustness to outliers in the data.
• ACA provided broader and more expressive uncertainty estimates compared to linear regression.
• A trade-off between model accuracy and interpretability was observed.

Conclusions:

• ACA is a promising method for enhancing ML-driven health decision support systems, particularly for complex chronic conditions like type 2 diabetes.
• ACA's ability to handle non-linearities and provide robust uncertainty estimates improves the utility of patient-generated data.
• Further research should consider the balance between predictive accuracy and interpretability for effective clinical implementation.