Exploring approaches to computational representation and classification of user-generated meal logs

Area of Science:

• Computational nutrition
• Digital health
• Machine learning applications in healthcare

Background:

• Patient-generated health data (PGHD) offers valuable insights into dietary habits.
• Analyzing free-text meal logs presents challenges due to unstructured data.
• Nutritional goal alignment is crucial for personalized health management.

Purpose of the Study:

• To evaluate machine learning (ML) models for classifying meal-goal alignment using free-text meal logs.
• To assess the impact of domain-specific enrichment on ML model performance.
• To compare ML-based classification with individual self-assessments.

Main Methods:

• Utilized a dataset of over 3000 free-text meal logs from 114 individuals.
• Employed text embeddings (TF-IDF, BERT) and domain-specific enrichment (ontologies, ingredient parsing, macronutrient content).
• Evaluated logistic regression and multilayer perceptron classifiers against dietitian-validated "gold-standard" data.

Main Results:

• ML models achieved higher accuracy (0.726-0.841) than self-assessments (0.576) even without enrichment.
• The best ML model with enrichment reached accuracies of 0.814-0.902.
• Enrichment with parsed ingredients, food entities, and macronutrients generally improved performance across various nutritional goals.

Conclusions:

• ML effectively classifies meal-goal alignment from unstructured meal logs, surpassing individual self-assessments.
• Integrating nutrition domain knowledge significantly enhances ML model accuracy.
• ML analysis of PGHD holds promise for delivering patient-centered nutrition guidance in precision healthcare.