Behavioral-clinical phenotyping with type 2 diabetes self-monitoring data

Matthew E. Levine, David J. Albers, Marissa Burgermaster, Patricia G. Davidson, Arlene Smaldone, Lena MamykinaarXiv 2018 · DOI · arXiv

Abstract

Objective: To evaluate unsupervised clustering methods for identifying individual-level behavioral-clinical phenotypes that relate personal biomarkers and behavioral traits in type 2 diabetes (T2DM) self-monitoring data. Materials and Methods: We used hierarchical clustering (HC) to identify groups of meals with similar nutrition and glycemic impact for 6 individuals with T2DM who collected self-monitoring data. We evaluated clusters on: 1) correspondence to gold standards generated by certified diabetes educators (CDEs) for 3 participants; 2) face validity, rated by CDEs, and 3) impact on CDEs' ability to identify patterns for another 3 participants. Results: Gold standard (GS) included 9 patterns across 3 participants. Of these, all 9 were re-discovered using HC: 4 GS patterns were consistent with patterns identified by HC (over 50% of meals in a cluster followed the pattern); another 5 were included as sub-groups in broader clusters. 50% (9/18) of clusters were rated over 3 on 5-point Likert scale for validity, significance, and being actionable. After reviewing clusters, CDEs identified patterns that were more consistent with data (70% reduction in contradictions between patterns and participants' records). Discussion: Hierarchical clustering of blood glucose and macronutrient consumption appears suitable for discovering behavioral-clinical phenotypes in T2DM. Most clusters corresponded to gold standard and were rated positively by CDEs for face validity. Cluster visualizations helped CDEs identify more robust patterns in nutrition and glycemic impact, creating new possibilities for visual analytic solutions. Conclusion: Machine learning methods can use diabetes self-monitoring data to create personalized behavioral-clinical phenotypes, which may prove useful for delivering personalized medicine.

Introduction

Managing Type 2 Diabetes (T2DM) is fundamentally a challenge of individual variability. Glycemic responses to dietary intake are highly personalized; a meal that is "safe" for one patient may cause a significant blood sugar spike in another. While traditional clinical phenotyping often occurs at the population level (e.g., distinguishing between Type 1 and Type 2), there is a critical need for individual-level phenotyping that integrates behavioral data with clinical outcomes. This paper explores the use of unsupervised machine learning to identify "behavioral-clinical phenotypes"—patterns that link specific nutritional behaviors to glycemic responses within a single individual.

Methods

Data Collection and Feature Engineering

The study utilized data from six adults with T2DM. Participants recorded meal descriptions and photos via a custom smartphone application and captured pre- and post-meal blood glucose (BG) readings. Five Registered Dietitians (RDs) then estimated the macronutrient composition (carbohydrates, protein, fat, and fiber) of these meals using the USDA National Nutrient Database. The final datasets ranged from 34 to 101 meals per participant.

The researchers employed Hierarchical Clustering (HC) to group these meals. Through iterative testing, they determined that the most effective feature set combined the absolute change in blood glucose (1–3 hours post-meal) with the absolute grams of macronutrients, rather than the percentage of total calories.

Validation Framework

To validate the unsupervised clusters, the study used two primary methods:

  1. Mathematized Gold Standard: Certified Diabetes Educators (CDEs) manually identified qualitative patterns (e.g., "High-fat meals cause late spikes"). These were translated into Boolean expressions to quantitatively measure how well the HC clusters matched expert intuition.
  2. Visual Analytics: Parallel coordinate plots were used to visualize cluster means, allowing experts to evaluate the relationships between multiple nutritional variables and glucose outcomes simultaneously.

Results

Pattern Rediscovery and Validity

The hierarchical clustering approach successfully rediscovered all nine of the initial gold-standard patterns identified by experts. Specifically, four patterns were highly consistent (where over 50% of meals in a cluster followed the expert pattern), while the remaining five were identified as distinct sub-groups within larger clusters. Furthermore, 50% of the generated clusters received high marks from medical experts (scoring >3 on a 5-point Likert scale) for face validity, clinical significance, and actionability.

Clinical "Force Multiplier" Effect

One of the most significant findings was the impact of machine learning on human expert performance. When CDEs were presented with the cluster visualizations, they were able to refine their own clinical observations. This led to a 70% reduction in contradictions between the experts' stated patterns and the actual participant data. This suggests that ML-driven phenotyping can serve as a powerful decision-support tool, helping clinicians move beyond "rules of thumb" to data-driven, patient-specific insights.

Discussion

This research highlights the importance of defining phenotypes relative to the individual in the early stages of precision medicine. By clustering behavior (nutrition) and clinical outcomes (glucose) together, the model uncovers reciprocal relationships that might be missed if these data streams were analyzed in isolation.

However, several limitations remain. The small sample size (n=6) means this study serves primarily as a proof-of-concept. Additionally, the participants had relatively stable HbA1c levels; the efficacy of this method in patients with less-controlled diabetes remains to be seen. The manual translation of qualitative expert observations into Boolean logic also presents a bottleneck for scaling this approach.

Conclusion

Unsupervised machine learning, specifically hierarchical clustering, is a viable method for identifying personalized behavioral-clinical phenotypes in T2DM. By integrating self-monitoring data with clinical biomarkers, this approach provides a pathway for more robust, personalized dietary interventions and demonstrates the potential for ML to enhance clinical expertise through visual analytics.

Summary updated

Abstract

Objective: To evaluate unsupervised clustering methods for identifying individual-level behavioral-clinical phenotypes that relate personal biomarkers and behavioral traits in type 2 diabetes (T2DM) self-monitoring data. Materials and Methods: We used hierarchical clustering (HC) to identify groups of meals with similar nutrition and glycemic impact for 6 individuals with T2DM who collected self-monitoring data. We evaluated clusters on: 1) correspondence to gold standards generated by certified diabetes educators (CDEs) for 3 participants; 2) face validity, rated by CDEs, and 3) impact on CDEs' ability to identify patterns for another 3 participants. Results: Gold standard (GS) included 9 patterns across 3 participants. Of these, all 9 were re-discovered using HC: 4 GS patterns were consistent with patterns identified by HC (over 50% of meals in a cluster followed the pattern); another 5 were included as sub-groups in broader clusers. 50% (9/18) of clusters were rated over 3 on 5-point Likert scale for validity, significance, and being actionable. After reviewing clusters, CDEs identified patterns that were more consistent with data (70% reduction in contradictions between patterns and participants' records). Discussion: Hierarchical clustering of blood glucose and macronutrient consumption appears suitable for discovering behavioral-clinical phenotypes in T2DM. Most clusters corresponded to gold standard and were rated positively by CDEs for face validity. Cluster visualizations helped CDEs identify more robust patterns in nutrition and glycemic impact, creating new possibilities for visual analytic solutions. Conclusion: Machine learning methods can use diabetes self-monitoring data to create personalized behavioral-clinical phenotypes, which may prove useful for delivering personalized medicine.
BibTeX
@misc{levine2018behavioral,
  author = {Matthew E. Levine and David J. Albers and Marissa Burgermaster and Patricia G. Davidson and Arlene Smaldone and Lena Mamykina},
  title = {{Behavioral-clinical phenotyping with type 2 diabetes self-monitoring data}},
  year = {2018},
  eprint = {1802.08761},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/1802.08761},
}
Figure