Subtyping patients with chronic disease using longitudinal BMI patterns

Technical Summary: Subtyping Patients with Chronic Disease Using Longitudinal BMI Patterns

Problem Statement
Obesity is a well-established risk factor for major chronic diseases, including diabetes, cancer, and cardiovascular conditions. While the relationship between obesity and these diseases has been extensively studied using cross-sectional BMI measurements (single time-point data), the role of longitudinal BMI trajectories (patterns of weight change over time) remains underexplored, particularly on a large scale. Existing literature suggests that historical obesity patterns are critical for understanding disease risk, yet traditional predictive methods often fail to capture these temporal dynamics, while advanced deep learning approaches frequently lack interpretability. This study addresses the gap by investigating whether longitudinal obesity patterns alone can predict an individual's future risk of developing 18 major chronic diseases and by subtyping patients based on these patterns.

Methodology
The study utilized a large, geographically diverse Electronic Health Record (EHR) dataset from the athenahealth system, covering approximately 1.53 million individuals (after filtering for multiple visits) over a six-year period (2013–2018).

1. Cohort Definition: The researchers defined 18 separate cohorts, one for each major chronic disease (e.g., diabetes, stroke, Alzheimer's, cancer, hypertension). Each cohort included patients diagnosed with the specific disease and an equal number of randomly selected healthy controls. A combined cohort of patients with any of the 18 diseases was also analyzed.
2. Feature Engineering: To capture temporal obesity patterns while maintaining interpretability, nine new features were engineered from individual BMI trajectories:
   • Weighted Mean BMI: Accounts for irregular time intervals between visits.
   • Trend: The weighted mean of differences between consecutive BMI values (first derivative).
   • Normalized Up-Count and Down-Count: Measures the frequency of weight gain and loss cycles (weight cycling).
   • Maximum BMI: The highest recorded BMI value.
   • Maximum Change in BMI: The largest spike or drop between consecutive visits.
   • Start and End BMI Categories: Categorical variables indicating the BMI classification (e.g., underweight, obese) at the beginning and end of the observation period.
   • Median BMI: The central tendency of the trajectory.
3. Clustering (Subtyping): The k-means clustering algorithm was applied to each cohort using the nine engineered features. The optimal number of clusters was determined using the elbow method, and cluster quality was validated using Silhouette and Calinski-Harabasz scores. The authors compared k-means against hierarchical clustering and Deep Temporal Clustering (DTC), finding k-means produced the most distinct, non-overlapping clusters.
4. Analysis: Identified clusters were analyzed for demographic (age, gender, race), socioeconomic (income, insurance), and physiological (HbA1c, blood pressure, LDL) characteristics. Relative risks of disease incidence were calculated for clusters dominated by positive cases versus negative cases.

Key Contributions

• Interpretable Feature Set: The authors defined a series of nine interpretable, evidence-based features derived from longitudinal BMI values. Unlike regression methods that may miss temporal patterns or deep learning models that act as "black boxes," these features explicitly capture trend, seasonality, and irregularity in BMI data.
• Validation of Temporal Relevance: The study demonstrated that features capturing temporal BMI patterns are directly associated with the incidence of specific diseases, including obesity, diabetes, cancer, hypertension, stroke, dementia, and others.
• Patient Subtyping: The research successfully identified distinct patient subtypes (clusters) for eight of the 18 studied diseases. These subtypes exhibit distinguishable patterns of disease incidence and are characterized by unique demographic, socioeconomic, and physiological profiles.

Results
The clustering analysis revealed noticeable, distinct patterns in 8 out of the 18 chronic disease cohorts. The remaining 10 diseases did not show distinct BMI-related subtypes, potentially due to insufficient positive cases or the absence of a strong BMI-only signal.

• Diabetes: Two clusters (2 and 3) showed high concentrations of diabetic patients (60.83% and 57.28%, respectively). These groups were characterized by higher physiological risks (elevated HbA1c, blood pressure, LDL), lower income, and BMI trends that remained above 30 with less decline.
• Cancer: Cluster 2 showed a higher risk of cancer (59.68% positive) with high BMI trends (>40), while Cluster 3 had mostly healthy individuals but with poorer physiological measurements and a lower BMI trend.
• Hypertension: Cluster 5 showed a 1.66 times higher risk of hypertension compared to Cluster 2. Cluster 5 had higher BMI trends, lower income, and a higher proportion of Black or African American patients.
• Alzheimer's/Dementia: Clusters 2 and 3, consisting of older adults with high BMI trends or steady high BMI, showed significantly higher risks (2.93 and 2.51 times) compared to Cluster 1.
• Obesity: The obesity cohort showed the most distinct patterns. Clusters 2 and 3 had extremely high risks of obesity (29.7 and 32.87 times higher than Cluster 1) and were characterized by younger/middle-aged individuals with high LDL and blood pressure.
• Hip Fracture/Osteoporosis: An inverse relationship was observed; clusters with higher positive cases (fractures) had comparatively healthier physiological measurements, while clusters with worse health conditions had fewer positive cases.
• Combined Cohort: Analysis of patients with any of the 18 diseases revealed that diseased clusters generally had worse physiological measurements and lower incomes compared to the asymptotic (healthy) cohort, with significant racial and geographic disparities.

Significance and Claims
The paper claims that longitudinal BMI patterns alone can effectively subtype individuals at risk for specific chronic diseases, re-establishing the direct relationship between obesity trajectories and diseases such as diabetes, hypertension, and dementia. The study highlights that patient subtyping based on BMI trajectories propounds the role of obesity as a factor in major chronic diseases, revealing distinct subgroups with specific demographic and physiological characteristics.

The authors maintain a modest stance, noting that while BMI patterns are significant, chronic diseases result from complex interactions of many covariates. They acknowledge limitations, including the lack of separate male/female cohorts to preserve sample size and the reliance on interpretable features rather than more complex, less interpretable deep learning models (like RNNs or transformers) which might reveal hidden patterns if larger datasets were available. The study concludes that these findings support existing guidelines (e.g., screening for diabetes in overweight individuals aged 40–59) and suggest that future research should explore BMI trends in conjunction with other covariates.