Latent Class Trajectory Phenotypes of Longitudinal Visit and Follow-up Patterns Among Patients with Hypertension

Technical Summary: Latent Class Trajectory Phenotypes of Longitudinal Visit and Follow-up Patterns Among Patients with Hypertension

Problem Statement
Hypertension management relies heavily on sustained retention in care, yet long-term follow-up patterns are heterogeneous, with many patients discontinuing visits early. Conventional summary measures (e.g., total visit counts) assume population homogeneity and fail to capture distinct longitudinal subgroups. Furthermore, Electronic Health Record (EHR) data captures "observed contact" (retention) but cannot directly measure patient engagement or medication adherence. A critical gap exists in characterizing the specific trajectories of visit retention in hypertension and understanding how administrative censoring (variable observation windows) influences these patterns and their association with clinical outcomes like blood pressure (BP) control.

Methodology
The authors conducted a retrospective cohort study using de-identified EHR data from the EvidenceNOW quality-improvement program.

• Cohort: 26,710 patients with hypertension and at least two recorded visits.
• Modeling Approach: The study employed Latent Class Trajectory Modeling (LCTM) of repeated binary visit indicators within a Bernoulli finite mixture framework. Unlike classical Group-Based Trajectory Modeling (GBTM) which fits polynomial curves, this approach estimates class-specific attendance probabilities independently at each time interval, allowing for flexible, non-parametric shapes.
• Temporal Resolutions: Models were fitted using 1-month intervals (primary) and 3-month intervals (robustness check) over a 36-month period.
• Model Selection: The number of latent classes ($K$) was determined using the Bayesian Information Criterion (BIC) and GRoLTS-recommended metrics: Average Posterior Probability (APP), Relative Entropy, and Odds of Correct Classification (OCC).
• Validation: Stability was assessed via 200 bootstrap resamples using the Adjusted Rand Index (ARI). A sensitivity analysis restricted the cohort to patients with $\ge$24 months of follow-up to assess the impact of administrative censoring.
• Statistical Analysis: Multivariable linear regression examined associations between trajectory phenotypes and treatment documentation/BP control, adjusting for age, sex, race, baseline BP, and comorbidity burden.

Key Contributions and Results

1. Identification of Four Phenotypes: The study identified four internally reproducible latent classes of observed EHR contact, consistent across both 1-month and 3-month resolutions:
   • Low Observed Follow-up (67–68%): Rapid decline in contact after the index visit, approaching zero by month 27.
   • Stable Low Observed Retention (12–18%): Sharp initial decline followed by a stable, low probability of attendance (~0.2).
   • Delayed Decline in Observed Visits (11–15%): Moderate contact maintained for ~12–18 months before a gradual decline.
   • Sustained Observed Retention (2–6%): Consistently high contact probability throughout the 36-month period.
2. Reproducibility and Stability: The four-class structure demonstrated excellent bootstrap reproducibility (mean ARI 0.966–0.969). Concordance between the 1-month and 3-month models was high for the extreme groups (Low Follow-up and Sustained Retention) but lower for intermediate groups, suggesting sensitivity to temporal aggregation.
3. Impact of Administrative Censoring: A major finding is that 79.2% of the cohort had less than 24 months of follow-up, with censoring heavily concentrated in the low-contact groups. The study concludes that the observed decline in contact for these groups likely reflects administrative observation limits rather than confirmed patient disengagement. The four-class structure persisted in the sensitivity analysis (patients with $\ge$24 months), though with reduced separation metrics.
4. Clinical Associations:
   • Treatment Documentation: Higher-contact phenotypes showed consistently and significantly higher rates of treatment documentation compared to the Low Observed Follow-up group, even after adjusting for comorbidities and demographics.
   • Blood Pressure Control: Associations between trajectory groups and BP control were small and inconsistent after adjustment. The "Sustained Observed Retention" group did not achieve the lowest BP at fixed time points (12, 24, 36 months) despite showing the largest first-to-last BP reduction. The authors attribute this to bias from differential observation window lengths; fixed-time analyses revealed negligible between-group differences in absolute BP levels.

Significance and Claims
The paper claims that latent class trajectory modeling of binary visit indicators provides a robust framework for characterizing the heterogeneity of observed EHR contact patterns in hypertension. The study emphasizes that these phenotypes are internally reproducible but substantially influenced by administrative censoring, cautioning against interpreting them as confirmed behavioral disengagement without further validation.

The authors assert that observed contact trajectories are more strongly associated with care process measures (specifically treatment documentation) than with absolute BP outcomes. They argue that the apparent improvement in BP for high-contact groups in paired analyses is an artifact of longer observation windows rather than superior treatment response. Consequently, the study concludes that while these phenotypes help identify subgroups with different retention patterns, their utility for risk stratification or targeting interventions requires future studies to distinguish administrative censoring from true loss to follow-up and to link these trajectories to hard cardiovascular endpoints. The paper does not claim these phenotypes are currently ready for clinical implementation but rather establishes a methodological foundation for future inquiry.