Automated, Component-Level Assessment of a National Inpatient Diabetes Guideline: Adherence and Its Association with Hospitalization Complications Across Six Hospitals

Technical Summary: Automated, Component-Level Assessment of a National Inpatient Diabetes Guideline

Problem Statement
While clinical guidelines for inpatient hyperglycemia exist, there is a lack of granular, automated methods to assess adherence at the level of individual recommendations across large hospital networks. Traditional assessments often rely on aggregate metrics or manual chart reviews, which fail to capture the temporal dynamics of care or identify specific gaps in implementation that correlate with patient outcomes. The challenge lies in translating complex, time-dependent clinical guidelines into computable logic to measure adherence continuously and link specific adherence patterns to subsequent hospitalization complications.

Methodology
The study employed a retrospective, multicenter design involving 48,354 admissions of 29,577 adults with diabetes or inpatient hyperglycemia across six Israeli general hospitals (August 2022–December 2025). The core technical innovation was the formalization of 14 specific guideline recommendations into quality-assessment temporal patterns. These patterns were applied to knowledge-based temporal abstractions derived from electronic health records (EHRs), allowing the system to score adherence wherever a recommendation was applicable to a patient's clinical course.

The analysis proceeded in two phases:

1. Adherence Measurement: Pre- and post-guideline issuance adherence was compared using Welch t-tests with false discovery rate correction.
2. Outcome Association: The study investigated the relationship between adherence and 10 specific complications. Associations were estimated using bias-reduced logistic regression over 24-, 48-, and 72-hour windows. To address confounding, inverse probability of treatment weighting (IPTW) was utilized to contrast patients with above-median adherence against those with at-or-below-median adherence.

Key Contributions

• Granular Automation: The paper demonstrates the feasibility of automating guideline adherence assessment at the component level (per recommendation) across a national scale, moving beyond binary "adherent/non-adherent" classifications.
• Temporal Logic: By utilizing temporal abstractions, the methodology captures the continuity of care (e.g., repeated insulin decisions) rather than just discrete admission actions.
• Causal Inference Framework: The application of IPTW and bias-reduced regression provides a robust framework for estimating the association between specific adherence patterns and clinical outcomes in an observational setting.

Results

• Adherence Patterns: Adherence varied significantly by recommendation type. While admission and medication-management actions showed near-universal adherence, substantial gaps were identified in ongoing insulin management, with mean adherence scores ranging from 0.15 to 0.28.
• Temporal Trends: Overall adherence increased modestly following the guideline's issuance (from 0.682 to 0.703, P < .001), though significant heterogeneity remained between hospitals.
• Outcome Associations: After weighting, above-median adherence was associated with:
  • A 44% reduction in 30-day mortality (OR 0.557; 95% CI, 0.514–0.603).
  • An 82% reduction in hyperglycemia (OR 0.175).
  • A 52% reduction in severe hyperglycemia (OR 0.485).
  • A 20% increase in the odds of hypoglycemia (OR 1.204).
• Correlation of Gaps and Outcomes: The recommendations with the lowest implementation rates were predominantly those that tracked most strongly with patient outcomes. Specifically, these were recommendations concerning repeated insulin decisions distributed throughout a hospital stay, as opposed to discrete actions taken only upon admission.

Significance and Claims
The paper claims that automated, component-level assessment is a viable tool for identifying specific areas where fuller guideline implementation would most plausibly impact patient care. The findings suggest that the "low-hanging fruit" of guideline adherence (discrete admission actions) has largely been addressed, while the more complex, continuous aspects of care (ongoing insulin management) remain the primary barriers to optimal outcomes. The study concludes that such automated assessments can effectively pinpoint where targeted interventions are needed to reduce complications and mortality in inpatient diabetes care.