Utility of glucose, lipid and kidney function Trajectory Measures for incident Cardiovascular Disease risk prediction for people living with Type 2 Diabetes: a case-study using Danish registry data

This Danish registry study of over 83,000 people with type 2 diabetes demonstrates that while incorporating trajectory measures of HbA1c, LDL-cholesterol, and eGFR (specifically dispersion and change rather than central tendency) only marginally improves the discrimination of incident cardiovascular disease risk, it offers modest gains in net reclassification, particularly for LDL-cholesterol variability.

The Gist

Imagine you are trying to predict who might get into a car accident in the next few years.

Most doctors currently look at a driver's single snapshot from today: "Is their speedometer reading 60 mph right now? Are they wearing a seatbelt?" Based on that one moment, they guess the risk.

This new study asks a different question: What if we looked at the driver's entire history? Did they speed up and down wildly? Did they gradually slow down over time? Did they have a pattern of erratic driving, even if their speed today looks normal?

Here is the breakdown of this research, translated into everyday language.

The Big Picture: The "Snapshot" vs. The "Movie"

People with Type 2 diabetes are at higher risk for heart problems (like heart attacks or strokes). Doctors usually use a "snapshot" to predict this risk. They look at your blood sugar, cholesterol, and kidney function just once (the most recent test) and plug those numbers into a formula.

The problem? A single test can be misleading. Maybe your blood sugar was high because you ate a donut yesterday, or maybe your cholesterol is low because you just started a new diet. One number doesn't tell the whole story.

This study, using massive health records from Denmark, tried to see if looking at the "movie" (the history of your numbers over three years) helps predict heart trouble better than just the "snapshot."

The Three Key Characters

The researchers tracked three specific "vital signs" for 83,000 people with diabetes:

1. HbA1c (The Sugar Gauge): How much sugar is in your blood on average.
2. LDL Cholesterol (The Gunk): The "bad" cholesterol that clogs arteries.
3. eGFR (The Kidney Filter): How well your kidneys are cleaning your blood.

The Experiment: Finding the "Rhythm"

Instead of just looking at the final number, the researchers calculated two new things for each person based on their past tests:

• The "Jitter" (Variability): How much did the numbers bounce around? (Did the sugar go from 100 to 200 and back to 100?)
• The "Trend" (Slope): Was the person getting better or worse over time? (Was the cholesterol slowly creeping up, even if it wasn't high today?)

What They Discovered

Here is the surprising part: The "Jitter" and the "Trend" mattered more than the average.

• The "Average" didn't help much: Knowing that your average sugar was 50 didn't predict heart trouble very well.
• The "Jitter" was a red flag: People whose numbers bounced around a lot (high variability) were at higher risk. It's like a car engine that revs up and down erratically; it's more likely to break than one that runs steadily, even if the average speed is the same.
• The "Trend" was crucial: People whose numbers were slowly getting worse over time were at higher risk.
• The "Cholesterol" was the star: The "jitter" and "trend" in LDL cholesterol were the strongest predictors. If your "bad" cholesterol was bouncing around or creeping up, your risk of a heart event went up significantly.

Did It Change the Prediction?

The researchers asked: "Does adding this 'movie' history actually make our predictions better?"

• The "C-Index" (The Score): Imagine a test where 0.5 is a coin flip and 1.0 is perfect. The old "snapshot" model scored about 0.67. Adding the "movie" history bumped it up to 0.673.
  • Translation: It's a tiny, tiny improvement. It's like adding a pinch of salt to a soup that's already pretty good. It doesn't make it a gourmet meal, but it adds a little flavor.
• The "Reclassification" (The Safety Net): This is where it got interesting. While the overall score didn't change much, the new model was better at catching the people who were about to get sick.
  • Analogy: Think of a fishing net. The old net caught most fish, but let some big ones slip through. The new "movie" net didn't catch more fish overall, but it caught the specific big fish that were trying to escape. It correctly identified about 3% to 9% more people who were at high risk, allowing doctors to treat them sooner.

The Bottom Line

The "Movie" is better than the "Snapshot," but it's not a magic wand.

• Good News: We don't need expensive new tests. We already have all this data in our medical records. We just need to start looking at the patterns and trends instead of just the latest number.
• Reality Check: The improvement is modest. It won't perfectly predict who will get a heart attack, but it helps doctors spot the "sneaky" risks that a single test misses.
• The Takeaway: If your numbers are bouncing around wildly or slowly creeping up, even if your current test looks "okay," you might be at higher risk. Doctors should pay attention to the journey, not just the destination.

In short: Don't just look at where you are today; look at how you got there. The path matters just as much as the destination.

Technical Summary

1. Problem Statement

Cardiovascular disease (CVD) is the leading cause of mortality in people with Type 2 Diabetes (T2D). Current clinical risk prediction models rely on single baseline measurements of risk factors (e.g., the most recent HbA1c, LDL-cholesterol, or eGFR). These models demonstrate only moderate predictive performance (C-indices $\approx$ 0.7) and often underestimate risk in specific subgroups (e.g., women and younger patients).

The authors hypothesize that longitudinal trajectory measures—derived from repeated clinical measurements over time—contain incremental predictive value regarding CVD risk that single time-point measurements miss. Specifically, they investigate whether metrics of dispersion (variability) and change (slope) in glucose, lipids, and kidney function improve risk stratification beyond standard baseline values.

2. Methodology

Study Design and Data Sources

• Design: A register-based multivariable prognostic prediction modeling study with internal validation.
• Population: 83,326 individuals with prevalent T2D from nationwide Danish registries.
  • Inclusion: Alive on Jan 1, 2015; diagnosed with T2D; no prior CVD history; $\ge$2 recorded measurements of HbA1c, LDL-C, and eGFR between 2012–2014.
  • Exclusion: Individuals without sufficient longitudinal data (often due to regional registry gaps in Southern Denmark).
• Data Linkage: Unique personal identifiers linked data from the Civil Registration System, National Patient Register, Register of Laboratory Results for Research (RLRR), Prescription Registry, and Death Registers.

Variables and Trajectory Construction

• Baseline (Index Date): Jan 1, 2015.
• Observation Window: 2012–2014 (3 years prior to baseline).
• Predictors:
  • Standard Baseline: Age, sex, T2D duration, HbA1c, LDL-C, HDL-C, eGFR, and medication use (BP, lipid, glucose-lowering, insulin).
  • Trajectory Measures: Calculated for each participant using three paired approaches for HbA1c, LDL-C, and eGFR:
    1. Mean & Standard Deviation (SD)
    2. Median & Interquartile Range (IQR)
    3. Intercept & Slope (from a linear mixed-effects growth model).
• Outcome: Major Adverse Cardiovascular Events (MACE), defined as non-fatal ischemic heart disease, stroke, heart failure, or CVD mortality, occurring between 2015–2020.

Statistical Analysis

• Reference Model: Cox proportional hazards regression using only baseline predictors.
• Trajectory Models: Sequentially added paired trajectory measures to the reference model using an offset procedure.
• Performance Metrics:
  • Discrimination: Change in C-index ($\Delta$C-index).
  • Reclassification: Continuous Net Reclassification Index (NRI).
  • Association: Hazard Ratios (HR) with 95% Confidence Intervals (CI).
• Software: R (v4.4.1) using packages survival, nricens, etc.

3. Key Results

Cohort Characteristics

• Demographics: Mean age 65 years; 48% female; median T2D duration 6 years.
• Baseline Levels: Median HbA1c 48 mmol/mol; LDL-C 2.1 mmol/L; eGFR 83 mL/min/1.73m².
• Events: During a median 6-year follow-up, 11,280 participants (13.5%) experienced a MACE event.

Association with CVD Risk

• Dispersion and Change vs. Centrality:
  • Measures of centrality (mean, median, intercept) showed no significant association with CVD events after adjusting for baseline values.
  • Measures of dispersion and change (SD, IQR, slope) were significantly associated with increased risk.
• Hazard Ratios (Adjusted):
  • HbA1c: Slope HR 1.16; SD HR 1.08.
  • LDL-Cholesterol: Slope HR 1.48 (strongest association); SD HR 1.15.
  • eGFR: SD HR 1.20; Slope HR 1.10.
  • Note: Higher variability or steeper negative/positive slopes (depending on the metric direction) correlated with higher risk.

Predictive Performance

• Discrimination (C-index): The reference model had a C-index of 0.67. Adding trajectory measures resulted in minimal gains ($\Delta$C-index: +0.001 to +0.003).
• Reclassification (NRI): Modest but consistent improvements were observed.
  • Continuous NRI ranged from +3.3% (LDL Median & IQR) to +7.9% (HbA1c Intercept & Slope).
  • A combined model (incorporating the best trajectory pairs for all three domains) achieved an NRI of +9.0% and a C-index of 0.673.
• Subgroups: Improvements in reclassification appeared more pronounced in patients >40 years, males, and those on specific medications, though results were exploratory.

4. Key Contributions

1. Specific Focus on T2D: Unlike previous studies in general populations, this is the first large-scale study to specifically evaluate trajectory measures of HbA1c, LDL-C, and eGFR in a T2D cohort.
2. Dispersion over Centrality: The study provides robust evidence that variability and rate of change in these biomarkers are prognostically informative, whereas the average level (centrality) adds little value once the baseline level is accounted for.
3. LDL-Cholesterol Volatility: Identified that variability in LDL-cholesterol is a particularly strong predictor of CVD in T2D (HR > 1.4 for slope), suggesting that unstable lipid control is a critical risk factor.
4. Pragmatic Implementation: Demonstrated that these trajectory measures can be derived from routinely collected Electronic Health Record (EHR) data without requiring new biomarkers or complex testing, offering a low-cost method to refine risk models.

5. Significance and Limitations

• Clinical Implication: While the improvement in discrimination (C-index) is statistically small, the improvement in risk reclassification (NRI) is clinically relevant. It suggests that trajectory measures can help identify high-risk patients who might be missed by static baseline models, potentially guiding more aggressive management.
• Feasibility: The approach leverages existing data infrastructure, making it immediately applicable in healthcare systems with longitudinal registries.
• Limitations:
  • Selection Bias: The requirement for $\ge$2 measurements may exclude sicker or less adherent patients, though the authors note a large portion of exclusions were due to geographic registry gaps rather than clinical status.
  • Modest Gains: The incremental predictive value, while statistically significant, is modest. The authors suggest that combining trajectory measures with novel biomarkers or advanced machine learning techniques may be necessary for substantial clinical impact.
  • Generalizability: Results are based on the Danish healthcare system; applicability to systems with different testing frequencies or healthcare access patterns requires validation.

Conclusion: Trajectory measures of dispersion and change in HbA1c, LDL-cholesterol, and eGFR provide independent prognostic information for CVD risk in T2D. While they do not drastically overhaul discrimination metrics, they offer a pragmatic, low-cost enhancement to existing risk prediction tools, particularly for reclassifying patients into more accurate risk categories.