Design-induced artifacts when 'disease clocks' are plugged into second-stage analyses of symptom onset

This study demonstrates that the reported predictive performance of plasma p-tau217 disease clocks for Alzheimer's disease symptom onset is largely an artifact of structural dependencies between baseline age and follow-up constraints, rather than a reflection of independent biomarker signal.

The Gist

The Big Idea: The "Fake" Clock

Imagine a group of scientists who built a fancy new Alzheimer's "Disease Clock."

This clock uses a blood test (measuring a protein called p-tau217) to guess two things:

1. When a person's brain first started showing signs of the disease (the "start time").
2. When that person will likely start showing symptoms (the "symptom time").

The scientists claimed their clock was amazing. They said, "If we look at your blood test, we can predict exactly how many years it will be before you get sick."

This paper argues that the clock isn't actually telling us anything new about the disease. Instead, it's just a mathematical trick that confuses "how old you are" with "when you get sick."

---

The Analogy: The "Running Race" Trap

To understand the problem, let's imagine a race.

The Setup:

• The Runners: People in a study.
• The Finish Line: The moment they develop Alzheimer's symptoms.
• The Stopwatch: The "Disease Clock" that tries to predict when they will cross the finish line.

The Flaw in the Study:
The scientists only looked at runners who actually finished the race during the study period. They ignored everyone who was still running or hadn't started yet.

Here is the trick:
If you start a race at age 80, and the race only lasts for 10 years, you cannot finish the race at age 95. You simply don't have enough time left in the study.

• If you are 80, you must finish by 90.
• If you are 85, you must finish by 90.

Because of this time limit, your starting age (baseline age) automatically predicts when you finish. It's not because you are a fast or slow runner; it's just because of the math of the time limit.

The "Clock" Mistake:
The scientists took their fancy "Disease Clock" (which is just your starting age minus a calculated time) and used it to predict the finish line.

• The Reality: The clock is mostly just repeating your starting age in a fancy way.
• The Result: The clock looks like it's working perfectly, but it's only working because it's repeating the obvious fact that "older people who start the study later will finish the study sooner."

The "Magic Trick" Proof

To prove this wasn't a real biological discovery, the authors did a magic trick:

1. The Real Clock: They used the actual blood test data to calculate the "time to sickness."
2. The Fake Clock: They threw away the blood test data and replaced it with random numbers (like rolling dice) to guess the time to sickness.

The Shocking Result:
The "Fake Clock" (random numbers) predicted the outcome just as well as the "Real Clock."

This proves that the blood test data wasn't doing the heavy lifting. The prediction was coming entirely from the structure of the study (the age limits and the time window), not from the biology of the disease.

The "Shared Room" Analogy

Think of the prediction model as a room with two people trying to explain why a door opened:

1. Person A (Baseline Age): "I pushed the door open because I was already standing right next to it."
2. Person B (The Blood Test Clock): "I pushed the door open because of my special disease timing."

The authors used a statistical tool (Commonality Analysis) to see who actually pushed the door.

• Person A pushed the door 90% of the time.
• Person B barely touched the door.
• The Problem: Because Person B was standing right next to Person A, it looked like they were both pushing together. But when you separate them, Person B is barely doing anything.

Why Does This Matter?

1. False Hope: If doctors use these "Disease Clocks" to tell patients, "You have 5 years until symptoms," they might be wrong. The clock is mostly just saying, "You are old, and the study ended soon."
2. Wasted Money: Researchers might spend millions trying to improve a clock that is mathematically broken, rather than looking for real biological signals.
3. The Real Signal: The authors aren't saying the blood test (p-tau217) is useless. It is a good marker for the disease. But using it to build a "timeline clock" in this specific way creates a hallucination of precision.

The Bottom Line

The paper is a warning label. It says: "Be careful when you mix 'Age' with 'Time to Event' in a study with a short deadline. You might think you've discovered a new law of physics, but you've just discovered a math trick."

The "Disease Clock" isn't telling us when the disease will strike; it's just telling us how much time is left on the calendar.

Technical Summary

1. Problem Statement

The paper critiques a recent study by Petersen et al. that utilized "disease clock" models (Sampled Iterative Local Approximation [SILA] and Temporal Integration of Rate Accumulation [TIRA]) to estimate the age at plasma %p-tau217 positivity. Petersen et al. claimed that this estimated age predicts the age at onset of symptomatic Alzheimer's Disease (AD).

The authors of this critique (Insel and Donohue) argue that the reported high predictive performance is not driven by the biological signal of the biomarker (plasma p-tau217) but is instead an artifactual result of the study design and statistical structure. Specifically, they identify two primary sources of bias:

1. Restricted Sampling: The analysis was limited to individuals who progressed to symptomatic AD within a fixed follow-up window.
2. Mathematical Coupling: The predictor (estimated age of positivity) and the outcome (age of symptom onset) share common time components (baseline age), creating a self-referential relationship that induces spurious correlations.

2. Methodology

The authors re-analyzed digitized data extracted from the published figures of the Petersen et al. study using the following approaches:

• Decomposition of Predictors: They broke down the "clock-derived" predictor (Estimated Age of %p-tau217 Positivity) into its constituent parts:
  • Baseline Age ($A_{base}$)
  • Estimated Time from Positivity ($T_{est}$)
  • Relationship: $Age_{pos} = A_{base} - T_{est}$
• Commonality Analysis: They performed a variance decomposition to quantify how much variance in the outcome (Age of Symptom Onset) was explained uniquely by Baseline Age, uniquely by the estimated time component, and how much was shared between them.
• Null Scenario (Randomization): To isolate structural artifacts from biomarker signal, they replaced the biomarker-derived time component ($T_{est}$) with randomly generated values drawn uniformly from the observed range. This preserved the statistical distribution of the predictor but removed all biological information.
• Re-evaluation of Alternative Outcomes: They examined analyses where the outcome was "duration from positivity to onset" (excluding age from the outcome) to demonstrate that the structural artifact persists, merely shifting the constraint from baseline age to the age of positivity.

3. Key Contributions

• Identification of Structural Artifacts: The paper provides a rigorous statistical demonstration that "disease clock" models, when applied to restricted cohorts with bounded follow-up, generate strong associations purely due to mathematical constraints rather than biological validity.
• Quantification of Biomarker Utility: By isolating the unique contribution of the biomarker timing component, the authors show it adds negligible predictive value compared to baseline age alone.
• Generalizability of the Critique: The authors argue this is not unique to p-tau217 but applies to any disease clock model using constructed predictors and outcomes with shared temporal components (e.g., amyloid PET clocks).

4. Key Results

• Dominance of Baseline Age: In the ADNI cohort, Baseline Age alone explained approximately 78% of the variance in age at symptom onset ($R^2 \approx 0.78$).
• Superiority of Simple Age over Clocks: The reported clock-derived predictors (TIRA and SILA) had significantly lower explanatory power ($R^2 \approx 0.337$ and $0.470$, respectively) compared to baseline age alone.
• Negligible Unique Contribution: Commonality analysis revealed that the estimated time from %p-tau217 positivity contributed minimally to the prediction:
  • Unique Variance (SILA): 6%
  • Unique Variance (TIRA): 3%
  • Unique Variance (Randomized): 0.07%
  • The vast majority of explained variance was either unique to baseline age or shared between baseline age and the clock predictor.
• The Randomization Test: When the biomarker component was replaced with random noise, the predictive performance remained nearly identical to the original clock models ($R^2 \approx 0.79$). This confirms that the association is driven by the structural relationship between the variables, not the biomarker data.
• Persistence in Duration Analyses: Analyses regressing "time from positivity to onset" against "age of positivity" also showed strong negative correlations. The authors demonstrate this is a geometric constraint: individuals who become positive at older ages physically cannot have long durations to onset within the study's observation window, creating a false negative correlation.

5. Significance and Implications

• Methodological Warning: The study serves as a critical warning against the uncritical use of "disease clock" models in longitudinal analyses where the outcome is constrained by the predictor's time components. It highlights the risk of overfitting to structural constraints rather than true biological signals.
• Clinical Interpretation: The findings suggest that current claims regarding the ability of plasma p-tau217 clocks to precisely predict the timing of symptom onset in individuals are likely overstated. The predictive power is largely a reflection of the patient's current age and the study's follow-up limits.
• Future Research Directions: The authors advocate for:
  • Evaluating predictors in "all-comers" populations rather than restricted progressor-only samples.
  • Using latent class approaches to distinguish between progressors and stable individuals, rather than assuming a universal declining trajectory.
  • Careful statistical evaluation of constructed predictors to ensure they do not induce spurious associations through shared variance.

In conclusion, the paper argues that the apparent success of plasma p-tau217 disease clocks in predicting symptom onset is a statistical illusion created by the study design, and that baseline age remains the dominant predictor in these specific analytical frameworks.