Estimating tau onset age from tau PET imaging in two longitudinal cohorts using sampled iterative local approximation

This study demonstrates that the Sampled Iterative Local Approximation (SILA) algorithm can accurately model longitudinal tau PET trajectories and estimate individual tau positivity onset ages in the meta-temporal region across two cohorts, though its accuracy in the entorhinal cortex is limited in individuals with dementia.

The Gist

The Big Picture: Reconstructing a Movie from a Few Frames

Imagine Alzheimer's disease is a long, slow-motion movie that plays out over 20 or 30 years. The "plot" involves two main villains: Amyloid (the first bad guy to show up) and Tau (the second bad guy who arrives later and causes the most damage).

The problem for scientists is that they can't watch the whole movie. Most people only show up to the cinema (get scanned) when they are already middle-aged or older, and they only get to see a few scenes (scans) over a few years. They miss the beginning of the movie entirely.

The Goal: The researchers wanted to build a "time machine" (a mathematical algorithm called SILA) that could look at the few scenes a person did see and accurately guess:

1. When the movie actually started for them.
2. What the scenes looked like before they walked into the theater.
3. How fast the plot is moving.

They already built this time machine for the first villain (Amyloid), and it worked great. Now, they wanted to see if it could work for the second villain (Tau).

---

The Experiment: Two Different Cameras, Two Different Groups

To test their time machine, the researchers used data from two different groups of people (cohorts) and two different types of "cameras" (PET tracers):

1. The ADNI Group: A large, national study with older participants. They used a camera called Flortaucipir.
2. The WISC Group: A Wisconsin-based study with slightly younger participants. They used a camera called MK-6240.

They looked at two specific "locations" in the brain where the Tau villain usually hides:

• The Entorhinal Cortex (EC): The "front door" of the brain. This is usually where Tau shows up first.
• The Meta-Temporal (MT) Region: A larger neighborhood further back in the brain where Tau spreads later.

---

The Results: How the Time Machine Worked

1. The "Backward" vs. "Forward" Trick

The researchers tested if their time machine could look backward (guessing what happened in the past) or forward (guessing what will happen in the future).

• The Analogy: Imagine you see a car driving down a hill.
  • Looking Backward: You can easily guess where it came from because the path is clear.
  • Looking Forward: It's harder to guess exactly where it will go next because the road might curve, or the driver might hit the brakes.
• The Finding: The time machine was excellent at looking backward. It could accurately guess a person's past brain scan results. However, it was less accurate at predicting the future, especially for people who hadn't developed the disease yet.

2. The "Front Door" vs. The "Neighborhood" Problem

This is the most interesting part of the study. The time machine worked differently depending on where in the brain they looked.

• The Meta-Temporal (MT) Region (The Neighborhood):

  • Performance: The time machine worked perfectly here. It could accurately estimate when Tau arrived and how fast it was spreading, regardless of whether the person had dementia or not.
  • Why? Once Tau spreads to this larger area, it follows a very predictable, steady path, like a river flowing downstream.

• The Entorhinal Cortex (The Front Door):

  • Performance: The time machine got confused here, but only for people who already had dementia.
  • The Analogy: Imagine the "Front Door" is a small, crowded room. When a person gets very sick (dementia), the room starts to shrink (brain atrophy) and the furniture gets rearranged. The camera sees the room getting smaller, so the "Tau signal" looks like it's going down instead of up. The time machine, expecting the signal to always go up, gets thrown off.
  • The Fix: The researchers found that if they only looked at people before they got dementia (the preclinical stage), the time machine worked great at the front door too. But once dementia sets in, the "Front Door" is too messy to use for precise timing.

3. The "Genetic Risk" Factor

The researchers also checked if the time machine respected known risk factors, specifically the APOE-e4 gene (a genetic variant that increases Alzheimer's risk).

• The Finding: The time machine confirmed what we already suspected: People with the risky gene got "invaded" by Tau about 4 to 5 years earlier than people without the gene.
• Why this matters: It proves the time machine is telling the truth. If the machine was broken, it wouldn't have noticed this difference.

---

The Takeaway: What Does This Mean for Us?

The Good News:
Scientists now have a reliable tool to "rewind the tape" for people with Alzheimer's. Even if a patient gets their first brain scan at age 70, this tool can estimate that their Tau buildup actually started around age 60. This helps researchers understand the timeline of the disease much better.

The Caveat:
The tool works best in the "neighborhood" (Meta-Temporal region) and for people who haven't reached the dementia stage yet. If you are looking at the "Front Door" (Entorhinal Cortex) of a patient with advanced dementia, the tool gets a bit shaky because the brain changes physically (shrinks) in ways that confuse the camera.

In Summary:
Think of SILA as a smart detective. It can look at a few clues (scans) and tell you exactly when the crime (Tau buildup) started and how fast it's spreading. It's a very good detective for the main crime scene, but it needs a little help when the crime scene is a bit messy (advanced dementia). This helps doctors and scientists plan better treatments and understand who is at risk and when.

Technical Summary

1. Problem Statement

Alzheimer's Disease (AD) biomarkers, specifically amyloid and tau, follow a stereotypic temporal sequence beginning decades before clinical symptoms. While methods exist to model longitudinal amyloid PET trajectories and estimate the age of amyloid positivity (A+ onset), there is a lack of validated methods for modeling longitudinal tau PET trajectories and estimating individual tau positivity (T+) onset ages.

• The Challenge: Longitudinal tau PET data is often sparse (5–10 years of follow-up), making it difficult to observe the exact transition from tau-negative (T-) to tau-positive (T+).
• The Need: A robust algorithm is required to retrospectively estimate T+ onset ages and model tau accumulation trajectories across different brain regions, tracers, and cohorts to facilitate clinical trial planning and understanding disease progression.

2. Methodology

The study validated the Sampled Iterative Local Approximation (SILA) algorithm, previously used for amyloid, on tau PET data.

• Cohorts & Data:
  • ADNI: 385 participants with longitudinal [18F]flortaucipir (FTP) tau PET scans.
  • WISC (Wisconsin): 288 participants with longitudinal [18F]MK-6240 tau PET scans.
  • Regions of Interest (ROIs): Entorhinal Cortex (EC) and a Meta-Temporal (MT) composite region.
• SILA Implementation:
  • Training: SILA uses a discrete sampling approach to generate a numeric curve representing the mean rate of change at evenly spaced SUVR values. This curve is smoothed and integrated (Euler's method) to create a SUVR vs. time trajectory.
  • Anchoring: The time axis is anchored such that 0 years corresponds to the T+ threshold (determined via Gaussian Mixture Models for each cohort/ROI).
  • Estimation: For individual participants, the model solves for "T+ time" based on an observed SUVR. This is subtracted from the participant's age at the scan to estimate the Estimated Tau Onset Age (ETOA).
• Validation Strategy (Aim 1):
  • 10-Fold Cross-Validation (KF): Trained on 90% of data, tested on 10%.
  • Whole-Sample (WH): Trained on the full dataset.
  • Prediction Schemes: Evaluated Forward (predicting future scans from the first scan) and Backward (predicting past scans from the last scan) accuracy.
  • Metrics: Root Mean Square Error (RMSE), Residual Sum of Squares (RSS), bias, and T+/– classification accuracy.
  • ETOA Accuracy: Compared SILA-estimated onset ages against the observed midpoint between the last T- and first T+ scans in converters.
• Risk Factor Analysis (Aim 2):
  • Examined associations between SILA-derived T+ status/onset and APOE-ε4 carrier status and dementia (CDR ≥ 1).

3. Key Contributions

• First Validation of SILA for Tau: Demonstrates that SILA, a method validated for amyloid, can accurately model longitudinal tau PET trajectories and estimate individual T+ onset ages.
• Multi-Cohort/Tracer Generalizability: Successfully applied across two distinct cohorts (ADNI and WISC) using two different tracers (FTP and MK-6240), proving the method's robustness.
• Regional Specificity: Identified critical differences in model performance between the Entorhinal Cortex (EC) and the Meta-Temporal (MT) region, particularly regarding the impact of dementia.
• Backward vs. Forward Prediction: Confirmed that retrospective (backward) estimation is significantly more accurate than prospective (forward) prediction for tau, especially in T- individuals.

4. Key Results

Model Performance (Aim 1)

• Trajectory Consistency: SILA produced monotonically increasing curves with low slopes below the T+ threshold and steeper slopes above.
• Prediction Accuracy:
  • Backward > Forward: Backward prediction consistently outperformed forward prediction across all cohorts and ROIs (Lower RMSE, higher T+/– accuracy: 90–97% vs. 88–92%).
  • Region Differences: The Meta-Temporal (MT) region showed superior performance (lower error, higher accuracy) compared to the Entorhinal Cortex (EC).
  • Dementia Impact: In the EC, model residuals were significantly associated with dementia status (ADNI). Dementia participants often showed negative within-person SUVR slopes at high values, degrading model fit. The MT region was less affected by dementia status.
• ETOA Accuracy: In T- to T+ converters, the difference between observed and estimated onset ages was minimal:
  • ADNI MT: 0.12 years difference.
  • WISC MT: -0.09 years difference.
  • Estimates fell within the observed T- to T+ timeframe in 72–86% of cases.

Risk Factor Associations (Aim 2)

• APOE-ε4: APOE-ε4 carriers had significantly earlier estimated T+ onset ages (approx. 4.6 years earlier in EC) and higher odds of being T+ compared to non-carriers. SILA-derived odds ratios closely matched observed threshold-based odds ratios.
• Dementia: Individuals with dementia had significantly longer estimated time spent as T+ (approx. 3–4 years longer) compared to non-demented individuals.
• Consistency: SILA-derived associations with risk factors did not differ significantly from those derived using raw observed PET data, validating the method's utility for epidemiological studies.

5. Significance and Implications

• Retrospective Estimation: SILA allows researchers to "fill in the gaps" in longitudinal studies, estimating tau burden and onset age for timepoints prior to the first available PET scan. This is crucial for studies where the T- to T+ transition was not directly observed.
• Clinical Trial Design: By providing a "tau clock" (time since T+ onset), SILA can help stratify participants in clinical trials based on disease stage rather than just cross-sectional biomarker levels.
• Regional Nuance: The study highlights that while the MT region is robust for modeling across the entire disease spectrum (preclinical to dementia), the EC is more sensitive to the effects of dementia (likely due to atrophy and partial volume effects). Therefore, EC modeling is best suited for preclinical research, while MT is preferred for broader clinical applications.
• Methodological Best Practices: The authors recommend using backward prediction for retrospective estimation and suggest that while whole-sample training is efficient, cross-validation is necessary to ensure convergence, particularly in populations with dementia.

Conclusion: The study establishes SILA as a valid, accurate tool for modeling longitudinal tau PET trajectories and estimating individual T+ onset ages, offering a powerful method to characterize the temporal dynamics of Alzheimer's pathology across diverse populations.