Streamlining Eligibility Assessment for Alzheimers Disease-Modifying Therapies: Prediction of MMSE Scores Using the Digital Clock and Recall

This study demonstrates that a machine learning model utilizing multimodal data from the rapid, digital Clock and Recall (DCR) assessment can accurately and equitably predict Mini-Mental State Examination (MMSE) scores, offering a scalable solution to streamline eligibility screening for Alzheimer's disease-modifying therapies while mitigating the educational and cultural biases of traditional testing.

The Gist

The Big Problem: A Traffic Jam at the Doctor's Office

Imagine a new, life-saving medicine has just been invented for Alzheimer's disease. It's like a "fire extinguisher" that can stop the fire of memory loss from spreading. But there's a catch: to get this medicine, you have to prove you are in the "early stages" of the fire.

Currently, doctors use an old, paper-based test called the MMSE (Mini-Mental State Examination) to decide who qualifies. Think of the MMSE as a handwritten map drawn by a human.

• It's slow: It takes a doctor 10–15 minutes to administer it.
• It's biased: The map is drawn in a specific language and culture. If you are from a different background or didn't go to school for as long, the map might look confusing to you, even if your brain is working just fine. This means many people get wrongly told they are "too sick" for the medicine, while others get wrongly told they are "too healthy."
• It's a bottleneck: With millions of people needing this test, there aren't enough doctors to give it out. It's like trying to get a ticket to a concert at a single, tiny window with a line stretching for miles.

The New Solution: A High-Speed Digital Scanner

The researchers at Linus Health built a new tool called the Digital Clock and Recall (DCR). Think of this as a high-tech, 3-minute video game played on an iPad.

Instead of just looking at the final answer (like "Did you draw the clock right?"), the DCR watches how you do it. It records:

• How fast your hand moves.
• Where you pause.
• The tone and speed of your voice when you speak.
• The pressure of your pen on the screen.

It's like a super-smart coach watching an athlete. A human coach might just say, "You missed the shot." The digital coach says, "You hesitated for 0.4 seconds before moving your arm, your grip tightened, and your voice cracked slightly. This tells us exactly how your brain is working."

The Magic Trick: The "Translator" Robot

The big question was: Can this fancy digital game replace the old paper test?

The researchers used Machine Learning (AI) to build a "translator." They fed the AI data from nearly 1,000 people who took both the old paper test and the new digital game. The AI learned the secret language connecting the two.

The Analogy: Imagine you have a friend who speaks only "Digital" and another who speaks only "Paper." The AI is the universal translator that listens to the Digital friend's complex story and instantly writes down the Paper friend's score.

The Results: Fast, Fair, and Accurate

The study found that the AI translator was incredibly good at its job:

1. It's Accurate: The difference between the AI's guess and the real human score was about 2 points.
   • Why is this good? Even when a human doctor gives the paper test to the same person twice, the score often changes by 2 to 4 points just because of tiredness or a bad day. The AI is just as reliable as a human doctor, but it never gets tired or biased.
2. It's Fair: The AI didn't care if you were White, Black, Hispanic, or had a lot of education. It gave accurate scores for everyone. It ignored the "noise" of culture and education that confuses the old paper test.
3. It's Fast: The whole thing takes 3 minutes and can be done by a nurse or even a patient themselves, freeing up the specialist doctors to focus on the people who really need help.

Why This Matters: Clearing the Road

By using this digital tool, we can create a "Digital Triage" system.

• Before: Everyone waits in a long line for a slow, biased paper test. Many get turned away unfairly.
• Now: The digital tool quickly scans everyone. It instantly says, "This person is likely in the 'Goldilocks zone' (not too sick, not too healthy) and needs to see a specialist."

This clears the traffic jam. It ensures that the life-saving Alzheimer's drugs go to the people who need them based on their actual brain health, not based on their accent, their education, or how long they've been waiting in line.

In short: The researchers built a fast, fair, digital bridge that connects modern technology to old medical rules, ensuring that the right patients get the right help, right now.

Technical Summary

1. Problem Statement

The recent FDA approval of anti-amyloid disease-modifying therapies (DMTs) like lecanemab and donanemab has created an urgent need to screen large populations of older adults to identify those eligible for treatment. Eligibility is currently gatekept by specific ranges on the Mini-Mental State Examination (MMSE) (e.g., scores between 20–29).

• Operational Bottleneck: The MMSE requires 10–15 minutes of direct clinician administration, which is unsustainable in primary care settings where visit times are short. This creates a massive bottleneck in patient triage, potentially delaying access to life-saving therapies.
• Systemic Bias: The MMSE is a pencil-and-paper test known to suffer from significant cultural, educational, and racial biases. It frequently misclassifies individuals from underrepresented racial/ethnic groups or those with lower educational attainment as cognitively impaired, threatening to exacerbate health disparities by denying them access to DMTs.
• Need for a Proxy: There is a critical need for a rapid, scalable, and equitable digital assessment that can accurately "crosswalk" to MMSE scores to facilitate efficient patient triage without relying on biased legacy instruments.

2. Methodology

The study utilized a retrospective analysis of data from two large, multi-site clinical studies coordinated by the Global Alzheimer's Platform Foundation (GAP).

• Datasets:
  • Training Cohort (Bio-Hermes-001): $N=945$ participants (Cognitively Unimpaired, Mild Cognitive Impairment [MCI], and probable Alzheimer's Dementia).
  • External Validation Cohort (Apheleia-001): $N=238$ participants (individuals with progressive cognitive concerns).
• Input Data (Digital Clock and Recall - DCR™):
  • The DCR is a 3-minute, FDA-listed digital assessment administered on an iPad.
  • Tasks: Immediate word recall, Digital Clock Drawing (Command and Copy conditions), and Delayed word recall.
  • Feature Extraction: Unlike traditional scoring which only looks at the final product, the DCR captures ~2,000 high-resolution process features, including:
    • Kinematics: Pen velocity, acceleration, stroke length, and pauses.
    • Temporal: Initiation time, inter-stroke intervals, and task duration.
    • Spatial: Clock face symmetry, numeral placement, and circularity.
    • Acoustic/Speech: Pitch, jitter, shimmer, speaking rate, and response latency during recall.
• Model Architecture:
  • Algorithm: A Poisson elastic net regression model (using the glmnet package).
  • Predictors: 307 DCR-derived metrics (drawing and speech features) plus Age.
  • Target Variable: The total MMSE score.
  • Training Strategy:
    • Data split: 70% training, 30% test.
    • Weighting: To address the sparse distribution of low MMSE scores in the dataset, the model applied inverse-rate weighting to scores below 23 and a secondary weight to scores above 27, ensuring the model learned the full spectrum of cognitive function.
    • Validation: 10-fold cross-validation on the training set, followed by external validation on the Apheleia cohort without retraining.

3. Key Contributions

• Digital Crosswalk: The study successfully developed a machine learning model that translates multimodal digital biomarkers from the DCR into a predicted MMSE score, bridging the gap between new digital tools and legacy regulatory requirements.
• Equity-Centric Design: The research explicitly tested for demographic fairness, demonstrating that the model's accuracy is not compromised by race or ethnicity, addressing a known failure point of the traditional MMSE.
• Process-Based Sensitivity: By leveraging millisecond-level kinematic and acoustic data rather than just final scores, the model captures subtle cognitive deficits that traditional scoring misses (e.g., "ceiling effects" in high-functioning individuals).

4. Results

• Prediction Accuracy:
  • Training/Test (Bio-Hermes): The model achieved a Root Mean Squared Error (RMSE) of 2.31.
  • External Validation (Apheleia): The model maintained robust generalizability with an RMSE of 2.62.
• Statistical Significance:
  • The prediction error (RMSE ~2.3) falls within the established test-retest reliability range of the manual MMSE itself (2–4 points). This indicates the digital prediction is statistically non-inferior to a human re-administration of the test.
  • The Reliable Change Index (RCI) for the model (4.22) was comparable to the RCI of the MMSE (4.10–4.42).
• Demographic Fairness:
  • The model exhibited consistent accuracy across racial and ethnic groups:
    • Race: White (RMSE = 2.34) vs. Non-White (RMSE = 2.14).
    • Ethnicity: Hispanic (RMSE = 2.26) vs. Non-Hispanic (RMSE = 2.31).
  • This contrasts sharply with the MMSE, which typically shows higher error rates for non-White and lower-educated populations.
• Feature Importance: The most predictive features included delayed recall accuracy, speech production rates, clock-drawing latency, and component placement, alongside age.

5. Significance and Clinical Implications

• Solving the Triage Bottleneck: The DCR takes <3 minutes and can be self-administered or administered by non-specialists. By providing an immediate, automated predicted MMSE score, it allows primary care providers to rapidly stratify patients into the "Goldilocks window" (MMSE 20–30) for DMT eligibility. This prioritizes referrals for confirmatory biomarker testing (PET/CSF) and specialist evaluation, potentially reducing wait times from years to months.
• Health Equity: By replacing a culturally biased paper test with a process-based digital assessment, the system reduces the risk of false positives for impairment in minority populations. This ensures that access to DMTs is determined by underlying pathology rather than sociodemographic factors.
• Scalability: The automated nature of the scoring allows for integration into Electronic Health Records (EHR) and high-throughput screening, essential for managing the millions of older adults who may be eligible for these new therapies.
• Future Outlook: While the study is cross-sectional, the authors posit that this "digital triage" engine represents a critical infrastructure upgrade for the era of disease-modifying Alzheimer's treatments, transforming eligibility assessment from a manual, biased bottleneck into a rapid, equitable, and data-driven process.