📊 statistics

Optimized questionnaire item selection for tracking the progression of motor symptoms in Parkinson's disease

This study demonstrates that using coordinate descent and adaptive selection algorithms to optimize item subsets for the MDS-UPDRS questionnaire significantly reduces uncertainty in estimating Parkinson's disease severity compared to traditional Fisher information ranking or random selection, offering a trade-off between methodological complexity and measurement precision.

Original authors: Karl Sigfrid, Ellinor Fackle-Fornius, Frank Miller

Published 2026-04-14

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Karl Sigfrid, Ellinor Fackle-Fornius, Frank Miller

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are a doctor trying to track how a patient's Parkinson's disease is progressing over time. To do this, you usually give them a very long, detailed questionnaire called the MDS-UPDRS. It's like a 30-minute marathon of questions about shaking, stiffness, walking, and daily tasks.

While this marathon gives you a very clear picture, it's exhausting for the patient (and the doctor). It's like asking someone to run a full marathon just to check if they can walk to the mailbox. The problem is, if the test takes too long, patients get tired, skip questions, or doctors might skip the test entirely in busy clinics.

The Goal: The authors of this paper wanted to find a way to shorten the test—maybe to just 5 or 10 questions—without losing the accuracy of the medical picture. They wanted to know: Which specific questions are the most important, and how do we pick them?

The Three Strategies (The "How-To" Guide)

The researchers tested three different ways to pick the best "short list" of questions from the original 34. Think of these strategies as three different ways to pack a suitcase for a trip where you only have space for a few items.

1. The "Star Player" Approach (Ranking by Fisher Information)

This is the most common method. Imagine you have a team of 34 players, and you want to pick the top 5. You look at their individual stats (how good they are at scoring) and just pick the 5 with the highest scores.

The Flaw: In this analogy, what if all 5 "star players" are all goalkeepers? You'd have a great defense, but no one to score goals. Similarly, this method picks questions that are all very good at measuring "average" patients, but they might all miss the patients who are very sick or very healthy. It's efficient, but it has blind spots.

2. The "Team Chemistry" Approach (Coordinate Descent)

This method is smarter. Instead of just picking the 5 best individual players, it looks at how the players work together. It asks: "If I pick Player A, who is the best partner to go with them to cover all bases?"

The Metaphor: It's like building a fantasy sports team. You don't just pick the top 5 scorers; you pick a mix of a striker, a defender, and a midfielder so your team is balanced and covers the whole field.
The Result: This method found a set of questions that gave a much clearer picture of the patient's condition, especially for those at the extremes (very mild or very severe symptoms), reducing the "guesswork" (uncertainty) by about 26% compared to just picking randomly.

3. The "Personal Trainer" Approach (Adaptive Selection)

This is the "best-case scenario" method. Imagine a personal trainer who watches you run, sees you struggling, and immediately hands you a different pair of shoes.

How it works: As the patient answers the first question, the computer calculates their score and instantly picks the perfect next question for that specific person.
The Catch: This is the "Holy Grail." It gives the best possible accuracy (reducing uncertainty by 34%), but it requires a complex, digital system that changes the test in real-time. It's like having a robot doctor who tailors the exam on the fly.

The Big Discovery

The researchers found a "sweet spot" in the math:

If you cut the test down to just 5 questions: Using the smart "Team Chemistry" method is a huge win. It's almost as good as the long test and much better than just picking questions at random.
If you keep 20 questions: The fancy methods don't help as much. If you have 20 questions, even a random pick is "okay," so the extra effort to find the "perfect" 20 isn't worth the trouble.

Why This Matters

Think of the original 34 questions as a giant, heavy toolbox.

Random Selection: Picking 5 tools at random from the box. You might get a hammer and a screwdriver, but maybe you miss the wrench you actually need.
The "Star Player" Method: Picking the 5 most expensive tools. They are great, but maybe they are all hammers.
The "Team Chemistry" Method: Picking 5 tools that cover every job you might need to do (hammer, screwdriver, wrench, pliers, tape measure).

The Takeaway:
You don't need to ask every single question to get a good diagnosis. By using a smart algorithm to pick a small, balanced set of questions (about 14 items instead of 34), doctors can get the same level of accuracy in about half the time. This reduces "survey fatigue" for patients, meaning they are more likely to stick with their treatment and get better care.

In short: Don't just pick the "best" questions; pick the questions that work best together.

1. Problem Statement

The Unified Parkinson's Disease Rating Scale (MDS-UPDRS) is the standard tool for tracking Parkinson's disease (PD) progression. However, the full questionnaire is time-consuming (approx. 30 minutes) and imposes a significant response burden on patients and clinicians, leading to survey fatigue and underutilization in non-research clinical settings.

The core challenge is to identify an optimal subset of items from the full 34-item motor function battery that minimizes the uncertainty of the latent trait estimate (disease severity) while significantly reducing the number of questions.

Key Theoretical Gap: Previous studies often selected items based on Expected Fisher Information (ranking items by individual informativeness). The authors argue that maximizing expected Fisher information does not necessarily minimize the Expected Standard Deviation (ESD) of the latent trait estimate. Fisher information is additive, but the reduction in estimation uncertainty is non-linear and depends on the complementarity of items across the entire latent trait distribution (especially in the tails).

2. Methodology

Data Source

Dataset: Parkinson's Progression Markers Initiative (PPMI) Clinical collection.
Sample: 4,283 patients (including diagnosed PD, healthy controls, and at-risk individuals) with longitudinal data.
Variables: 34 motor function items from MDS-UPDRS Part II (Patient-reported) and Part III (Clinician-rated). Responses are on a 5-level ordinal scale (0–4).
Structure: Longitudinal data with multiple visits per patient over several years.

Statistical Modeling

The authors employed a Longitudinal Graded Response Model (GRM) integrated within a Generalized Linear Mixed Model (GLMM) framework.

Latent Trait ( $\theta$ ): Modeled as a function of time, including a global intercept/slope and individual random intercepts/slopes to capture patient-specific progression.
Estimation: A two-stage Bayesian approach using Stan (MCMC sampling):
1. Stage 1 (Pre-estimation): Estimated item parameters (discrimination and lower thresholds) using only baseline data via the ltm package in R.
2. Stage 2 (Full GLMM): Estimated the full model with random effects, treating the pre-estimated lower thresholds and discrimination parameters as fixed to ensure convergence and identifiability.
Model Diagnostics: Convergence was verified using $\hat{R}$ statistics (mean 1.0003) and Effective Sample Size (ESS).

Item Selection Algorithms

The study compared three optimization strategies to select a subset of $K$ items ( $K \in \{1, \dots, 34\}$ ):

Ranking by Expected Fisher Information:
- Calculates the expected information for each item across the population distribution.
- Selects the top $K$ items.
- Limitation: Ignores item redundancy; tends to select items that cluster in the same region of the latent trait (usually the mean).
Coordinate Descent Local Search (Minimizing Expected Standard Deviation):
- Objective: Directly minimize the Expected Standard Deviation (ESD) of the latent trait estimate.
- Algorithm: A Fedorov-type exchange algorithm. It starts with a random subset of $K$ items and iteratively swaps items to find a local optimum where no single swap reduces the ESD further.
- Advantage: Accounts for the joint properties of the item set, ensuring coverage across the distribution (including tails).
Adaptive Selection (Best-Case Limit):
- Assumes the true latent trait score is known for each patient at the time of testing.
- Selects the $K$ items that maximize information conditioned on that specific score.
- Significance: Represents the theoretical upper bound of performance for any adaptive testing system.

3. Key Contributions

Theoretical Distinction: Demonstrated mathematically and empirically that maximizing Expected Fisher Information $\neq$ minimizing Expected Standard Deviation. The former favors items informative near the population mean, while the latter requires items that cover the tails of the distribution to reduce overall uncertainty.
Algorithmic Comparison: Introduced and compared a coordinate descent local search algorithm against standard ranking methods for longitudinal IRT data.
Longitudinal IRT Application: Successfully applied a complex, high-dimensional GLMM (8,739 parameters) to model PD progression, providing a robust framework for future item selection in longitudinal studies.
Practical Trade-off Analysis: Quantified the trade-off between methodological complexity (adaptive vs. static) and precision gains.

4. Results

Parameter Estimates

The model confirmed that symptom severity generally increases over time (mean slope $\approx 0.075$ ).
Item information functions varied significantly. Notably, items like "Posture" provided broad information across the distribution (flat, wide curves), whereas others were highly informative only in specific ranges.

Performance of Selection Methods

The study evaluated the reduction in Expected Standard Deviation (ESD) compared to random item selection for subsets of size 5, 10, and 20.

Subset Size ( $K$ )	Method	% Reduction in ESD vs. Random Selection
5 Items	Max Fisher Info (Ranking)	14%
	Coordinate Descent (Min ESD)	26%
	Adaptive Selection	34%
10 Items	Max Fisher Info	8%
	Coordinate Descent	22%
	Adaptive Selection	27%
20 Items	Max Fisher Info	6%
	Coordinate Descent	13%
	Adaptive Selection	16%

Key Findings:

Small Subsets: For small item sets (e.g., $K=5$ ), the Coordinate Descent method significantly outperforms simple ranking (26% vs. 14% reduction in error).
Diminishing Returns: As the subset size increases, the performance gap between methods narrows. With 20 items, the difference between the best and worst method is minimal.
Adaptive Limits: The adaptive method offers only a marginal gain over the static Coordinate Descent method, suggesting that the complexity of real-time adaptive testing may not be justified for this specific application.
Item Composition: The optimal sets for minimizing ESD included items with "flatter" information functions (e.g., Posture) that cover the tails, whereas the Fisher Information ranking selected items with high peaks near the mean.

5. Significance and Implications

Clinical Efficiency: The study proves that a reduced questionnaire (e.g., 14 items) can achieve the same precision as the full 34-item test if items are selected via Coordinate Descent. This could reduce administration time by ~60%, facilitating more frequent monitoring in clinical practice.
Methodological Guidance: The authors recommend Coordinate Descent as the optimal balance between precision and complexity. It is superior to simple ranking but avoids the logistical overhead of fully adaptive testing.
Future Directions: The paper suggests that for very small item sets, advanced optimization is critical. However, for larger sets, the benefit of complex algorithms diminishes. The methodology is applicable to any longitudinal health assessment where reducing patient burden is a priority.
Limitations: The adaptive results represent a "best-case" scenario assuming perfect knowledge of the latent trait. In practice, adaptive selection relies on provisional estimates, which may reduce the actual gain. Additionally, the current item pool lacks items that effectively measure very mild symptoms (left tail of the distribution).