Confidence Filtering in Video-Based Gait Assessment: Measurement Error versus Selection Shift

David Regev; Eran Beeri Bamani

AI summary · arXiv: 2610.39704v1 · cs.CV

Confidence Filtering in Video-Based Gait Assessment: Measurement Error versus Selection Shift

David Regev, Eran Beeri Bamani

Original paper Submitted to arXiv on 29 Oct. 2026 6 pages + 2-page appendix, 4 figures, 2 supplementary figures, 2 tables. Submitted to the 2027 IEEE International Conference on Rehabilitation Robotics (ICORR) License: not stated (all rights reserved) Also in: fr

How to read this summary? Illustrated guide

The essentials

Confidence filtering slightly shifted assessed movement without improving clinical prediction in a Parkinson-related cohort; simulations showed that its effects depend on how confidence relates to movement and error.

Based on the abstract only: check the details in the original paper.

Question

Does retaining high-confidence observations in video-based gait assessment reduce measurement error without changing the movement being assessed? Can unchanged clinical prediction error establish that filtering preserves the assessed movement?

Method

The authors compared confidence selection with budget-matched random selection in a Parkinson-related cohort of 96 participants and an NF2-related replication cohort of 18 participants. They assessed movement summaries and held-out Mini-BESTest error, examined a policy preserving motion-group proportions, and decomposed and simulated measurement error and selection shift using the cohort's window structure.

Findings

Result In the Parkinson-related cohort, confidence selection shifted the movement summary by +0.05 SD while held-out Mini-BESTest error remained essentially unchanged.
See the passage in the paper
In the Parkinson-related cohort, confidence selection produced a small but consistent shift in the observed movement summary (+0.05 SD, 95% CI 0.03 to 0.07), while held-out Mini-BESTest error stayed flat (-0.04 points, -0.13 to 0.05); a policy preserving motion-group proportions reduced the shift, while NF2 showed no shift and its predictor did not beat a constant baseline.
Result Preserving motion-group proportions reduced the shift, whereas the NF2 cohort showed no shift and its predictor failed to outperform a constant baseline.
See the passage in the paper
In the Parkinson-related cohort, confidence selection produced a small but consistent shift in the observed movement summary (+0.05 SD, 95% CI 0.03 to 0.07), while held-out Mini-BESTest error stayed flat (-0.04 points, -0.13 to 0.05); a policy preserving motion-group proportions reduced the shift, while NF2 showed no shift and its predictor did not beat a constant baseline.
Result The decomposition and simulations indicated that filtering reduces local error when error depends on confidence but biases the endpoint when confidence tracks movement more strongly than error.
See the passage in the paper
Filtering lowers local error when error depends on confidence, but biases the endpoint once confidence tracks movement more than error.
Result Confidence was weakly associated with movement in the Parkinson-related data, with rho = 0.14.
See the passage in the paper
In the Parkinson-related data, confidence was only weakly tied to movement (rho = 0.14).
Authors’ interpretation The authors conclude that unchanged clinical prediction error in small cohorts cannot establish that confidence filtering preserves the movement being assessed.
See the passage in the paper
Unchanged clinical error from such small-cohort predictors therefore cannot show that quality filtering preserves the assessed movement.

Limitations

The NF2 replication predictor did not outperform a constant baseline, limiting what its clinical prediction performance can establish about filtering.
See the passage in the paper
In the Parkinson-related cohort, confidence selection produced a small but consistent shift in the observed movement summary (+0.05 SD, 95% CI 0.03 to 0.07), while held-out Mini-BESTest error stayed flat (-0.04 points, -0.13 to 0.05); a policy preserving motion-group proportions reduced the shift, while NF2 showed no shift and its predictor did not beat a constant baseline.
The conclusion that filtering is harmless depends on well-tracked windows being unbiased; an alternative systematic-error assumption allowed filtering to add up to 0.58 degrees.
See the passage in the paper
There, filtering is harmless if well-tracked windows are unbiased. If instead systematic error only redistributes within participants, filtering can add up to 0.58 degrees.

Why it matters

The authors distinguish reducing local measurement error from preserving the movement represented by a gait endpoint. They argue that stable clinical prediction performance alone is insufficient evidence that confidence filtering leaves the assessed movement unchanged.

Written by AI from the paper. Passages are checked word for word against the text we analysed; the rest may contain errors.

AI-generated keywords: Confidence filtering; Video-based gait assessment; Measurement error; Selection shift; Clinical prediction

Read further

Comprehensive Summary
Layman's Summary
Blog article
Original abstract

David Regev and Eran Beeri Bamani’s paper, “Confidence Filtering in Video-Based Gait Assessment: Measurement Error versus Selection Shift,” examines whether retaining only high-confidence pose observations improves gait assessment without altering the movement being assessed. Video-based systems often discard low-confidence observations on the assumption that they are less accurate. However, this filtering may also preferentially retain particular types of movement. The central question is whether confidence filtering merely reduces measurement error or also changes the composition of the observations used to summarize gait. This summary is based on the supplied abstract. The full text is unavailable because of licensing restrictions, so detailed methods and additional analyses cannot be assessed. The authors compared confidence-based selection with budget-matched random selection in a Parkinson-related cohort of 96 participants and an NF2-related replication cohort of 18 participants. Matching the number of retained observations helps distinguish the effects of confidence-based selection from those of simply using less data. The authors examined changes in the observed movement summary and held-out prediction error for the Mini-BESTest, a clinical balance assessment. They also evaluated a selection policy designed to preserve the proportions of different motion groups. To investigate the mechanisms underlying the observed changes, they decomposed measurement error and selection shift and simulated both using the cohorts’ window structure. The abstract does not specify the movement-summary definition, filtering thresholds, predictor specifications, or simulation procedures. In the Parkinson-related cohort, confidence filtering produced a small but consistent increase in the observed movement summary: 0.05 standard deviations, with a 95% confidence interval of 0.03 to 0.07. By contrast, the change in held-out Mini-BESTest prediction error was close to zero: −0.04 points, with an interval of −0.13 to 0.05. Thus, the movement summary shifted measurably without a clear improvement or deterioration in clinical prediction error. Preserving motion-group proportions reduced this shift, suggesting that the composition of retained observations matters. In the NF2-related cohort, the authors found no shift, but the predictor did not outperform a constant baseline. This replication therefore offers limited reassurance about the relationship between filtering and clinically useful prediction. The decomposition and simulations distinguish conditions under which filtering helps from those under which it introduces bias. Filtering reduces local measurement error when confidence is informative about that error. However, it can bias the final endpoint when confidence tracks movement more strongly than it tracks measurement error. In the Parkinson-related data, confidence was only weakly associated with movement (rho = 0.14). Under the assumption that well-tracked windows are unbiased, filtering was harmless in the authors’ analysis. Under an alternative assumption that systematic error only redistributes within participants, filtering could add up to 0.58 degrees. The main conclusion is that stable clinical prediction error does not establish that confidence filtering preserves the movement being assessed. Filtering may alter an endpoint both by affecting measurement error and by changing which movements remain represented. Important limitations include the modest cohort sizes, particularly the 18-person NF2 replication cohort; the weak NF2 predictor; and the sensitivity of simulated effects to assumptions about systematic error. The abstract supports a conditional conclusion rather than a blanket rejection of filtering: its consequences depend on how confidence relates to both accuracy and movement.

When Cleaner Video Data Changes the Gait You Measure

Video-based gait assessment promises a practical way to study how people walk and move. But turning video into reliable measurements involves a consequential decision: which observations should count? Pose-estimation systems often assign confidence scores to their observations, and discarding low-confidence data can seem like an obvious quality-control step. Yet cleaner-looking data may not represent the same movement as the original recording.

That tension is the focus of “Confidence Filtering in Video-Based Gait Assessment: Measurement Error versus Selection Shift,” by David Regev and Eran Beeri Bamani. Their central question is whether confidence filtering simply removes measurement error—or also changes the mix of movements used to summarize gait. The findings suggest that these effects need to be evaluated separately.

Better Measurements, or a Different Sample?

Confidence filtering rests on an appealing assumption: observations that a system tracks confidently are more accurate. If that is true, removing uncertain observations could improve the resulting assessment. But confidence may also depend on the movement itself. Some movements may be easier to track than others, meaning that filtering could preferentially retain certain kinds of motion.

This creates two distinct mechanisms. Measurement error concerns how accurately retained observations describe movement. Selection shift concerns which movements remain represented after filtering. A filter could reduce error in individual observations while making the overall movement summary less representative.

As an illustration, imagine a recording containing two types of movement, one consistently tracked with greater confidence. Selecting only the highest-confidence observations would change their relative proportions. Even if each retained observation were accurate, the final summary might describe a narrower slice of the person’s movement.

How the Researchers Tested Filtering

The authors compared confidence-based selection with budget-matched random selection in a Parkinson-related cohort of 96 participants and an NF2-related replication cohort of 18 participants. Budget matching means retaining the same number of observations in each approach. This helps distinguish the effect of selecting by confidence from the effect of simply using less data.

They examined changes in the observed movement summary and held-out prediction error for the Mini-BESTest, a clinical balance assessment. They also tested a selection policy designed to preserve the proportions of different motion groups. To explore possible mechanisms, they decomposed measurement error and selection shift and simulated both using the cohorts’ window structure.

This account is based on the supplied abstract. The full text is unavailable because of licensing restrictions, so detailed methods and additional analyses cannot be assessed. The abstract does not specify the movement-summary definition, filtering thresholds, predictor specifications, or simulation procedures.

A Movement Shift Without a Clear Prediction Benefit

In the Parkinson-related cohort, confidence filtering increased the observed movement summary by 0.05 standard deviations, with a 95% confidence interval of 0.03 to 0.07. This was a small but consistent shift.

Meanwhile, the change in held-out Mini-BESTest prediction error was close to zero: −0.04 points, with an interval of −0.13 to 0.05. There was therefore no clear improvement or deterioration in clinical prediction error, despite a measurable change in the movement summary.

That distinction is the study’s central lesson. Stable prediction error does not establish that filtering preserves the movement being assessed. A clinical prediction task and a movement endpoint capture different aspects of performance; stability in one does not guarantee stability in the other.

Why Movement Composition Matters

The policy that preserved motion-group proportions reduced the shift. This supports the interpretation that the composition of retained observations matters, not just their individual quality. It also points toward a potentially useful safeguard: evaluating whether filtering disproportionately removes particular movement groups.

In the NF2-related cohort, the authors found no shift. However, the predictor did not outperform a constant baseline. Combined with the small sample of 18 participants, this limits how much reassurance the replication offers about filtering and clinically useful prediction.

When Filtering Helps—and When It Can Bias Results

The decomposition and simulations support a conditional account. Filtering can reduce local measurement error when confidence is informative about that error. But it can bias the final endpoint when confidence tracks movement more strongly than it tracks measurement error.

In the Parkinson-related data, confidence was only weakly associated with movement, with rho = 0.14. Even so, the implications depended on assumptions about systematic error. Under the assumption that well-tracked windows were unbiased, filtering was harmless in the authors’ analysis. Under an alternative assumption that systematic error only redistributes within participants, filtering could add up to 0.58 degrees. That figure is assumption-dependent, not a universally observed filtering effect.

The Takeaway: Validate the Filter, Not Just the Predictor

This paper does not justify rejecting confidence filtering outright. Instead, it challenges the idea that confidence automatically identifies the best data for every assessment. Modest cohort sizes, a weak NF2 predictor, and simulation sensitivity all warrant caution. The practical message is clear: assess both accuracy and representativeness. A filter should be judged not only by whether predictions remain stable, but also by whether it changes the movement the assessment is intended to describe.

Created on 02 Nov. 2026

Was this synthesis useful?

Score: 0

Ask questions about this paper to our AI assistant

How to chat with this paper? Illustrated guide

Compare with the PDF

You can also chat with multiple papers at once here.

⚠The license of the paper does not allow us to build upon its content and the AI assistant only knows about the paper metadata rather than the full article.

AI assistant instructions?

Tools for this paper

Explain it differently 3 free rewrites left today

Save to a collection Group papers in collections to compare them, chat with them and export them. Create a free account Log in

Export BibTeX · RIS (Zotero, Mendeley, EndNote) · Obsidian (Markdown) · Notion (Markdown)

Advanced tools Premium

For researchers who follow a field closely: citation alerts, journal-club slides, version comparison and Zotero export. Discover Premium

Zotero

Citation alerts Citation alerts guide

Who works on it? The people and institutions who publish on its topic

Weekly letter Presented in "Artificial intelligence this week" of 2 November 2026. Artificial intelligence this week: when better scores hide the wrong answer

Journal club Premium+ Slides to present it Propose for the journal club

Versions What changed between versions

Talk about it in public Premium+ Press kit: release, questions, hype check Press kit guide

Disclaimer: The AI-based summarization tool and virtual assistant provided on this website may not always provide accurate and complete summaries or responses. We encourage you to carefully review and evaluate the generated content to ensure its quality and relevance to your needs.

Confidence Filtering in Video-Based Gait Assessment: Measurement Error versus Selection Shift

The essentials

Question

Method

Findings

Limitations

Why it matters

Read further

Definitions

When Cleaner Video Data Changes the Gait You Measure

Better Measurements, or a Different Sample?

How the Researchers Tested Filtering

A Movement Shift Without a Clear Prediction Benefit

Why Movement Composition Matters

When Filtering Helps—and When It Can Bias Results

The Takeaway: Validate the Filter, Not Just the Predictor

Ask questions about this paper to our AI assistant

Tools for this paper

Advanced tools Premium

Papers around this one

Closest papers

Prior work it builds on

Go further

Cite this paper (BibTeX)