Inverse Reinforcement Learning for Interpretable Keystroke Biomarkers in Parkinson's Disease

Navin Bondade

REVIEW 3 major objections 4 minor 1 cited by

Inverse Reinforcement Learning for Interpretable Keystroke Biomarkers in Parkinson's Disease

T0 review · 3 major / 4 minor · reviewed 2026-08-02 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read A per-subject speed-preference weight recovered from raw keystroke timings via IRL correlates with Parkinson's disease motor severity at r = -0.607, adding interpretable signal beyond raw typing speed.

desk verdict First IRL application to keystroke dynamics with a robust internal correlation, but the abstract's mPower external validation claim directly contradicts the full text and must be fixed before this is publishable. read the letter →

arxiv 2606.25270 v3 pith:RLLLXKMK submitted 2026-06-24 cs.LG

Navin Bondade This is my paper

classification cs.LG

keywords inversereinforcementlearningkeystrokedynamicsParkinson'sdiseasedigitalbiomarkerUPDRS-IIImaximumentropyrewardmotorseverity

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper tries to establish that a person's underlying preference for typing speed—recovered from raw keystroke timings without any clinical label—is a meaningful, interpretable marker of Parkinson's disease motor severity. Treating each keystroke transition as a discrete choice over flight-time bins, the authors fit a per-subject reward function by maximum-entropy inverse reinforcement learning. The recovered speed-preference weight correlates with UPDRS-III at r = -0.607, replicates across two sub-cohorts, survives controlling for raw typing speed, and shows high test-retest reliability. If true, this would make passive typing data a viable window into motor decline, with the interpretability of a reward weight rather than a black-box classifier. The paper is as much about validation discipline as the coefficient: it documents an identifiability fix, two implementation bugs with quantified negligible impact, and two non-surviving weights as negative results.

What carries the argument

The central mechanism is maximum-entropy inverse reinforcement learning over keystroke transitions. Flight times are discretized into K = 5 bins (quintiles of the pooled distribution); each transition is modeled as a discrete choice with reward R(s,a) = w·φ(s,a), where φ includes (1) negative flight time (speed preference), (2) negative deviation from a local context estimate (consistency), and (3) cross-hand cost (hand-alternation). Per-subject weights w are fit by maximizing the log-likelihood of observed bin choices against unchosen alternatives—the standard MaxEnt IRL gradient. The paper's key methodological move is diagnosing that an initial four-feature version had two features correla

What would settle it

Run the identical IRL pipeline on an independent keystroke dataset with clinician-rated UPDRS-III and no pooling leakage (e.g., building bins on a training set only); if the w_speed–severity correlation drops to null or changes sign, the central claim fails. A cheaper check: recompute all analyses with leave-one-subject-out bin construction on the same data and compare against r = -0.607.

Watch

Extended reading notes

Core claim

The central claim is that inverse reinforcement learning—normally used to infer what reward an agent is optimizing—can be applied to keystroke timings to recover a per-subject speed-preference weight, w_speed, that reflects the implicit cost the subject assigns to moving quickly between keys. In the neuroQWERTY MIT-CSXPD dataset (85 subjects, 42 with PD), w_speed correlates with clinician-rated UPDRS-III severity at r = -0.607 (95% CI [-0.770,-0.364], p < 0.001): a stronger preference for fast typing goes with less motor impairment. The result replicates in two independent sub-cohorts, remains significant after adjusting for mean and SD of flight time, adds incremental explained variance ove

Load-bearing premise

The load-bearing premise is that the pooled-flight-time quintile bins used to define the discrete choices are not materially corrupted by each subject's own data; the paper itself calls this a mild, disclosed form of data leakage rather than a corrected one.

Editorial extensions

If this is right

w_speed can be computed per subject from raw keystrokes alone, with no clinical labels during fitting, making it a candidate passive digital biomarker.
The weight adds significant variance beyond raw typing speed (R² 0.194 to 0.338), so it captures motor-relevant information not reducible to simple speed.
Test-retest reliability ICC(2,1) = 0.903 across clinic sessions means the measure is stable enough for longitudinal monitoring.
The model outputs an interpretable reward weight (why the person types as they do), unlike classifier scores.
The same pipeline recovered the signal on independent smartphone tapping data (r = -0.639), suggesting modality-transferability.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

The IRL framing suggests a family of interpretable digital biomarkers: any repeated motor action (mouse movement, touchscreen gestures, speech timing) could be cast as a discrete-choice reward-recovery problem.
If external keystroke replication succeeds, w_speed could serve as a progression endpoint in PD trials; the reported reliability is the precondition for that use.
The disclosed pooled-bin leakage means the exact correlation magnitude may shift slightly under leave-one-subject-out bin construction; a clean re-analysis would quantify this.
Given residual collinearity between speed and consistency features (r = 0.946), 'speed preference' may partly encode rhythmic consistency; the separable signal is real, but the named component may be a composite.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

The genuinely new thing here is using maximum-entropy IRL to recover a per-subject speed-preference weight from keystroke timing, along with an honest account of an identifiability failure, two implementation bugs found via self-audit, and two negative results. That kind of transparency is rare and worth respecting. The headline correlation (r=-0.607, n=42) holds up under per-cohort replication, partial correlations, sensitivity analyses, and permutation tests. The incremental R2 over raw typing speed is modest but real. This is a solid methods application, not a paradigm shift.

The soft spots are serious, though. The abstract advertises cross-modality external validation on mPower with n=200, r=-0.639, p=2.52e-24, and OR=14.19. The full text says explicitly in Section 7 that external replication has not yet been performed. There is no mPower analysis anywhere in the body. The abstract also reports AUC 0.605 and 0.750 that do not appear in the results section. That is an internal contradiction in the paper's central advertised evidence, not a minor presentational issue. It needs to be withdrawn or supported.

The reader's worry about pooled quintile bin boundaries is real but secondary—it is disclosed in Section 7 as a mild leakage, and the robustness checks suggest it is unlikely to flip the result. The residual collinearity between speed and consistency (r=0.946) is also disclosed, and the differential correlation pattern is some evidence of separable signal. The lack of code is a practical limitation, not a scientific flaw.

Who is this for? Researchers working on digital PD biomarkers and anyone interested in applying IRL to behavioral time series. They will find the identifiability correction and the bug-audit narrative instructive. But the paper is not ready as advertised. The author should either add the mPower analysis or revise the abstract to match the body. Without that fix, the claims outrun the evidence.

Recommendation: send to peer review, because the methodology and the internal validation are worth referee attention. But the editor should require the author to resolve the abstract/body contradiction before acceptance. A serious reviewer will catch it immediately; the author should be given the chance to fix it.

Referee Report

3 major / 4 minor

Summary. The paper applies maximum-entropy inverse reinforcement learning (IRL) to keystroke flight-time choices in the neuroQWERTY MIT-CSXPD dataset, fitting per-subject reward weights for speed preference, consistency, and hand-alternation. The central empirical claim is that the recovered speed-preference weight w_speed correlates with UPDRS-III motor severity (r=-0.607, p<0.001, n=42), with robustness checks including per-cohort replication, control for raw typing speed, sensitivity analyses, permutation tests, and cross-checks against other clinical scores. The authors also document an identifiability failure in an initial four-parameter model, two implementation bugs found by adversarial code review, and two negative results. The abstract additionally advertises cross-modality external validation on mPower smartphone tapping data (n=200), several AUC values, SHAP/LASSO comparisons, and an ICC reliability value, none of which appear in the full text. Section 7 explicitly states that external replication has not yet been performed, creating a direct internal contradiction.

Significance. If the single-dataset correlation is sustained, the work is a valuable methodological contribution: it is the first IRL formulation for keystroke dynamics, provides an interpretable per-subject biomarker, and the robustness/audit practices (documented identifiability failures, bug disclosure, sensitivity analyses) are exemplary and responsive to the field's replication problems. The disclosed limitations and negative results are honest. However, the current abstract materially overstates the evidence by claiming cross-modal external validation and reporting several metrics that are absent from the body. The body alone supports at most a well-validated internal, single-dataset finding. The significance as advertised is therefore not yet established, though the core single-dataset result is plausible and worth publishing after substantial revision.

major comments (3)

[Abstract; Section 7] The abstract claims 'Cross-modality external validation on independent mPower smartphone tapping data (n=200) recovers the same signal (r=-0.639, p=2.52e-24, OR=14.19),' and also reports model-free AUC=0.605, LOO-CV AUC=0.750, SHAP/LASSO comparisons, and ICC(2,1)=0.903. None of these results appears in the body, Methods, Results, Table 1, or Discussion. Section 7 explicitly states: 'external replication on an independent cohort is the natural next step and has not yet been performed.' This is a direct internal contradiction on the paper's strongest advertised evidence. The abstract must be revised to report only results that appear in the body, or the missing analyses must be added with full methods and results. As submitted, readers cannot verify the cross-modal, cross-device, cross-country validation claim, and the paper's advertised significance is unsupported.
[Section 5.2; Section 7] Action bins are constructed from pooled flight-time data across all 85 subjects, and each subject's own data contribute to the bin boundaries used in their own reward fitting. Section 7 discloses this as 'a mild, disclosed form of data leakage rather than a corrected one' with an estimated ~1% per-subject influence. However, this estimate is informal; per-subject contributions to quantile boundaries can be non-uniform, especially under heavy-tailed flight-time distributions, and could systematically bias w_speed in a way that inflates the UPDRS correlation. Please provide a leave-one-subject-out (or nested/out-of-fold) bin-construction analysis and report r, confidence interval, and p for the w_speed-UPDRS correlation under that procedure. If the correlation is materially unchanged, the concern is resolved; if not, the claim must be qualified.
[Section 5.3; Section 7] After merging the two collinear terms, the three-parameter model still has r=0.946 collinearity between phi_speed and phi_consistency. The paper interprets the divergent correlations with UPDRS (w_speed r=-0.607 vs. w_consistency r=0.164) as evidence that 'genuine separable signal survives,' but with collinearity this high, individual weight estimates can be unstable and poorly identified. A weight's correlation with an external variable does not, by itself, establish that the weight is a reliable standalone parameter. Please provide a condition-number or profile-likelihood diagnostic, or a regularization/perturbation stability analysis, demonstrating that w_speed is identifiable enough to serve as a per-subject biomarker. Without such an analysis, the claim that the identifiability problem is 'resolved' is too strong.

minor comments (4)

[Abstract; Table 1] The full-text abstract states stability across 'nine sensitivity configurations,' but Table 1 reports '5 configurations' for bin count and window sensitivity. Please reconcile these numbers.
[Abstract; Table 1] The abstract reports a 95% CI of [-0.770, -0.364] for r=-0.607, while Table 1 reports a bootstrap 95% CI of [-0.767, -0.369]. Specify which CI is reported (e.g., Fisher z-transformed vs. bootstrap) and use consistent values.
[Section 3] The Tappy negative result is based on an '80-subject balanced sample' drawn from a pool of 227 subjects. The selection criteria are not stated; if the sample was not the full available set, the negative result may not be representative. Please clarify the selection procedure.
[Section 5.2] The hand-assignment feature uses a standard QWERTY touch-typing key assignment, but many users do not touch-type. A brief acknowledgment of this limitation (or a sensitivity check using per-subject alternation patterns) would strengthen the manuscript.

Circularity Check

0 steps flagged · score 0.0 of 10

No circular derivation: the IRL weight is fit from keystroke timing without clinical labels, so the UPDRS correlation is an external association, not a fitted prediction. The disclosed bin-boundary leakage is a bias concern, not a circularity.

full rationale

The central derivation is not circular. The per-subject speed-preference weight w_speed is recovered by maximum-entropy IRL from raw keystroke timing only; UPDRS-III labels are not used during fitting (Section 5.2, Section 6.1), so the reported correlation with severity is an external association rather than a fitted input renamed as a prediction. The mild leakage disclosed in Section 7 — action-bin boundaries from pooled flight-time data including each subject's ~1% contribution — is a preprocessing bias that could affect the correlation's magnitude, but it does not make the correlation equal to the fitted quantity by construction; it is not a definitional reduction. There are no load-bearing self-citations, no imported uniqueness theorem from the authors' own prior work, and no ansatz smuggled in via citation. The most serious issue in the manuscript is not circularity: the Abstract claims cross-modality external validation on mPower tapping data (n=200, r=-0.639, OR=14.19) and model-free AUC/LOO-CV AUC values that do not appear anywhere in the body, while Section 7 states external replication 'has not yet been performed.' This internal contradiction is a reporting/integrity problem that should be corrected or retracted, but it falls outside the circularity patterns defined here. Therefore the circularity score is 0.

Assumptions & free parameters 4 free parameters · 5 assumptions · 1 invented entities

The central model relies on standard MaxEnt IRL assumptions plus a handful of dataset-specific choices (binning, thresholds). No new physical entities are postulated; the only invented construct is the interpretable reward weight, whose external validity is currently unsubstantiated.

free parameters (4)

Number of flight-time bins K = 5
Discretization granularity; sensitivity checked with K=4,5,6.
Rolling window size = 10
Context window for rolling mean; sensitivity checked with 5,10,15.
Flight-time outlier threshold = 3 seconds
Threshold for removing flight-time outliers; chosen to match prior work.
Pooled flight-time quintile bin boundaries = Not reported numerically
Bin edges estimated from all 85 subjects' pooled data; each subject's contribution ~1%; used to define action bins and representative values.

assumptions (5)

domain assumption Keystroke transitions follow a maximum-entropy softmax policy over the defined reward features (Section 5.2).
Assumes the generative model of keystroke timing is equivalent to an agent maximizing reward with softmax noise.
domain assumption The three reward features (speed, consistency, hand-alternation) capture the relevant motor decision process.
If important features are missing, recovered weights may be biased.
domain assumption UPDRS-III is a valid and reliable measure of PD motor severity.
Used as ground truth for correlation; measurement error would attenuate or distort the association.
domain assumption Data cleaning (removing rows with hold time outside [0,5) s, flight time >3 s) does not introduce systematic selection bias.
If removed rows are not random, the recovered weights could be biased.
ad hoc to paper The pooled bin boundaries are only mildly influenced by each subject's own data (~1% per subject), so the disclosed leakage is negligible.
This is the paper's own stated assumption, not independently verified.

invented entities (1)

IRL-derived speed-preference weight (w_speed)
purpose: Latent variable summarizing an individual's implicit cost assigned to fast typing
Correlates with UPDRS within the same dataset, but no external validation is actually provided in the full text; the abstract's mPower claim is absent from the manuscript. Thus it lacks a falsifiable handle outside this paper.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Inverse Reinforcement Learning for Interpretable Keystroke Biomarkers in Parkinson's Disease." pith.science (2026). https://pith.science/paper/RLLLXKMK

@misc{pith2026260625270,
  author       = {Pith},
  title        = {Pith review of: Inverse Reinforcement Learning for Interpretable Keystroke Biomarkers in Parkinson's Disease},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/RLLLXKMK}},
  note         = {Machine review of arXiv:2606.25270}
}

read the original abstract

Keystroke dynamics offer a passive window into motor function, but existing work extracts aggregate typing statistics and trains classifiers for PD/control discrimination, foregoing interpretability and rarely reporting reliability. We instead apply maximum-entropy inverse reinforcement learning (IRL) to raw keystroke timing, recovering a per-subject speed-preference weight (w_speed) reflecting the implicit cost assigned to fast movement, without any clinical label during fitting. On the neuroQWERTY MIT-CSXPD dataset (85 subjects, 42 PD), we diagnose and correct a feature collinearity failure in an initial four-parameter decomposition, yielding an identifiable three-parameter model. The recovered w_speed correlates with UPDRS-III motor severity at r=-0.607 (95% CI [-0.770,-0.364], p<0.001, n=42), replicates across two independent sub-cohorts (r=-0.720, r=-0.588), and retains significant partial correlation after controlling for mean and SD of flight time (r=-0.371, p=0.016). It outperforms SHAP and LASSO on the same proxy features (r=+0.362 and r=+0.410) while additionally providing per-subject, interpretable output. A model-free AUC of 0.605 and LOO-CV AUC of 0.750 (95% CI [0.644,0.847]) confirm discriminative value. Test-retest reliability across clinic sessions yields ICC(2,1)=0.903 (95% CI [0.842,0.971]); no prior keystroke-PD study has reported formal reliability. Two other recovered weights (consistency, hand-alternation) did not survive confound checks, strengthening credibility of the surviving signal. Cross-modality external validation on independent mPower smartphone tapping data (n=200) recovers the same signal (r=-0.639, p=2.52e-24, OR=14.19), confirming convergent validity across modality, device, and country.

Figures

Figures reproduced from arXiv: 2606.25270 by the authors.

Review history (3 revisions) →

Discussion (0). Sign in to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

Selective Impairment of Motor Recovery from Typing Errors in Parkinson's Disease: A Survival Analysis
q-bio.NC 2026-06 conditional novelty 6.0 of 10

Post-error recovery time after backspace typing errors correlates with Parkinson's disease severity, while pre-error keystroke instability does not.

Reference graph

Works this paper leans on

12 extracted references · 1 linked inside Pith · cited by 1 Pith paper

[1]

Adams, W.R. (2017). High-accuracy detection of early Parkinson's Disease using multiple characteristics of finger movement while typing. PLOS ONE, 12(11), e0188226

2017
[2]

Alfalahi, H., Khandoker, A.H., Chowdhury, N., Iakovakis, D., Dias, S.B., Chaudhuri, K.R., & Hadjileontiadis, L.J. (2022). Diagnostic accuracy of keystroke dynamics as digital biomarkers for fine motor decline in neuropsychiatric disorders: A systematic review and meta-analysis. Scientific Reports, 12, 7690

2022
[3]

Arroyo-Gallego, T., Ledesma-Carbayo, M.J., Butterworth, I., Matarazzo, M., Montero-Escribano, P., Puertas-Martin, V., Gray, M.L., Sanchez-Ferro, A., & Giancardo, L. (2017). Detecting motor impairment in early Parkinson's disease via natural typing interaction with keyboards: validation of the neuroQWERTY approach in an uncontrolled at-home setting. Journa...

2017
[4]

Dorsey, E.R., et al. (2018). Global, regional, and national burden of Parkinson's disease, 1990--2016: a systematic analysis for the Global Burden of Disease Study 2016. The Lancet Neurology, 17(11), 939--953

2018
[5]

Francesconi, A., Cappetta, D., Rebecchi, F., Soda, P., Guarrasi, V., & Sicilia, R. (2025). Cross-dataset Multivariate Time-series Model for Parkinson's Diagnosis via Keyboard Dynamics. arXiv preprint

2025
[6]

Giancardo, L., S\'anchez-Ferro, A., Arroyo-Gallego, T., Butterworth, I., Mendoza, C.S., Montero, P., Matarazzo, M., Obeso, J.A., Gray, M.L., & San Jos\'e Est\'epar, R. (2016). Computer keyboard interaction as an indicator of early Parkinson's disease. Scientific Reports, 6, 34468

2016
[7]

Goldberger, A., Amaral, L., Glass, L., Hausdorff, J., Ivanov, P.C., Mark, R., Mietus, J.E., Moody, G.B., Peng, C.K., & Stanley, H.E. (2000). PhysioBank, PhysioToolkit, and PhysioNet: Components of a new research resource for complex physiologic signals. Circulation, 101(23), e215--e220

2000
[8]

Iakovakis, D., et al. (2018). Touchscreen typing-pattern analysis for detecting fine motor skills decline in early-stage Parkinson's disease. Scientific Reports, 8, 1--13

2018

Show all 12 references

[9]

Ng, A.Y., & Russell, S.J. (2000). Algorithms for inverse reinforcement learning. In Proceedings of the Seventeenth International Conference on Machine Learning

2000
[10]

arXiv preprint arXiv:2507.06326

Sample-Efficient Reinforcement Learning Controller for Deep Brain Stimulation in Parkinson's Disease (2025). arXiv preprint arXiv:2507.06326

2025 arXiv
[11]

Tripathi, S., Arroyo-Gallego, T., & Giancardo, L. (2022). Keystroke-Dynamics for Parkinson's Disease Signs Detection in an At-Home Uncontrolled Population: A New Benchmark and Method. IEEE Transactions on Biomedical Engineering

2022
[12]

Ziebart, B.D., Maas, A., Bagnell, J.A., & Dey, A.K. (2008). Maximum entropy inverse reinforcement learning. In Proceedings of the 23rd AAAI Conference on Artificial Intelligence

2008

Pith tools

Reviewed August 2, 2026 · model on record in the stance chip above.

[1] [1]

Adams, W.R. (2017). High-accuracy detection of early Parkinson's Disease using multiple characteristics of finger movement while typing. PLOS ONE, 12(11), e0188226

2017

[2] [2]

Alfalahi, H., Khandoker, A.H., Chowdhury, N., Iakovakis, D., Dias, S.B., Chaudhuri, K.R., & Hadjileontiadis, L.J. (2022). Diagnostic accuracy of keystroke dynamics as digital biomarkers for fine motor decline in neuropsychiatric disorders: A systematic review and meta-analysis. Scientific Reports, 12, 7690

2022

[3] [3]

Arroyo-Gallego, T., Ledesma-Carbayo, M.J., Butterworth, I., Matarazzo, M., Montero-Escribano, P., Puertas-Martin, V., Gray, M.L., Sanchez-Ferro, A., & Giancardo, L. (2017). Detecting motor impairment in early Parkinson's disease via natural typing interaction with keyboards: validation of the neuroQWERTY approach in an uncontrolled at-home setting. Journa...

2017

[4] [4]

Dorsey, E.R., et al. (2018). Global, regional, and national burden of Parkinson's disease, 1990--2016: a systematic analysis for the Global Burden of Disease Study 2016. The Lancet Neurology, 17(11), 939--953

2018

[5] [5]

Francesconi, A., Cappetta, D., Rebecchi, F., Soda, P., Guarrasi, V., & Sicilia, R. (2025). Cross-dataset Multivariate Time-series Model for Parkinson's Diagnosis via Keyboard Dynamics. arXiv preprint

2025

[6] [6]

Giancardo, L., S\'anchez-Ferro, A., Arroyo-Gallego, T., Butterworth, I., Mendoza, C.S., Montero, P., Matarazzo, M., Obeso, J.A., Gray, M.L., & San Jos\'e Est\'epar, R. (2016). Computer keyboard interaction as an indicator of early Parkinson's disease. Scientific Reports, 6, 34468

2016

[7] [7]

Goldberger, A., Amaral, L., Glass, L., Hausdorff, J., Ivanov, P.C., Mark, R., Mietus, J.E., Moody, G.B., Peng, C.K., & Stanley, H.E. (2000). PhysioBank, PhysioToolkit, and PhysioNet: Components of a new research resource for complex physiologic signals. Circulation, 101(23), e215--e220

2000

[8] [8]

Iakovakis, D., et al. (2018). Touchscreen typing-pattern analysis for detecting fine motor skills decline in early-stage Parkinson's disease. Scientific Reports, 8, 1--13

2018