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

Abstract: Keystroke dynamics have been explored extensively as a passive digital biomarker for Parkinson's disease (PD), typically by extracting summary statistics from typing timing and training a classifier to discriminate PD from healthy controls. We instead apply inverse reinforcement learning (IRL) to keystroke data, modeling each keystroke as a discrete choice over typing speed and recovering, per subject, an interpretable reward function that explains their observed timing behavior. To our knowledge this is the first application of IRL to keystroke dynamics. On the public neuroQWERTY MIT-CSXPD dataset (85 subjects, 42 with PD), an initial four-parameter reward decomposition (speed, effort, smoothness, hand-alternation cost) was found to suffer severe feature collinearity between two terms ($r=1.000$ in typical contexts); we diagnose and correct this, yielding an identifiable three-parameter model. The recovered speed-preference weight correlates with UPDRS-III severity at $r=-0.607$ ($p<0.001$, $n=42$), replicates independently across two sub-cohorts, is stable across nine sensitivity configurations, and retains a statistically significant contribution beyond raw typing speed alone (incremental $R^2$ from 0.194 to 0.338, $p=0.006$). Two other recovered weights (consistency, hand-alternation) did not survive confound checks and are reported as negative results. We document two implementation bugs found during adversarial code review (session-boundary contamination, a rolling-window data leakage) and show the headline result is materially unchanged after fixing both. We discuss this result in the context of a literature where reported accuracies vary widely between studies (pooled AUC 0.85, I^2=94% in a 2022 meta-analysis), and argue that the validation process itself, not only the correlation coefficient, is part of the contribution.

• The paper applies inverse reinforcement learning to keystroke dynamics to extract latent reward functions as interpretable motor biomarkers for Parkinson’s disease.
• The approach recovers a speed preference weight that correlates with UPDRS-III motor severity (r = -0.607) and boosts model explanatory power from an R² of 0.194 to 0.338.
• The study addresses feature collinearity by merging effort and smoothness into a 'consistency' term, highlighting robust validation and transparency in digital biomarker research.

Inverse Reinforcement Learning for Interpretable Keystroke Biomarkers in Parkinson's Disease: An Expert Summary

Motivation and Problem Formulation

Conventional research on digital biomarkers for Parkinson’s Disease (PD) using keystroke dynamics relies predominantly on summary statistics extracted from timing features and applies standard supervised learning pipelines to diagnose PD or estimate severity. While numerous high-accuracy results have been reported in controlled settings, these models offer limited interpretability, do not explain why observed behavior is discriminative, and often exhibit wide performance variability across studies and datasets, raising concerns about generalization and mechanistic insight.

This work advances the field by applying inverse reinforcement learning (IRL) to the analysis of keystroke timings in PD. Here, each keystroke is formulated as a sequential decision over discretized speed actions, and the IRL framework is used to recover, per subject, the latent reward function that best explains their typing patterns. The recovered reward parameters afford direct interpretability—quantifying individual preferences for speed, rhythmic consistency, and hand-alternation efficiency as behavioral biomarkers.

Methodological Contributions

A maximum-entropy IRL approach is applied to the neuroQWERTY MIT-CSXPD dataset (85 subjects, 42 with PD) with raw keystroke data, using key reward features:

• Speed: Mean keystroke flight time, capturing bradykinesia
• Consistency: Deviation from a subject-specific rolling mean (combining initial 'effort' and 'smoothness' terms into a single, identifiable feature after feature collinearity was diagnosed)
• Hand-Alternation Cost: Cost of using the same vs. opposite hand per standard QWERTY typing

The paper identifies and corrects a critical identifiability issue: the initially independent effort and smoothness terms in the reward function exhibited perfect collinearity (r = 1.000 in representative contexts), precluding unique recovery of their respective weights. This was resolved through their merger into a unified 'consistency' term, which restored numerical stability and sign consistency for the learned parameters.

An intensive validation suite includes fixing two implementation bugs found during adversarial review (session-boundary contamination and a data leakage windowing error), demonstrating that the primary findings are robust to these corrections.

Key Results and Numerical Claims

The principal empirical finding is that the IRL-recovered speed preference weight ($w_{\text{speed}}$) correlates with clinical UPDRS-III motor severity at $r = -0.607$ (p < 0.001, n = 42 across the PD cohort). This relationship is both highly robust across two independent sub-cohorts (CS1: r = -0.720, CS2: r = -0.588) and invariant under a range of sensitivity analyses and confounder adjustments:

• The model's incremental explanatory power: Adding $w_{\text{speed}}$ to a model with raw typing speed raises $R^2$ from 0.194 to 0.338 (p = 0.006), indicating that IRL recovers information beyond basic timing statistics.
• Bootstrap and permutation tests confirm statistical significance (CI excludes zero, permuted p < 0.0001).
• Additional reward terms (consistency, hand-alternation) do not independently correlate with PD severity after confound control, and are reported as negative results.

The approach fails to yield significant discrimination in the Tappy dataset, a common challenge in the literature, underscoring the impact of dataset collection protocol on keystroke biomarker efficacy.

Implications and Theoretical Insights

This study demonstrates, for the first time, that interpretable reward-based models inferred via IRL can capture clinically relevant motor impairment biomarkers from passive digital trace data. The explicit modeling of behavior as optimized under latent preferences offers substantial interpretability and a bridge to formal decision-theoretic understanding of fine-motor decline in PD. Feature identifiability is directly addressed and documented, highlighting the importance of robust, transparent validation in high-stakes medical ML contexts.

The results emphasize that much of the discriminatory power in keystroke timing for PD arises from behavioral components interpretable as a preference for or aversion to slowness under constraint—closely aligned with pathophysiological expectations in PD-related motor bradykinesia. However, the paper also makes clear that although IRL-derived weights carry statistically significant explanatory value beyond conventional features, their effect size is partial and does not account for the majority of severity variance.

Limitations and Future Directions

Interpretation is necessarily tempered by several constraints:

• All results are obtained on a single, clinic-collected dataset without demographic covariates; external replication is required to establish generalizability.
• Some degree of residual feature collinearity persists in the reward formulation (r = 0.946 between speed and consistency in typical contexts).
• Data pooling in action bin boundary selection introduces a mild but disclosed data leakage.
• No claims are made regarding diagnostic performance (binary classification of PD vs. control), focusing instead on severity correlation within the PD cohort.

Future work should focus on (1) external validation in longitudinal and ecologically valid at-home datasets; (2) extension to more granular or demographically controlled cohorts; (3) investigations into adaptive or personalized binning for keystroke actions; and (4) cross-modal integration with other digital or physiological signals.

Conclusion

This study establishes a new paradigm for interpretable digital biomarker discovery in fine-motor disorders, explicitly modeling keystroke timing through the lens of IRL. By formalizing latent behavioral preferences, it provides robust, clinically significant markers of PD motor severity beyond what is attainable with summary timing statistics, and sets a methodological precedent for rigorous validation and error transparency in AI for health applications.