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

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

Problem Statement
Parkinson's disease (PD) is typically diagnosed after significant motor neuron loss has occurred, creating a need for passive digital biomarkers to support earlier detection and longitudinal monitoring. While keystroke dynamics (timing of key presses and releases) have been explored as a candidate biomarker, existing literature relies almost exclusively on supervised classification. These approaches extract fixed summary statistics (means, variances, higher moments) from typing signals to discriminate PD patients from healthy controls. While effective at answering whether a user's typing resembles a PD patient's, they fail to explain why or what specific behavioral changes have occurred. Furthermore, reported accuracies in this field vary widely, with a 2022 meta-analysis showing high heterogeneity ($I^2 = 94\%$), suggesting potential overfitting or methodological inconsistencies in prior studies.

Methodology
This paper proposes a shift from supervised classification to Inverse Reinforcement Learning (IRL) to model keystroke dynamics. Instead of training a classifier, the authors treat each keystroke transition as a discrete choice over typing speed and infer the latent reward function that best explains an agent's observed sequential decisions.

• Data Source: The study utilizes the public neuroQWERTY MIT-CSXPD dataset (85 subjects: 42 PD, 43 controls), collected in a clinic setting during free-text typing. The authors explicitly rejected the Tappy dataset in preliminary analysis, where standard statistical features failed to show significant group differences (AUC $\approx$ 0.377), consistent with literature noting Tappy's difficulty due to uncontrolled home-collection protocols.
• Feature Engineering: The model discretizes flight times into $K=5$ bins based on pooled quintiles. The reward function $R(s, a)$ is defined as a weighted sum of feature functions $\phi(s, a)$:
  1. Speed: $\phi_{speed}(a) = -v_a$ (preference for faster typing).
  2. Consistency: $\phi_{consistency}(t, a) = -|v_a - c_t|$, where $c_t$ is a rolling mean context.
  3. Hand-Alternation: $\phi_{hand}(t, a) = -\mathbb{I}[\text{cross-hand}] \cdot v_a$.
• Optimization: Subject-specific weights $w$ are recovered via Maximum-Entropy IRL, maximizing the log-likelihood of observed bin choices against unchosen alternatives using L-BFGS-B.
• Identifiability Correction: An initial four-parameter model (separating "effort" and "smoothness") failed due to perfect feature collinearity ($r=1.000$) between the two terms in typical contexts. The authors corrected this by merging them into a single "consistency" term, resulting in a stable three-parameter model.
• Validation Rigor: The pipeline underwent an adversarial code review, identifying and fixing two implementation bugs: (1) failure to reset rolling windows at session boundaries for multi-session subjects, and (2) data leakage where the rolling mean included the current observation. Both fixes resulted in negligible changes to the primary correlation coefficient.

Key Results
The primary finding is that the recovered speed-preference weight ($w_{speed}$) correlates robustly with clinical PD severity (UPDRS-III score):

• Correlation: $r = -0.607$ ($p < 0.001$, $n=42$). A more negative weight (indicating a stronger implicit reward for typing fast, all else equal) is associated with greater motor impairment.
• Incremental Value: While raw typing speed alone explains 19.4% of UPDRS variance ($R^2 = 0.194$), adding the IRL-recovered speed preference increases this to 33.8% ($R^2 = 0.338$, $p = 0.006$).
• Robustness: The result holds across:
  • Two independent sub-cohorts (CS1 and CS2).
  • Multiple sensitivity configurations (varying bin counts and window sizes).
  • Bootstrap confidence intervals (95% CI: $[-0.767, -0.369]$).
  • Permutation tests ($p < 0.0001$).
• Negative Results: The recovered weights for "consistency" and "hand-alternation" did not survive confound checks (specifically controlling for raw typing speed) and are reported as non-significant.

Significance and Claims
The authors position this work as the first application of IRL to keystroke dynamics. The paper's significance is framed not merely by the correlation coefficient, but by the validation process itself. The authors argue that the field suffers from unexplained heterogeneity and that rigorous validation—including the identification of identifiability failures, the correction of implementation bugs, and the reporting of negative results—is a critical contribution.

The study demonstrates that IRL can recover an interpretable, per-subject behavioral preference (speed) that correlates with clinical severity and provides explanatory power beyond simple summary statistics. However, the authors remain modest, acknowledging that the model is not perfectly identified due to residual collinearity between speed and consistency terms, and that external replication on independent cohorts is the necessary next step.