Automated screening of Parkinson's Disease using Smooth Pursuit Eye Movements

Introduction

• Oculomotor alterations can occur early in PD, and smooth pursuit has a neurophysiological rationale because its control involves basal-ganglia circuits.An electro-oculographic study reported oculomotor alterations in two-thirds of patients with PD; abnormal saccades during pursuit were reported in 43.8% of patients versus 10.0% of healthy controls.Source: Introduction · p. 1
• Conventional pursuit measures describe specific gaze–target relationships without capturing the signal’s temporal organization and dynamics, motivating complexity and nonlinear analysis.Nonlinear time-series analysis is presented as an indirect way to study system structure and dynamics.Source: Introduction · p. 2
• Evidence for complexity measures in continuous smooth pursuit from participants with PD remains limited, including only one preliminary screening study with moderate discrimination.Other eye-movement screening studies used different tasks, technologies, units of analysis, and validation designs, preventing direct performance comparisons.Source: Introduction · p. 2
• Blink-related missing blocks, baseline wander, and inter-eye coordination can alter gaze sequences, so preprocessing choices may change both interpretation and classifier input.The cited work identifies a need to study the effects of these preprocessing strategies.Source: Introduction · pp. 2–3
• The study proposes extracting complexity and nonlinear measures from preprocessed gaze sequences, selecting a common feature subset, and combining task- and axis-specific SVM outputs with a mixture of experts.The pipeline is intended to support screening and biomarker research, not replace clinical diagnosis.Source: Introduction · p. 3

Participants and clinical assessments

• 52 participants with early PD and 48 healthy controls were recruited; 11 participants with PD and one control were excluded, leaving 41 and 47 for analysis.The results apply to participants able to complete the tasks and provide technically valid recordings.Source: Participants and clinical assessments · p. 3
• Data were collected over two years at two hospitals using the same protocol, equipment, and clinical team..Source: Participants and clinical assessments · p. 3
• Clinical assessments included UPDRS, MoCA, and Beck Depression Inventory measures, and all participants with PD were assessed in their ON-medication state.The evaluation also covered medical history, diagnosis timing, symptoms, and complications.Source: Participants and clinical assessments · p. 3

Recording and smooth pursuit tasks

• A 1000-Hz infrared video eye tracker recorded binocular X and Y gaze positions during 12 fixed-order tasks with horizontal, vertical, or lemniscate target trajectories.Each eye-axis pairing formed a univariate gaze-position sequence, yielding four sequences per recording.Source: Recording and smooth pursuit tasks · p. 3

Methods

• The development set included 70 participants, while 18 participants formed the external test set; a 12-person young-control cohort was used only for descriptive predictions.The development set comprised 33 participants with Parkinson’s disease and 37 healthy controls; the external test set comprised 8 and 10, respectively.Source: Methods · p. 3

Preprocessing

• Blink correction was applied to every gaze sequence, while baseline-wander removal and common-eye-related-movement extraction were varied across four experimental frameworks.Blink gaps were reconstructed with Singular Spectrum Analysis; baseline wander was estimated with an Empirical Wavelet Transform, and ICA extracted the CERM component.Source: Preprocessing · pp. 3–4

Feature extraction

• Thirty-eight measures characterized gaze-sequence temporal and spatial organization for each participant, task, eye, and axis; four delay strategies were compared for embedded measures.Embedding dimension was estimated with False Nearest Neighbours; the delay strategies used a fixed delay or minima or decay points from target- or gaze-position mutual-information curves.Source: Feature extraction · pp. 4–5

Experimental setup

• The four frameworks and four delay strategies yielded 16 experimental settings, allowing preprocessing and delay strategy to be compared separately.The settings were formed by crossing each framework with each delay strategy.Source: Experimental setup · p. 6, Table 4 · p. 7

Preliminary selection of the experimental setting and organisation of the base models

• Preliminary selection compared delay strategies, base-model organizations, and frameworks using AUC, with a shared logistic-regression classifier and participant-level leave-one-out validation.Framework effects on sequence dispersion and local variation were described separately and were not used to select the framework.Source: Preliminary selection of the… · pp. 6–7

Classification pipeline and validation strategy

• Hierarchical selection removed redundant measures before forward selection; SVMs for each task and axis were combined as experts and evaluated through repeated holdouts and an external test.Twenty repeated holdouts tuned the feature subset, SVM, and combiner; the final threshold was set at the equal-error-rate point using validation predictions.Source: Classification pipeline and validation… · pp. 7–8

Selection of the experimental setting and organisation of the base models

• Preliminary analyses selected EF2–DS1 and task-by-axis models; X-axis separation reached AUC 0.68 versus 0.66 pooled, while Y remained near chance.Baseline-wander removal reduced sequence standard deviation by 83.00% and first-difference standard deviation by 3.00%; CERM extraction sometimes reversed effect direction.Source: Selection of the experimental setting… · p. 10, Table 5 · p. 11

Feature selection

• HFS reduced 38 measures to 20, then FFS retained 13; HFS+FFS had the highest mean AUC and G-mean and lowest G-mean variability among feature sets.The retained subset's development G-mean was 0.83, above the mean of 0.78 for 20 random subsets. Removing redundancy shifted contribution from grouped measures (51.22% to 22.21%) toward measures without redundant partners (48.78% to 77.79%).Source: Feature selection · p. 10, Table 6 · p. 11

Base models and features of the selected pipeline

• X-axis base models accounted for 88.58% of total contribution and generally outperformed Y-axis models, including during vertical target movement.The largest contributions came from SPT5, SPT4, SPT9, and SPT1.Source: Base models and features of the… · pp. 10–11, Table 7 · p. 12

Performance of the selected pipeline

• External testing yielded AUC 0.86 (95% confidence interval: 0.65–1.00), sensitivity 0.75, specificity 0.80, and G-mean 0.77.Six of 8 participants with PD and 8 of 10 HCs were correctly classified.Source: Performance of the selected pipeline · p. 10, Table 6 · p. 11, Figure 5 · p. 11

Base models and features of the selected pipeline

• Eight measures contributed 75.34% overall; ordinal and permutation entropy contributed 48.45%, with left-eye HE higher in HCs and wPE higher in PD.Left-eye HE was the top individual feature; higher values were associated with HCs in 14 of 15 active base models.Source: Base models and features of the… · pp. 11–12, Table 8 · p. 12

Discussion and Conclusions

• 0.86 AUC was obtained by the selected pipeline on the external test set.The test set contained 18 participants, and the estimate had a wide confidence interval.Source: Discussion and Conclusions · pp. 12–13
• Short delay vectors spanning 50–75 ms had higher AUC than strategies spanning hundreds of milliseconds to seconds, while baseline-wander removal reduced global dispersion and largely preserved local variation.The 50–75 ms span may combine parts of a catch-up saccade with surrounding pursuit; the source of discriminative fluctuations could not be determined.Source: Discussion and Conclusions · p. 12
• CERM extraction coincided with lower discrimination in EF3 and EF4, but these associations do not establish that the transformations caused the reduction.The observed SampEn reversal is consistent with some irregular local fluctuations being specific to each eye and removed with the non-shared component.Source: Discussion and Conclusions · pp. 12–13
• Feature selection removed correlated-measure redundancy without reducing development performance, and the selected subset outperformed same-sized random subsets.This made feature contributions more interpretable, although correlated features could otherwise accumulate or split contribution.Source: Discussion and Conclusions · p. 13
• Lower left-eye HE and higher wPE, DiffEn, and KFD described less persistent, more irregular gaze sequences in PD.The measures are consistent with shorter periods of stable pursuit or frequent small corrections, but do not identify the events involved.Source: Discussion and Conclusions · p. 13
• The X axis was the most informative source under this protocol, including during vertical pursuit, but this does not imply preferential horizontal impairment.Because the 12 tasks were administered in fixed order, task differences may also reflect adaptation, attention, fatigue, or learning.Source: Discussion and Conclusions · p. 13
• Generalisation is supported only under comparable acquisition conditions; larger independent studies and other systems are needed before transportability or diagnostic specificity is established.The authors position the pipeline as a candidate for screening and biomarker discovery, not a replacement for clinical diagnosis.Source: Discussion and Conclusions · pp. 13–14

• The appendix compares EDel, EDim, and temporal span across delay strategies to explain the final pipeline’s use of DS1.Source: A Embedding delay and dimension… · p. 15

A.1 Illustration of the delay strategies

• DS1 used a 25 ms delay in both participants, whereas DS2–DS4 produced delays of 173–789 ms.The data-driven strategies placed consecutive delay-vector coordinates hundreds of milliseconds apart, while DS1 kept them within a few tens of milliseconds.Source: §A.1 · p. 15
• Figure A.1 shows AMI curves and three-dimensional phase-space reconstructions for HG060 (HC) and HF014 (PD) in SPT11, using the left eye and X axis.The sequences are shown before baseline-wander removal; the figure marks selected delays on the gaze AMI curves and a shared DS2 delay on the target curve.Source: Figure A.1 · p. 15

A.2 Embedding delay of the data-driven strategies

• DS2’s first target-position AMI minimum yielded delays of 558 ms–1.02 s, or 20%–22% of the target period across every task and moving-target axis.This strategy tied the delay to the target period rather than to the gaze-position sequence.Source: §A.2 · p. 16
• DS3 delays had medians of 448–775 ms, while DS4 delays had medians of 115–472 ms.Both strategies produced delays about 5–30 times longer than DS1, varying by participant and thus describing different time scales across participants.Source: §A.2 · p. 16
• Figure A.2 displays DS3 and DS4 EDel distributions by task and axis, with EF1–EF4 shown in colours and DS2 as dotted lines.The horizontal axis uses a logarithmic scale in milliseconds.Source: Figure A.2 · p. 16

A.3 Embedding dimension and temporal span

• Median temporal spans were 50–75 ms for DS1, 2.02–2.24 s for DS2, 2.03–3.20 s for DS3, and 468 ms–1.60 s for DS4.FNN estimated EDim values of 2–4 with DS1 and 2–9 with the other strategies.Source: §A.3 · p. 16
• Figure A.3 shows EDim distributions for DS1–DS4, with colours marking EF1–EF4 and bars indicating the percentage of sequences at each dimension.Source: Figure A.3 · p. 17
• DS1 alone produced delay vectors spanning tens of milliseconds; the other strategies spanned substantial fractions of the target period.At those longer scales, the delay vectors are expected to reflect mainly the target-imposed slow trajectory.Source: §A.3 · p. 17

S1. Preliminary selection of the experimental setting

• Table S1 compares development-set PD-classification AUCs across delay strategies on the X axis, averaged across four experimental frameworks.Source: Table S1 · p. 22
• Separate-eye Y-axis AUCs are omitted from Table S2 because they were close to chance, approximately 0.50.Source: Table S2 · p. 22
• Table S3 reports median paired after-to-before ratios with 95% confidence intervals; ratios below 1 indicate reductions after baseline-wander removal or CERM extraction.Source: Table S3 · p. 22
• Table S4 reports representative Cohen’s d values, where positive values indicate higher PD feature values and negative values higher healthy-control values; bold rows mark direction reversals after CERM extraction.Source: Table S4 · p. 22

S2. Discriminative ability of the individual base models

• Table S5 reports base-model AUCs by smooth pursuit task, target movement, and axis.Source: Table S5 · p. 23
• Each Table S5 cell gives the mean and standard deviation across 20 holdouts, followed by the external-test AUC.Source: Table S5 · p. 23
• Bold Table S5 values identify the three highest development-set AUCs.Source: Table S5 · p. 23