What Do Longitudinal Voice Measurements Sample? Patient Specificity, Temporal Sampling and Disease-State Reconstruction in Parkinson’s Disease

Abstract

Background: Remote voice monitoring provides detailed long-term observations in Parkinson’s Disease (PD), but frequent sampling does not necessarily mean accurate tracking of disease progression within patients.

Objective: To evaluate patient specificity, temporal sampling and disease-state reconstructability in the Oxford Parkinson’s Disease Telemonitoring dataset as an empirical proof of concept for biomedical sampling-neutrality and reconstruction underdetermination.

Methods: There were 5875 voice samples, which 42 people delivered. As the same subjects were recorded multiple times at the same time points, 2,501 visit-level observations were generated by averaging the voice measures within each subject-time point. These were used for the primary analyses. The analysis included the examination of patient-specific UPDRS trajectories, Intraclass Correlations (ICC), inter- and intra-patient voice associations, intra-patient principal components, subject-grouped cross-validation, forward temporal prediction and random temporal thinning of directly measured voice variables.

Results: The motor and total UPDRS tests depend a lot on the patient with good reliability (ICC=0.927 and 0.943). The slopes for motor UPDRS were positive in 28 out of 42 participants and negative in 14 out of 42 participants. The total UPDRS slopes were positive in 28 out of 42 participants and negative in 14 out of 42 participants. The overall correlation between voice and UPDRS was moderate. The maximum of both worked-out and total correlation achieved |rho|=0.204 and |rho|=0.179 whereas patient-wise correlation was small with a maximum value of |rho|=0.054 and |rho|=0.041. The first within-patient voice component accounted for 66.9% of voice variance; it was almost unrelated to within-patient motor UPDRS (rho=-0.022), total UPDRS (rho=-0.022) and elapsed time (rho=-0.008). Analysis of voice-only ridge models for motor and total UPDRS in unseen-patient grouped validation yield negative R2.

The final 30% of observations were predicted well (R²=0.797 and 0.838) by known-patient baseline models, showing no improvement with voice variables. When applying random thinning at a period of 0.046, the median voice-feature reconstruction correlation drops from 0.192 to 0.046.

Conclusion: Repeated voice measurements contained substantial patient-specific structure but limited evidence of within-patient tracking of interpolated UPDRS progression in this dataset. Dense observations therefore should not be equated with unique disease-state reconstruction.

Keywords: Parkinson’s disease, Digital biomarkers, Voice, Longitudinal monitoring, Sampling, Within-person analysis, Disease-state reconstruction, Telemonitoring

Abbreviations: APQ11: 11-point amplitude perturbation quotient; BioMeTs: Biometric Monitoring Technologies;

Biomedical Sampling-Neutrality: Methodological proposition that sampling density and temporal discretization characterize the observation protocol rather than the disease’s intrinsic temporal granularity; DFA: Detrended Fluctuation Analysis; Disease-State Reconstruction Underdetermination: Methodological proposition that a clinical state may not be uniquely reconstructible from the available measurement domain; GroupKFold: Grouped K-Fold Cross-Validation; HNR: Harmonics-To-Noise Ratio; ICC: Intraclass Correlation Coefficient; IQR: Interquartile Range; MAE: Mean Absolute Error; MAE: Mean absolute error; MDS-UPDRS: Movement Disorder Society-sponsored revision of the Unified Parkinson’s Disease Rating Scale; NHR: Noise-To-Harmonics Ratio; PC1: First Principal Component; PCA: Principal Component Analysis; PD: Parkinson’s Disease; PPE: Pitch Period Entropy; R²: Coefficient of Determination; RMSE: Root Mean Squared Error; RPDE: Recurrence Period Density Entropy; SD: Standard Deviation; UPDRS: Unified Parkinson’s Disease Rating Scale; V3: Verification, Analytical Validation And Clinical Validation; ρ (rho): Spearman Rank Correlation Coefficient

1. Introduction

Parkinson’s Disease (PD) is classified as a progressive neurodegenerative disorder wherein its clinical severity is routinely assessed with the help of structured rating scales like the Unified Parkinson’s Disease Rating Scale and its Movement Disorder Society revision¹. The episodic nature of the in-clinic assessment has given rise to remote digital measurements that can capture the symptoms with greater fidelity and in more naturalistic settings^2,3.

Remote monitoring of voice is especially attractive because dysphonia is common in PD and acoustic measures can be collected noninvasively. According to Tsanas, et al.⁴, the combinations of voice features shown in the Oxford telemonitoring study could estimate the clinician-derived values of UPDRS and help show that speech-based remote monitoring may be feasible. Previous research demonstrated that in a study⁵, nonlinear and conventional dysphonia measures, like pitch period entropy and harmonics-to-noise measures, could discriminate PD-related voice impairment. Over the last years, researchers continue to endorse speech as a promising digital biomarker. Moreover, heterogeneity, validation requirements, as well as longitudinal evidence were emphasized^6,7.

A factor that differentiates individuals whose average severity of disease differs does not necessarily correlate with changes over time within the same individual. Longitudinal methodology clearly categorizes between-person effects and within-person effects since aggregation across individuals can produce associations that are weak, absent or reversed at the individual level⁸.

The present study is a test of two active propositions as reflected by the Oxford Parkinson telemonitoring dataset. The Biomedical Sampling-Neutrality theorem describes an observation protocol consisting of dense or discrete time-stamped measurements; but this does not, by itself, establish the temporal structure or reconstructability of the disease. The phenomenon of Disease-State Reconstruction Underdetermination suggests that a clinical state may be not uniquely reconstructible from an available measurement domain in new patients.

The purpose was not to surpass previously published models for UPDRS prediction. Rather, how strongly do the observed UPDRS levels reflect patient-specificity? What percentage of the voice-UPDRS associations persists between separating out of the within-patient and between-patient information? How well do voice measures generalize to unseen patients as opposed to future measurements of already observed patients? How does temporal thinning affect the reconstruction of the voice state that was directly measured?

2. Materials and Methods

2.1. Dataset and preprocessing

Tsanas et al.⁴ described the Oxford Parkinson’s Disease Telemonitoring dataset used in the analysis. The dataset contains a total of 5875 recordings from 42 subjects in the early stages of PD. Each of the 5875 recordings has a wealth of information. The person’s subject identifier, age, sex and time from recruitment. Also, we have linearly interpolated motor and total UPDRS and a total of 16 acoustic voice measurements. The dataset documentation clearly indicates that motor UPDRS and total UPDRS are clinician scores through linear interpolation. There were no missing values.

Multiple recordings of the same song were often made at the same subject-specific test time. To prevent treating same-time duplicates as separate temporal visits, primary analyses collapsed the data to 2,501 unique subject-time observations by averaging out voice variables within a subject-time duplicate.

2.2. Patient-specific trajectory analysis

The immediate change for each participant was defined as the least-squares slope of the motor and total UPDRS versus test time. That slope was converted to a change per 30 days. We calculated Spearman correlations for each UPDRS outcome in relation to time. We obtained ICCs with random-intercept mixed models.

2.3. Between-person and within-person voice associations

The subject mean was defined as the between-person component for each acoustic variable X and the visit-specific deviation from that mean was defined as the within-person component. Spearman correlation was calculated before and after person-mean centering. We utilized cluster-robust regression models that included age, sex, centered time, a standardized between-person voice component and a standardized within-person component⁵.

2.4. Latent voice-state analysis

The voice variables of 16 people were standardized and underwent PCA. We used the correlations of PC1 with within-person motor UPDRS, total UPDRS and elapsed time to evaluate whether the dominant acoustic state axis operated like a simple progression coordinate.

2.5. Patient-aware predictive validation

Group K Fold cross-validation with seven folds was applied with a generalization to unseen patients whereby all visits of each patient were assigned to that patient’s fold. The analysis of age+sex+time, voice-only and combined predictors found utility of ridge regression and histogram gradient boosting. Independent of each participant’s visits is what forward prediction of known patients is trained on.

2.6. Temporal thinning of directly measured voice data

Re-interpolating the thinned UPDRS data would be circular because the dataset already contains linear interpolated UPDRS outcomes. The voice features used in the sampling experiment were all measured directly. To create a smaller set of voices, we maintained the first and last observation and thinned out visits by 75%, 50%, 25% and 10%. Linear interpolation was used to construct omitted values. Each retention condition was executed 50 times.

2.7. Statistical interpretation

Focus was on the magnitude of the effect, within versus between decomposition, patient level validation and sensitivity patterns. Cluster-robust models give two-sided p values for reference. The analysis employed Python, pandas, SciPy, statsmodels and scikit-learn.

3. Results

3.1. Sample structure

There were 2,501 unique subject-time visits, derived from 5,875 recordings. Study participants made a median of 59 visits (range 39-78) over a median follow-up of 170.5 days (Table 1).

Table 1: Dataset and participant characteristics.

3.2. UPDRS trajectories were strongly patient-specific

The interclass correlation coefficient for motor UPDRS was 0.927 and for total UPDRS it was 0.943. The change in direction varied between the participants. A total of 28 participants had positive motor slopes and another 14 had negative motor slopes. The same number of participants had a total UPDRS score that was either positive or negative (Table 2, Figure 1 and 2).

Table 2: Patient-specific UPDRS trajectory heterogeneity.

Figure 1: Motor UPDRS trajectories across the 42 participants. Each line represents one participant.

Figure 2: Patient-specific 30-day motor and total UPDRS slopes. The wide spread illustrates heterogeneous temporal direction and magnitude.

3.3. Overall voice associations largely disappeared within patients

The strongest connection with a motor assessment of the UPDRS was with Shimmer: APQ11 (rho=0.179), reflecting the association of the visit level. The highest within-patient correlations observed were merely 0.054 (Table 3 and Figure 3).

Table 3: Cluster-robust decomposition of selected voice-UPDRS associations.

Figure 3: Overall, versus within-patient Spearman correlations between selected voice measures and motor UPDRS.

3.4. A dominant latent voice axis did not behave as a disease-time coordinate

PC1 explained 66.9% of within-patient acoustic variance and the first six components explained 95.6%. PC1 was essentially unrelated to centered motor UPDRS (rho=-0.022), centered total UPDRS (rho=-0.022) or centered elapsed time (rho=-0.008).

3.5. Unseen-patient reconstruction and known-patient future prediction differed

Ridge modelling using voice features yielded negative R-squared for both outcomes in unseen-patient grouped validation. After availability of a subject-specific baseline, visiting frequency and PD-status baseline-model predicted the last 30 per cent of visits in PD-subjects quite well (R²=0.797 for motor and 0.838 for total UPDRS). The introduction of centered time led to modest improvements; the incorporation of all 16 features had a minimal effect on the R² value (Table 4 and Figure 4).

Table 4: Patient-aware predictive validation.

Figure 4: Out-of-fold R² for ridge models evaluated in previously unseen patients.

3.6. Temporal thinning reduced voice-state reconstruction

From 75% retention to 10% retention, the median feature-level reconstruction correlation went down from 0.192 to 0.046. The normalized RMSE remained about one SD within person, suggesting that even at quite dense sampling interpolation was not complete (Table 5 and Figure 5).

Table 5: Temporal thinning experiment for directly measured voice features (50 random repetitions).

Figure 5: Mean median feature-level reconstruction correlation under random temporal thinning; error bars show the standard deviation across 50 repetitions.

4. Discussion

Tsanas, et al.,⁴ predicted that the UPDRS would be more or less predictable using an Oxford study. The current analysis tackles a different question: Do those dense repeated measurements act as genuine within-patient progression trackers and does increased sampling uniquely reconstruct disease state?

One principal finding was a distinct separation of patient specificity and within-patient tracking. Stable between-person differences were the main source of UPDRS variance. There were some overall associations with the acoustic features but they were fairly modest and they disappeared after person-mean centering. This illustrates why progression biomarkers should not only be assessed using pooled associations^9,10.

As digital biomarkers evolve from infrequent clinical tests to home-based repeated measurements, this distinction becomes increasingly relevant^2,3,11-14. The V3 framework asserts that clinical validation is contextual since a measure that is technically valid for discrimination may not be so for progression^10,15,16.

The other major finding is disagreement between unseen-patient and known-patient reconstruction. The voice models were not able to generalize to new participants well, whereas the baseline of the patient in the past gives a lot of predictive information about the future observations. This implies that a digital measure's information content is partially dependent on its reference history - the history in which we interpret it^9,17.

The third finding was temporal density. As UPDRS values were already interpolated, the appropriate application of the sampling experiment was conducted to the directly measured voice features. Although interpolation improved due to better retention, correlations were still weak. The improved measurement density can give better access to the observed signal without guaranteeing full reconstruction of the underlying state^7,13,14.

The aforementioned results motivate two methodological precautions. Biomedical Sampling-Neutrality characterizes an observation protocol by sampling density and temporal discretization, not the disease’s intrinsic temporal granularity. The state of the disease is reconstructed based on the finite measurement domain, which may underdetermine if clinical state can be inferred uniquely. This also holds for patient across the board. Both propositions do not provide a new biological law. They are analytical restrictions on our inferences^10,15,16.

The results should not be read as evidence against the speech biomarkers. Recent literature provides evidence for longitudinal and treatment-sensitive speech changes examined with more up-to-date protocols, including natural speech, richer acoustic features and treatment-state contrasts^6,18,19. According to recent systematic reviews, however, there is still a stronger validation gap in the longitudinal progression sensitivity across digital technologies^3,7.

Future studies of digital biomarkers should explicitly report progression performance separately for between-person discrimination, within-person sensitivity to change, unseen-patient generalization and incremental value beyond patient-specific history. If the aim is to deploy the model to new individuals, data splitting at the patient level should be considered essential^9,17,20-22.

5. Limitations

Only 42 people in the cohort made between-person inference and external validity limited. Clinician scored motor and total UPDRS values are obtained through linear interpolation rather than independent clinical examinations at each voice visit. Lack of medication status, duration of disease, change in treatment, language, context and detailed recording from the analytic file. Acoustic variables are closely connected and some have a mathematical relationship. The temporal thinning experiment is about evaluating the reconstructability of observed acoustic features and not the unobserved biological disease process. The negative and unseen patient R-values supersede this data set and modelling design and shouldn’t be generalized to modern multimodal or deep learning speech systems. The proposed conceptual principles are methodological constructs supported by proof-of-concept analysis not stand-alone clinical laws.

6. Conclusion

The longitudinal voice measurements of the Parkinson telemonitoring dataset have a lot of patient-specific structure but little evidence of simple within-patient tracking of interpolated UPDRS progression. When centering within-person measures, cross-sectional associations weakened, while voice-based models did not generalize to new patients. Rather, future prediction was largely determined by patient-specific baseline knowledge. Through denser sampling reconstruction of states improves but is not guaranteed as seen through temporal thinning of acoustic features. What is more appropriate is not how often a signal can be sampled, but rather how much clinically relevant state can be identified from the accessible measurement domain for longitudinal digital biomarkers.

7. References

1. Goetz CG, Tilley BC, Shaftman SR, et al. Movement Disorder Society-sponsored revision of the Unified Parkinson’s Disease Rating Scale (MDS-UPDRS): Scale presentation and clinimetric testing results. Movement Disorders. 2008;23(15): 2129-2170.
2. Espay AJ, Hausdorff JM, Sánchez-Ferro Á, et al. A roadmap for implementation of patient-centered digital outcome measures in Parkinson’s disease obtained using mobile health technologies. Movement Disorders, 2019;34(5): 657-663.
3. Sun YM, Wang ZY, Liang YY, et al. Digital biomarkers for precision diagnosis and monitoring in Parkinson’s disease. npj Digital Medicine. 2024;7: 218.
4. Tsanas A, Little MA, McSharry PE, et al. Accurate telemonitoring of Parkinson’s disease progression by noninvasive speech tests. IEEE Transactions on Biomedical Engineering. 2010;57(4): 884-893.
5. Little MA, McSharry PE, Hunter EJ, et al. Suitability of dysphonia measurements for telemonitoring of Parkinson’s disease. IEEE Transactions on Biomedical Engineering, 2009;56(4): 1015-1022.
6. Rusz J, Krack P, Tripoliti E. From prodromal stages to clinical trials: The promise of digital speech biomarkers in Parkinson’s disease. Neuroscience & Biobehavioral Reviews. 2024;167: 105922.
7. Rábano-Suárez P, Del Campo N, Benatru I, et al. Digital outcomes as biomarkers of disease progression in early Parkinson’s disease: A systematic review. Movement Disorders. 2025;40(2): 184-203.
8. Wang LP, Maxwell SE. On disaggregating between-person and within-person effects with longitudinal data using multilevel models. Psychological Methods. 2015;20(1): 63-83.
9. Chaibub Neto E, Pratap A, Perumal TM, et al. Detecting the impact of subject characteristics on machine learning-based diagnostic applications. npj Digital Medicine. 2019;2: 99.
10. Roussos G, Ruiz Herrero T, Hill DL, et al. Identifying and characterizing sources of variability in digital outcome measures in Parkinson’s disease. npj Digital Medicine. 2022;5: 93.
11. Bot BM, Suver C, Chaibub Neto E, et al. The mPower study, Parkinson disease mobile data collected using ResearchKit. Scientific Data. 2016;3: 160011.
12. Zhan A, Mohan S, Tarolli C, et al. Using smartphones and machine learning to quantify Parkinson disease severity: The Mobile Parkinson Disease Score. JAMA Neurology. 2018;75(7): 876-880.
13. Burq M, Rainaldi E, Ho KC, et al. Virtual exam for Parkinson’s disease enables frequent and reliable remote measurements of motor function. npj Digital Medicine. 2022;5: 65.
14. Czech MD, Badley D, Yang L, et al. Improved measurement of disease progression in people living with early Parkinson’s disease using digital health technologies. Communications Medicine. 2024;4: 49.
15. Coravos A, Khozin S, Mandl KD. Developing and adopting safe and effective digital biomarkers to improve patient outcomes. npj Digital Medicine. 2019;2: 14.
16. Goldsack JC, Coravos A, Bakker JP, et al. Verification, analytical validation and clinical validation (V3): The foundation of determining fit-for-purpose for Biometric Monitoring Technologies (BioMeTs). npj Digital Medicine. 2020;3: 55.
17. Karbalaie A, Abtahi F, Häger CK. Participant-aware model validation for repeated-measures data: Comparative cross-validation study. JMIR AI. 2026;5: 87728.
18. Sousa M, Krýže P, Amstutz D, et al. Digital speech biomarkers can measure acute effects of levodopa in Parkinson’s disease. npj Parkinson’s Disease. 2025;11: 184.
19. Šimek M, Tykalová T, Novotný M, et al. Speech biomarkers from smartphone calls track progression in REM sleep behavior disorder and Parkinson’s disease. Annals of Neurology. 2026;99(4): 935-948.
20. Farajollahi B, Saffarian A, Habibi SAH, et al. Diagnosis of Parkinson’s disease based on voice and speech using machine learning and deep learning: A systematic umbrella review. Journal of Voice. Advance online publication, 2026.
21. FDA-NIH Biomarker Working Group. (2016-2025). BEST (Biomarkers, EndpointS and other Tools) Resource. U.S. Food and Drug Administration and National Institutes of Health.
22. Rovini E, Maremmani C, Cavallo F. How wearable sensors can support Parkinson’s disease diagnosis and treatment: A systematic review. Frontiers in Neuroscience. 2017;11: 555.