Drug development efforts in Parkinson’s disease have accelerated and include promising therapeutic modalities aimed at proximal disease mechanisms1. A major need in the development of disease modifying therapies are tools to sensitively measure the behavioral response of a therapeutic, and increase the efficiency of clinical trials2. This challenge is amplified in Parkinson’s disease where there is heterogeneity across the population and an individual’s signs and symptoms can fluctuate within a day and across days. Task-based motor assessments, such as the MDS-UPDRS3 Part III commonly used in trials, sample behavior infrequently and over a short duration and cannot account for day-to-day variance to reliably determine an individual’s trajectory over time and its change in response to a therapy. Furthermore, clinical rating scales are not designed to measure subtle behavioral changes in the prodromal stages of disease, which is an ideal point of intervention for disease modifying therapies4.

Digital motor assessments performed frequently at home can partly account for day-to-day variability5,6. Potential limitations of task-based approaches include learning effects, reliance on the individual’s ability to perform the task and repeat it consistently over time, and the task’s relevance to everyday motor function. Sensors worn continuously5,6 or positioned in the home7 to capture natural behaviors throughout the day have the potential to overcome some of these limitations and produce sensitive and reliable measures by aggregating over many samples and over fluctuations in symptoms8. It was previously shown that submovement-based measures derived from accelerometers worn continuously at home are reliable, sensitively capture disease progression over time, and reflect motor function in other neurodegenerative diseases9,10,11,12. Submovements (SM) are movement building blocks characterized by bell-shaped velocity-time curves that begin and end with zero velocity13. It has been shown that individuals with PD, while performing pointing movements in a laboratory setting, have velocity profiles with lower peak velocity, a short acceleration phase, and less smoothness14, and submovements that are smaller, slower, and more frequent15. We hypothesized that submovements during natural behavior become progressively slower and smaller in Parkinson’s disease, reflecting hallmark clinical features of bradykinesia and hypokinesia.

There were 161 individuals in the Manifest-PD group, 106 individuals in the Prodromal-PD group, and 269 individuals in the Combined-PD group (including two individuals who did not fit in Manifest or Prodromal groups, see Methods). The Combined-PD group had a median age of 67 years, a median of 44 sessions spanning 1.3 years, and a median UPDRS-III score of 11 (see Supplementary Table 1). The Control group included 26 individuals with a median age of 65. Longitudinal analysis was conducted on the first one year of data from a subset 76 individuals from the Combined-PD group (49 from the Prodromal-PD group, see Methods) with a median age of 68 and a median UPDRS-III score of 2 (range 0–52) with 21 individuals on dopaminergic therapy. The longitudinal control group included 10 individuals with a median age of 68.

Individuals in the Manifest-PD group had smaller and less variable SMs with lower peak velocity and lower peak acceleration compared to the Control group (Table 1). Individuals in the Prodromal-PD group also had smaller and slower SMs with less variability, but with smaller effect sizes (Table 1). The two composite models showed large differences between Manifest-PD and Control groups (|e.s.| = 2.05 for the UPDRS model, |e.s.| = 1.45 for the pairwise model). The pairwise model showed the largest difference between Prodromal-PD and Control groups across all individual and composite measures (|e.s.| = 0.63, Table 1). For the pairwise model, mean SM distance (long duration, PC1 direction) and mean SM peak velocity (long duration, PC2 direction) were the two most salient features. For the UPDRS model, mean activity intensity (AI) of activity bouts was the most salient feature. AI is based on the variance in acceleration in each of the three axes of motion and has been shown to correlate with energy expenditure and be able to classify different levels of physical activity, such as sedentary and light intensity activity (e.g., doing laundry) versus moderate and vigorous activity (e.g., walking)16.

Table 1 Cross-sectional and longitudinal properties of wrist sensor features and models
Full size table

All individual SM measures were significantly correlated with UPDRS-III score and UPDRS-II score with medium effect size, and with the valence of the relationship matching the PD versus Control group comparisons: SMs became smaller, slower, and less variable with increasing disease severity (Table 1). The UPDRS and pairwise composite models had significant correlations with UPDRS-III (|r | = 0.71 and |r | = 0.42) and UPDRS-II (|r | = 0.57 and |r | = 0.36) scores. The vast majority of sensor-based measures demonstrated stronger correlations with UPDRS-II fine motor subscore compared with gross motor subscore (Table 1). The only statistically significant exception was the pairwise model where correlation with gross motor subscore was slightly stronger than fine motor subscore (95% CI: |r | = 0.33 to 0.37 versus |r | = 0.27 to 0.32). The majority of individual SM measures had high within-week reliability with a median ICC of 0.88 (Table 1). The UPDRS and pairwise composite models had ICCs of 0.91 and 0.77, respectively.

One-year longitudinal analysis of the Combined PD group (N = 76) demonstrated that all SM measures had a negative mean slope, and all but one was significantly different from zero (Table 1). The UPDRS model did not have statistically significant change over time. The pairwise model had the steepest mean slope (−0.32 SD/year), largest mean-to-SD ratio (−0.53), and smallest trial size estimate (414) across all measures. For the control group (N = 10), the pairwise model did not show significant change over time (Supplementary Table 2). The pairwise model progressed fastest for individuals in the Prodromal-PD group (N = 49), with sample size estimate of 223 (see Fig. 1), whereas the Manifest-PD group (N = 27) progressed slower and did not reach significance (Supplementary Table 3). The pairwise model was significantly more sensitive than the UPDRS-III clinical score in capturing disease progression in the Prodromal-PD group (mean slope −0.36 vs. 0.03 SD/year, p < 0.0005) and was no different in capturing disease progression in the Manifest-PD group. Males progressed faster than females for both the pairwise model and the UPDRS-III clinical score (Supplementary Table 3).

Fig. 1: Longitudinal data from the Prodromal-PD group.
figure 1

A Progression of the sensor-based pairwise composite score over a 1-year period in the Prodromal-PD and Control groups. Decreasing values indicate disease progression. B Progression of MDS-UPDRS Part III in the two groups. Increasing values indicate disease progression. C An example trajectory from one individual with prodromal PD. Top panel shows the pairwise composite score trajectory. Bottom panel shows the MDS-UPDRS Part III trajectory. Horizontal dotted lines indicate the baseline value and vertical dotted lines on the bottom panel indicate the period aligned with data on the top panel.

Full size image

In summary, movement measures derived from a single wrist sensor worn continuously at home were different in individuals with PD and prodromal PD compared to controls, correlated with UPDRS clinical scores, and captured disease progression over time. A consistent pattern was seen in individuals with PD: submovements were smaller, slower, and less variable. This pattern progressed over time and with increasing disease severity. The pairwise composite model score showed highest sensitivity for measuring progression in the Prodromal-PD group (slope mean-to-SD ratio: −0.72, sample size estimate: 223). The pairwise model, optimized to detect within-subject changes over time, also separated the Prodromal-PD group from the Control group with the largest effect size.

The pairwise model most strongly distinguished Prodromal-PD from Control groups and most sensitively captured disease progression, however the model only had good within-week reliability (ICC = 0.77). This contrasts with prior work in ALS and adult ataxias where the pairwise models trained on wrist data had ICCs ≥ 0.911,12. The lower within-week reliability of the PD-trained pairwise model may reflect true Parkinson’s-related phenotypic variability. When PD model weights were applied to wrist sensor data collected in a separate study in adults with ataxias10,12 the ICC was 0.87 for both the pairwise model and the UPDRS model, supporting that the lower ICC value of the pairwise model in this PD dataset arose from PD-specific characteristics. The pairwise model’s sensitivity for capturing early PD movement patterns and progression over time further indicate the model’s ability to measure small changes in movement. Similarly, the lower correlation between the pairwise model and the UPDRS-III clinical score, compared to the UPDRS model, likely reflects its training focus on detecting small within-subject changes rather than between-subject differences. Both the pairwise model and UPDRS-III notably did not significantly capture disease progression in the Manifest-PD group. This may be due in part to the small sample size (N = 27) and heterogeneity of this group, the potential inability to account for all medication changes, and the possibility that the pairwise model learned more about prodromal PD trajectories due to the larger amount of longitudinal data available for prodromal PD (N = 49).

An ideal digital measure of motor behavior is sensitive for detecting early disease signs and measuring change, clinically interpretable, and reflective of functionally relevant aspects of behavior. This work highlights that these objectives can be conflicting. The pairwise model was the most sensitive for detecting early signs and measuring disease progression but correlated less strongly with UPDRS-II and UPDRS-III than the UPDRS model, which did not have sensitivity for disease progression and had weak differences between Prodromal-PD and Controls. These two composite measures were composed of interpretable features– the most salient features for the pairwise model were SM distance and SM peak velocity– but the aggregation can reduce overall interpretability. Single measures are more interpretable, however as shown in this study, single measures are unable to match the performance of composite measures. These tradeoffs are important to consider when selecting digital measures for a particular use case.

For phase I and II trials, trials involving early-stage populations, and trials in rare diseases with limited sample sizes, a sensitive wrist sensor composite measure such as the pairwise model is a promising approach to detect small behavioral responses that could be missed by clinical scales, patient self-report, or single digital measures. Future work would benefit from including an ankle sensor worn at home, as it may yield sensitive measures of gross motor function that closely align with patient-reported function17 and complement upper-limb metrics. Additionally, incorporating emerging staging systems that integrate biological and clinical signs18 could enable a more precise definition of PD stages and clarify the utility of sensor-based outcomes within each stage. Future studies should also consider advanced time-series analysis approaches, particularly deep-learning methods capable of identifying novel and informative temporal patterns from wearable sensor data.

There are some limitations to this study. There were a relatively small number of participants (especially controls) for longitudinal analysis. Different time periods were used for estimating UPDRS-III slopes and sensor-based slopes. Given the relatively short (1-year) period of longitudinal analysis and limited sampling frequency of the UPDRS, we used a linear model to estimate the rate of progression and compare the rate of change in sensor-based measures with the UPDRS. However, future studies that collect frequent wearable sensor data over a longer duration could better characterize non-linear trajectories of disease progression in PD.

Methods

Recruitment and consent

Data used in the preparation of this article was obtained on 2024-09-18 from the Parkinson’s Progression Markers Initiative (PPMI) database (www.ppmi-info.org/access-dataspecimens/download-data), RRID:SCR_006431. PPMI is a prospective, longitudinal, observational, international multicenter study that aims to identify biomarkers for the progression of PD19. For up-to-date information on the study, visit www.ppmi-info.org.

PPMI – a public-private partnership – is funded by the Michael J. Fox Foundation for Parkinson’s Research, and funding partners; including 4D Pharma, Abbvie, AcureX, Allergan, Amathus Therapeutics, Aligning Science Across Parkinson’s, AskBio, Avid Radiopharmaceuticals, BIAL, BioArctic, Biogen, Biohaven, BioLegend, BlueRock Therapeutics, Bristol-Myers Squibb, Calico Labs, Capsida Biotherapeutics, Celgene, Cerevel Therapeutics, Coave Therapeutics, DaCapo Brainscience, Denali, Edmond J. Safra Foundation, Eli Lilly, Gain Therapeutics, GE HealthCare, Genentech, GSK, Golub Capital, Handl Therapeutics, Insitro, Jazz Pharmaceuticals, Johnson & Johnson Innovative Medicine, Lundbeck, Merck, Meso Scale Discovery, Mission Therapeutics, Neurocrine Biosciences, Neuron23, Neuropore, Pfizer, Piramal, Prevail Therapeutics, Roche, Sanofi, Servier, Sun Pharma Advanced Research Company, Takeda, Teva, UCB, Vanqua Bio, Verily, Voyager Therapeutics, the Weston Family Foundation and Yumanity Therapeutics.

This research study was approved by the institutional review board at each PPMI site and participants provided written informed consent.

Participants

Individuals in the PPMI study with 1) a diagnosis of PD, Prodromal-PD, or healthy control (based on PPMI cohort and most recent study visit, see https://www.ppmi-info.org/study-design/study-cohorts), 2) who participated in the Verily Study Watch substudy, and 3) had at least one MDS-UPDRS performed within 90 days of wearable sensor data collection were identified. The prodromal PD cohort in PPMI includes individuals who are at risk of PD based on clinical features (e.g., REM sleep behavior disorder, hyposmia), genetic variants, and positive dopamine transporter (DAT) SPECT (see https://www.ppmi-info.org/study-design/study-cohorts).

Detailed information about inclusion criteria, demographics, and study design for the overall PPMI study can be obtained from the PPMI website. For the Verily Study Watch substudy, the inclusion criteria were active participation in PPMI at a United States site. History of nickel allergy, sensitivity to metal jewelry, pregnancy, and breastfeeding were exclusion criteria.

Healthy controls with an MDS-UPDRS Part III total score (“UPDRS-III score”) > 4 were excluded. Most individuals with a diagnosis of PD were assigned to the Manifest-PD group, except for one individual with a UPDRS-III score <= 4. Most individuals with a Prodromal-PD diagnosis were assigned to the Prodromal-PD group, with the exception of 16 individuals with a UPDRS-III score > 4 (reassigned to the Manifest-PD group), and one individual with a Prodromal-PD diagnosis on dopaminergic therapy. The Manifest-PD and Prodromal-PD groups were used for comparisons with the Control group. The Combined-PD group was used for reliability analysis of the sensor-based measures and to evaluate relationships with UPDRS scores.

Longitudinal analysis was conducted in individuals who wore the sensor an average of at least 3 weeks per month for one year, during which there were no documented changes in levodopa equivalent daily dosage (LEDD). The first year of data from each participant was used in longitudinal analysis. 76 individuals in the Combined-PD group and 10 individuals in the Control group met this criterion.

Wearable sensor data collection and processing

We analyzed wearable sensor data collected in a Parkinson’s Progression Markers Initiative (PPMI)19 substudy conducted between 2018 and 2021. In the substudy, single wrist sensor (Verily Study Watch20) capturing 100 Hz triaxial accelerometer data was worn continuously at home over a period of up to 2.5 years. Participants were asked to wear the device on their preferred wrist and then to wear the device on that same wrist throughout the study. They were asked to wear the device up to 23 h per day, placing the watch in the Study Hub for about one hour each day to charge and transmit data to Verily.

Using a fully automated analysis pipeline without manual quality control steps, accelerometer data were chunked into one-week periods and these “sessions” were included in analysis if the sensor was worn for at least 12 h per day for at least 4 out of 7 days. As in prior work11 where the size of the dataset precluded manual partitioning of the data into day and night segments, segments were automatically partitioned to include data collected between 7:21 am and 11:27 pm21, while accounting for each individual’s time zone. Data analysis focused on the daytime segments. Gravity and high frequency noise were removed using a sixth order Butterworth filter with cutoff frequencies of 0.1 and 20 Hz9,10,22,23. Daytime accelerometer data within each session was processed and analyzed as previously described9,10,11 to produce 85 movement measures. These included total power in the 0.1–5 Hz frequency range and features based on the distribution of activity intensity computed in 1 s time bins. Features were also extracted from “activity bouts” and from submovements. Supplementary Table 4 provides a description of the 85 features extracted from wrist sensor data. Analysis of individual measures was conducted on the 24 submovement kinematic features related to SM distance, velocity, and acceleration.

Additionally, two composite measures were generated by training two different machine learning models. The first approach utilized a lasso regression model, which was trained to estimate MDS-UPDRS Part III total based on the 85 movement features (UPDRS model). The second model did not rely on clinical information and was trained on longitudinal sensor data to identify patterns of progression in Parkinson’s disease by learning within-subject changes across two time points (pairwise model11). The pairwise model was trained as previously described11. Briefly, for each individual, there were \(n* (n-1)/2\) possible pairwise comparisons, where \(n\) represents the number of one-week periods or sessions for that individual. The model takes two 85-dimensional samples (S1 and S2) from a single individual as input, representing that individual’s movement measures at two different points in time (tm and tn). The element-wise difference between the two vectors is computed (S1-S2), representing the direction of change of movement measures. This difference vector was the input to a binary classifier (logistic regression) which learned to predict whether the direction of change reflected disease progression (S2 was temporally after S1) or reflected disease improvement (S2 was before S1). The learned logistic regression model parameters (representing the direction of disease progression) were then applied as linear weights to the original movement measures to generate a score for each session that reflected how far in the direction of disease progression the individual had traveled at that moment in time.

Five-fold cross-validation was used to train and evaluate both approaches. All sessions for an individual were either entirely in the train set or test set for a given fold. For both models, each feature was z-score transformed prior to model training such that feature value ranges and model weights were comparable. To understand which individual features were the most salient in the two models, we identified features that were in the top 10 (out of 85) in feature importance across all 5 cross-validation folds. Pearson correlation coefficient was used to assess convergent validity, with each model compared with MDS-UPDRS Part II and Part III.

Clinical data

MDS-UPDRS assessments were obtained from the PPMI study and were associated with week-long wearable sensor sessions that were collected within 3 months of the assessment. Some individuals had MDS-UPDRS Part III scores completed in both the on and off state. For cross-sectional analyses comparing sensor-based measures with Part III scores, the average of on and off state scores were used (when both were available) to have a clinical measure that reflects some of the day-to-day variability in movement. However, cross-sectional correlations between Part III scores and sensor measures were similar if only on states or only off states were used. MDS-UPDRS Part II total scores as well as fine motor and gross motor subscores were used for cross-sectional correlation analysis. The Part II fine motor subscore was constructed by combining eating, dressing, and handwriting questions. The gross motor subscore was constructed by combining rising, walking, and freezing questions. For longitudinal analysis, only on-state data was used to compute MDS-UPDRS Part III slopes. This was selected to have as consistent scores as possible over time and because there were more on-state assessments performed than off-state assessments. Due to the relatively few MDS-UPDRS assessments available to compute slopes during the wearable sensor substudy, the time span for computing slopes was extended by 1.5 years before and after the sensor study period for each individual.

Statistical analyses

Statistical analyses were completed in MATLAB version R2022a (Mathworks, Natick, MA) and paralleled the statistical analyses conducted in a recent study in ALS11. The non-parametric Mann-Whitney U-test was used to determine individual feature differences between disease and control groups and Cohen’s d was used to quantify effect size. The Benjamini-Hochberg method was used to adjust for multiple comparisons and corrected p-values are reported24. Corrected p-values < 0.05 were considered significant. Single measure intraclass correlation coefficients (ICCs) were used to determine the test-retest reliability of features and composite scores. To evaluate reliability of sensor-based features, sessions with at least 6 days of data were used and features were computed from data recorded in first half of the days in the session (e.g., days 1–3) and the second half of the days in the session (e.g., days 4–6), separately, and ICCs were computed using a 2-way mixed effects model25. Pearson correlation coefficients and p-values were used to evaluate the relationship between sensor-based features and MDS-UPDRS Part II and Part III scores. As above, the Benjamini-Hochberg method was used to adjust for multiple comparisons.

For longitudinal analysis, each measure was first converted to z-scores allowing rate of change over time to be expressed in standard deviations per year (SD/year), thus supporting direct comparison of different measures with different value ranges. Each participant’s progression rate for a measure was determined by fitting a linear regression model to the individual’s longitudinal data for the measure and using the slope to represent progression rate26. The mean and standard deviation of the slope for each measure was computed across all participants with PD. The Wilcoxon signed-rank test was used to assess whether the computed slopes differed from a distribution with a median of zero. The Mann-Whitney U-test was used to compare pairwise model and MDS-UPDRS Part III slopes, however given differences in valence (i.e., negative pairwise slopes and positive UPDRS slopes indicate progression), pairwise model slopes were first multiplied by negative one.

The mean rate of change for the pairwise model and MDS-UPDRS Part III for the Prodromal-PD and Control groups were plotted in Fig. 1. 95% confidence intervals for mean slopes were obtained using the bootstrap method. A single participant from the Prodromal-PD group was randomly selected and their individual-level trajectory was displayed in Fig. 1. Since the pairwise model assessment was conducted at up to weekly intervals and the Part III assessment was approximately every six months, a Gaussian filter with a width of three months was applied to smooth the pairwise model assessment over time for visualization purposes.

For hypothetical clinical trial sample size estimates, we used a one sample model for a continuous outcome27 as described in Rutkove et al.28 with the same model parameters: 90% power to detect a 30% mean change in progression rate, with two-sided P values and a significance level of 0.05. As only the first year of longitudinal wearable sensor data was used for each participant, the sensor-based estimates represent the sample sizes required for a 1-year trial.