Parkinson’s disease (PD) is a progressive neurological disorder characterized by the cardinal motor symptoms of bradykinesia, rigidity, tremor and postural instability (Jankovic, 2008; Postuma et al., 2015). The Movement Disorder Society-Unified Parkinson’s Disease Rating Scale Part III (MDS-UPDRS III) remains the most widely used method to evaluate motor symptom severity (Goetz et al., 2008). There are several limitations associated with its use; most fundamentally it does not directly assess aspects of health which are of importance to people with Parkinson’s (PwP), such as preservation of their everyday physical mobility (Deane et al., 2014; Delgado-Ortiz et al., 2023; Mammen et al., 2025). Digital mobility outcomes (DMOs), such as walking speed and step count, can be robustly measured in real-world settings with wearables (Kirk et al., 2024b; Micó-Amigo et al., 2023). While wearables have shown promise in complementing existing clinical assessments of PD (Kirk et al., 2025), practical challenges remain for certain devices. For example, some devices consist of multiple sensors, requiring higher sampling frequency, limiting battery life. These may also require re-equipping after accidental removal, and adherence may be more difficult in specific subgroups, such as individuals with cognitive impairment who may forget to consistently wear or charge the device. In addition, a broader challenge across all wearables lies in the contextual interpretation of large, continuously collected datasets

Real-world mobility is assessed across diverse indoor and outdoor environments that place different demands on motor control, such as navigation of tight spaces and obstacles, walking for prolonged periods and negotiating changes in the walking surface from weather (Kim and Brown, 2022; Toda et al., 2020), or terrain (Kowalsky et al., 2021). PwP may choose to walk outdoors primarily during “ON” medication periods (Del Din et al., 2016), while more confined indoor environments may capture freezing of gait (FOG) disturbances (Zhang et al., 2021). Without contextual awareness, it is not clear whether changes in gait parameters are pathological or may instead reflect normal adaptations to environmental conditions (Marcianò et al., 2025). At present the context of real-world DMOs is solely inferred based upon walking bout (WB) thresholds. WBs are defined as walking sequences containing at least two consecutive strides of both feet. The start and end of a WB are determined by a resting period or any other activity (non-walking period) (Kluge et al., 2021). It is speculated that short WBs (<30 s) are more likely to occur indoors and longer WBs (>30 s) over open outdoor spaces. Estimation of context is possible with use of Global Positioning System (GPS) or Global Navigation Satellite System receivers, embedded within specialist devices, smart watches and smart phones, which despite broad applications have yet to be combined with continuous real-world mobility data (Bayat et al., 2020; Breasail et al., 2021; Jankowska et al., 2015). Differentiating between contexts could provide clinicians with richer insight into how mobility limitations manifest in daily life, beyond what can be observed during short, structured in-clinic assessments.

Alongside contextual factors, novel analytical approaches may offer valuable insights into how PwP adapt their walking across real-world environments and medication states (Althoff et al., 2017). DMOs are typically summarized as singular values, which may obscure clinically important patterns within its distribution. Walking speed provides a global measure of mobility, prone to variability from environmental and physiological demands, being influenced by cognitive impairment (Rochester et al., 2014), FOG (Zhang et al., 2021), medication effects (Corrà et al., 2021), and deficits in balance and muscle function (Pääsuke et al., 2004; Pelicioni et al., 2021; Raccagni et al., 2020). This may create a fluctuating pattern in the distribution of walking speed, consisting of multiple modes where slower modes may capture cautious walking or mobility limitations, while faster modes may reflect more optimal performance during “ON” periods or in less challenging environments. A greater number of modes could reflect greater adaptability, or fluctuating pattern influenced by gait-disturbances or medication (Atrsaei et al., 2021; Mancini et al., 2018). Despite its potential, there is currently no standard approach for modeling such patterns, as existing research has focused upon estimating a fixed bimodal distribution of walking speed. Gaussian Mixture Modeling (GMM), an unsupervised machine-learning clustering technique, could objectively estimate the number of walking speed modes within an individual’s distribution potentially providing novel understanding of how PD affects real-world mobility. However, this method has yet to be applied to continuous real-world mobility data, and its clinical utility remains to be established.

The aims of this study were to: (i) investigate the potential of indoor and outdoor contextual location information to enhance understanding of real-world digital mobility outcomes in PwP and control participants and (ii) explore the utility of a novel application of a data-driven machine-learning approach modeling the number of modes within the distribution of walking speed.

Thirty-nine PwP were included from the CiC study, and 19 controls from the TVS. In comparison to controls PwP were significantly taller, younger with a larger proportion of males (62% vs. 47%). Motor scores in the PwP were low (median MDS-UPDRS = 10 points) (Table 1).

The primary aim of this study was to explore the utility of indoor and outdoor location information in its ability to contextualize real-world mobility in PwP and control participants. Despite data loss, this study demonstrated feasibility of smartphones to capture indoor outdoor contextual factors which showed promise in providing novel insight into both walking behavior and functional gait DMOs. Most notably, both cohorts engaged in more walking indoors than outdoors, with shorter WBs occurring predominantly indoors; only the control participants adapted their walking speed differently between locations. Group differences emerged only in DMOs estimated from outdoor walking, where people with PD took significantly fewer steps and accumulated fewer minutes walked per day compared to controls. This study also examined a novel approach to characterizing walking behavior by identifying walking speed modes, which was similar across cohorts with individual variation. Walking speed modes showed moderate associations with motor severity, though further research is needed to establish their potential as a clinically meaningful feature in PD.

This study is the first to combine GPS-derived indoor outdoor contextual data with wearable-derived DMOs. 82% of participants had 7-days of usable contextual data demonstrating feasibility and aligning with published recommendations (Buekers et al., 2024). PwP had a lower proportion of WBs with valid contextual data (PD: 65% vs. controls: 80%). The specific reasons for invalid data were not captured but may relate to common GPS limitations rather than PD. Determining location from GPS is subject to noise, caused by occlusion (Kerr et al., 2011) and reflection that is particularly an issue is urban environments. In addition, the approach relies on mapping data, which suffers from omissions and update delays. Data loss could also be due to practical issues such as smartphones running out of battery if not placed on the Bluetooth charger. In the CiC cohort, Debelle et al. (2023) reported that selected participants found the smartphone a “nuisance,” being “bulky,” or “too big” to carry, finding limited benefit from the device. Future studies may benefit from smaller GPS-enabled devices and from capturing detailed metadata on participant’s living environments (urban vs. rural), which could help clarify sources of data loss, whether it’s due to technical issues or the presence of PD.

Contextual factors showed promise in improving our interpretation of real-world walking behaviors. Both cohorts undertook significantly more walking indoors in keeping with our hypothesis, based upon previous self-reported surveys that have determined humans spend 90% of their time indoors (Diffey, 2011). Additionally, short WBs predominantly took place indoors suggesting that activity here likely represents walking in the context of everyday activities completed in the household or work setting. Additionally, the mean daily WB duration was significantly longer outdoors than indoors for both cohorts, partially supporting our hypothesis that most long WBs would occur outdoors. However, both cohorts performed more WBs > 30 s indoors, reaching statistical significance only for PwP. For WBs > 60 s, only controls showed a non-significant trend toward more outdoor bouts. The overall number of long WBs was low, and the effect sizes at this duration were small (Supplementary Figure 1), suggesting limited statistical power to detect differences. These results may also reflect GPS limitations in distinguishing truly indoor from near-indoor locations (e.g., gardens) or transitions between indoors and outdoors, which GPS cannot reliably capture. Both cohorts undertook a similar amount of indoor walking. However, outdoors PwP undertook significantly less steps, minutes walked per day with a shorter mean and maximum WB duration. This may reflect environmental factors (e.g., controls living in more walkable areas with green spaces) or lifestyle factors such as commuting or greater physical activity. Conversely, people with PD may leave the house only when ‘ON’ medication or accompanied, with fatigue, motor severity, and wearing-off effects limiting outdoor activity. This highlights the complex interplay of factors shaping real-world mobility and the importance of contextual information for more precise clinical insight. As physical activity is neuroprotective in PD (Speelman et al., 2011), maintaining outdoor mobility is especially important.

Contextual factors also showed potential to improve interpretation of gait DMOs. At the gait-level adaptation between locations in walking speed and stride length were only observed for controls. It was expected that walking speed would be slower in more confined and cluttered indoor spaces (Wettstein et al., 2013; Wettstein et al., 2015), requiring a different adaption to outdoors. No location differences in walking speed among PwP may reflect PD-related factors such as impaired muscle contraction, rigidity, and postural instability affecting adaptability (Mirelman et al., 2019), or unmeasured environmental and lifestyle factors, such as more walkable environments that support longer steady-state gait or purposeful walking (e.g., commuting). When comparing cohorts, PwP trended to walk slower outdoors in comparison to controls across all WB categories, though differences were not significant. Perhaps lack of significance was driven by different group numbers and low sample sizes (Supplementary Figure 4). Stride length was significantly shorter during long WBs (>30 s), suggesting that PwP may be unable to generate the same spatial propulsion as OAs during longer, more optimal walking. PwP also showed significantly higher cadence during short outdoor WBs, indicating a more shuffling gait during intermittent outdoor walking. Cadence is not as widely quantified in studies as step or stride duration (Kirk et al., 2025), however this finding is consistent with recent literature (Yarnall et al., 2025). Stride duration tended to be lower in PwP across WBs, but differences were not significant. Overall, contextual factors revealed no strong group differences, likely influenced by unequal group sizes and mild motor severity in PwP. Larger, more evenly matched cohorts are needed to fully explore these differences across contexts.

Additionally, this study explored the utility of an unsupervised machine-learning technique to estimate the number of walking speed modes. The largest proportion of PwP and control participants had three walking speed modes across each duration threshold. Within a single WB, walking speed was expected to demonstrate three distinct phases: acceleration, steady-state, and deceleration, with additional variability to adapt to real-world demands. These findings support this hypothesis and expand upon existing studies which typically fit a GMM of two modes (faster and slower) (Atrsaei et al., 2021; Brodie et al., 2017; Van Ancum et al., 2019). More recently, Baroudi et al. (2022) employed a multi-level modeling approach and identified an optimal number of five walking speed modes across all individuals. However, comparisons are limited as their population consisted of young athletes who were assessed across 14 days. The limited differences in walking speed modes observed here may reflect the relatively mild motor severity of the PwP, as most participants were unlikely to experience pronounced medication fluctuations or FOG that typically drive greater variability (Mirelman et al., 2019; Packer et al., 2025). Furthermore, walking speed modes remained consistent across WB duration thresholds, suggesting stable gait modulation patterns across real-world activities. Although re-sampling participants to include larger step counts (7,500 and 10,000) increased the number of detected modes, this likely reflects the influence of larger sample sizes improving model resolution rather than true behavioral differences (Burnham and Anderson, 2004). Therefore, the relationship between data volume and estimated modes should be interpreted cautiously, as it may not indicate genuine changes in walking adaptability.

Clinical predictors only emerged with a greater number of walking speed modes estimated from walking bouts > 30 s, specifically higher medication dosage (LEDD) and greater motor severity (MDS-UPDRS III) (Supplementary Figures 5, 6). Walking speed in longer bouts reflects optimal steady-state gait, similar to laboratory walking, and is characterized by greater consistency and lower variability (Corrà et al., 2021). In contrast, higher LEDD often indicates longer disease duration and more advanced symptoms, both of which can fluctuate within and between days. Similarly, higher MDS-UPDRS III scores reflect more severe motor impairments, which can cause inconsistency in ability to regulate walking speed and stride length within a single WB. Together, these factors increase variability in walking speed during prolonged bouts, resulting in larger number of walking sped modes. Prolonged bouts provide sufficient temporal resolution to capture these fluctuations, whereas shorter bouts are typically too brief to reveal multiple modes. Previous studies have demonstrated ON–OFF medication differences in walking speed, with one (Corrà et al., 2021) reporting greater impairment in short, rather than long WBs; however, that analysis relied on singular mean and maximum values. Atrsaei et al. (2021) employed a bimodal approach and found correlations between LEDD and walking speed extracted from the highest mode in the distribution, which alongside the current approach, suggests a possible link between walking speed modes and medication. Nevertheless, caution is advised, as participants in the CiC study were in the early stages of motor symptom progression, and some may not have experienced medication-related fluctuations. Replication in a larger, more advanced PD cohort is needed before drawing stronger conclusions.

For limitations, this study had relatively low numbers overall, particularly for control participants which may have affected statistical comparisons. Additionally, PwP were assessed in a singular location, had mild motor severity and limited to H&Y stage 1–3, which limits generalizability of results. Effect sizes of specific DMOs were close to “0” suggesting statistical analysis was underpowered (Supplementary Figures 1–4). Future studies are needed with larger more diverse cohorts. Despite the exciting potential of GPS technology, a comprehensive technical validation considering algorithms, GPS receiver and device wear-location (Kerr et al., 2011; Schipperijn et al., 2014) is required. Recent approaches have adopted machine-learning and deep learning to estimate context based solely upon magnetometer data demonstrating high accuracy in comparison to GPS measurements (Marcianò et al., 2025). Additional DMOs related to turns captured indoors would provide novel insight into how PD affects real-world mobility (Mancini et al., 2018). Walking speed modes was not extended to other DMOs, as our initial aim was to evaluate the feasibility of the approach on a single, representative outcome. Walking speed was selected as it is derived from multiple underlying DMOs and is widely recognized as a global indicator of motor performance (Kirk et al., 2024b; Middleton et al., 2015). Walking speed modes must be replicated in larger and more diverse cohorts of PwP and is complex in its interpretation. Thus, a more comprehensive explanatory analysis is required to establish clinical meaningfulness including information related to falls risk, muscle force (Pelicioni et al., 2021), balance measures (Mancini et al., 2011; Mancini et al., 2012) and psychological factors such as anxiety and fear.

Integrating GPS-derived contextual information with DMOs offers a promising approach to enhance understanding of the factors influencing real-world mobility in PD and to explore how mobility is modulated across indoor and outdoor environments. Further technical validation is needed to refine the collection and analysis of GPS data and to evaluate its applicability in larger, more heterogeneous PD cohorts. Emerging features such as walking speed modes show promise for capturing the fluctuating and adaptive nature of mobility in PD, but their clinical relevance and robustness require confirmation through clinical validation in independent datasets.