Studies selection

The electronic database searches identified 446 records (Fig 1). Ninety-eight duplicates were removed and the remaining 348 articles were screened (229 records excluded). Following full text assessment (95 records excluded) a total of 24 studies were included in the review [31–54].

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 1. PRISMA flow diagram.

https://doi.org/10.1371/journal.pone.0246528.g001

Risk of bias assessment

The appraisal tool yielded six studies with medium and 18 with low risk of bias. Authors reported clear aims and objectives (Q1, 95.8%), study designs (Q2, 95.8%) and selection processes (Q4, 83.3%), however, the sample size was inadequate in 37.5% of the cases (Q3). The outcome variables were appropriate to the aims (Q5, 100%) and measured with the correct instruments (Q6, 100%), while statistics and general methods were reported adequately (Q7, 87.5%; Q8, 79.1%). Results were presented in depth (Q9, 87.5%) and described in the methods (Q10, 87.5%). Discussions and conclusions were justified by the results (Q11, 100%) with no conflicts of interests (Q13, 100%), yet, 37.5% of the authors omitted or did not fully investigate the study’s limitations (Q12). Detailed scores for each level of bias and each individual study are presented in S1 and S2 Tables.

Characteristics of included studies

Among the 24 identified papers, 15 included sessions that were recorded both at home and in the lab during one [33, 35, 38–45, 51], two [31, 32], or three visits [36, 37]; these studies included an initial calibration/validation in a supervised environment for the development of ad-hoc algorithms and then subsequent implementation/testing in an unsupervised setting. Besides the number of testing days and lab sessions, the exact duration of each lab assessment was not always specified, thus hindering the reproducibility of the protocol. In contrast, nine articles described data collection that was exclusively undertaken at home [34, 46–50, 52–54] (Table 2).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 2. Characteristics of the studies.

https://doi.org/10.1371/journal.pone.0246528.t002

Sample sizes ranged from seven [31, 41] to 125 [51] PwP, and from nine [31] to 67 [39] controls in laboratory environments. The same numbers for at-home tests ranged from one [33, 52] to 170 [50] PwP, and from one [33] to 172 [50] controls. In 13 cases, PwP were asked to fill a diary in order to track activities[31, 36, 37, 53], medication intake [31, 32, 34, 35] and symptoms [34, 41, 45–47, 52, 54] (Table 2).

Data were collected by accelerometers alone (in 46.7% of the studies recording in labs and 50% of the studies at-home) [33, 35–37, 41, 44–47, 49–51, 53], in combination with gyroscopes (lab: 33.3%, home: 33.3%) [31–33, 39, 42, 43, 48, 52] or along with gyroscopes and magnetometers (lab: 20%, home: 16.7%) [34, 38, 40, 51, 54]. Authors used off-the-shelf devices such as the AX3 (Axivity, York, UK) [44, 45, 50, 51], DynaPort (McRoberts, The Hague, Netherlands) [33, 39, 42, 43, 48], GT3X (ActiGraph, Pensacola, USA) [53], Mimamori-gait system (Mitsubishi Chemical, Tokyo, Japan) [36, 37], Mobi8 (TMSI International, Oldenzaal, The Netherlands) [33], MOX5 (Maastricht Instruments, Maastricht, The Netherlands) [52], Opal (APDM, Portland, USA) [34, 38, 40, 51, 54] and Parkinson’s Kinetigraph (Global Kinetics Corporation, Melbourne, Australia) [35, 36]. In five studies, volunteers wore prototype sensors [31, 32, 41, 47, 49] (Fig 2). Data collection frequently lasted for a week and ranged from one [31, 36, 37, 41, 46, 47, 49] to 14 days [52] (Fig 3 and Table 2).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 2. Device–Number of studies.

https://doi.org/10.1371/journal.pone.0246528.g002

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 3. Data collection–Number of days.

https://doi.org/10.1371/journal.pone.0246528.g003

Fourteen works investigated gait impairments, eight of which focused on walking [31, 33, 36, 37, 42–44, 51], three on turning [38, 40, 48], two on falls [39, 50] and one on physical activity [53]. Ten articles examined symptoms, side-effects of treatments, and their fluctuations, including two on bradykinesia [32, 35], four on tremor [34, 47, 52, 54], four on dyskinesia [34, 35, 45, 46] and four on the on/off state [41, 45, 46, 49] (Table 2). During gait impairment monitoring, sensors were typically placed at the lower back, in 63.6% and 57.1% of the works taking place in the lab or at-home, respectively [33, 38, 39, 42–44, 48, 50, 51]. Lower back sensors were also combined with IMUs at the top of each foot (lab: 9.1%, home: 14.3%) [38, 40], waist (lab: 18.2%, home: 21.4%) [36, 37, 53], and left shank (lab: 9.1%, home: 7.1%) [31]. To monitor symptoms and their fluctuations, typical sensor positions included the waist (lab: 25%, home: 20%) [41, 49], wrist (lab: 25%, home: 20%) [35, 46, 47], both wrists (lab: 25%, home: 30%) [45, 52, 54] or in a combination of both ankles and the waist (lab: 25%, home: 20%) [32, 34] (Table 2).

Aims, outcome measures, and types of analysis

Fourteen articles investigated gait impairments with the aim of assessing different mobility tendencies and habits in daily life (Tables 2 and 3). Kinematics [31, 33, 36–38, 40, 43, 44], also in combination with frequency measures [33, 39, 42, 51] were computed to study PD and healthy subjects [31, 33, 36–38, 40, 43, 44, 53], or different PD populations such as recently and previously diagnosed patients [31], fallers and non-fallers [39, 50], and subjects with or without freezing-of-gait [42, 48, 50]. In order to extract kinematic and frequency parameters, walk detection algorithms were implemented in seven cases [31, 36, 37, 39, 42, 44, 50, 51], while turning algorithms in three [38, 40, 48].

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 3. Aim, outcome measures, type of analyses, and results of the studies.

https://doi.org/10.1371/journal.pone.0246528.t003

Ten articles studied symptoms and their fluctuations with the intention of detecting bradykinesia [32, 35, 41, 49], tremor [34, 47, 52, 54], dyskinesia [34, 35, 41, 45, 46, 49], and on/off state episodes [41, 45, 46, 49]. Supervised machine learning approaches, such as Artificial Neural Networks (ANN) [45], Fuzzy logic algorithms [35], linear regression [52] and Support Vector Machine (SVM) [32] models were used in this context. One publication used multiple instance learning algorithms [34], namely, the Diverse Density (DD), Expectation Maximization version of Diverse Density (EM-DD), Discriminative variant of the axis-parallel hyper-rectangle (ID-APR), Multiple instance learning k-Nearest Neighbor (MIL-kNN) and Multiple Instance Support Vector Machine (MI-SVM). Finally, four studies used thresholds and analyses of frequency patterns [41, 47, 49, 54]. Walk [41, 49, 54] and activity recognition [32] algorithms were also employed in order to assess symptoms during specific patients’ actions (Tables 2 and 3).

Results of the included studies

Yoneyama et al. (2013/2014) [36, 37] found that the average duration of the gait cycle was longer in PwP (1.16 ± 0.20 s) compared to controls (1.08 ± 0.19 s; p < 0.001). Similarly, Del Din et al. (2016) [44] reported that Parkinsonians walked with slower and shorter steps (step velocity:1.038 ± 0.422 m/s and 1.103 ± 0.411 m/s for PD and controls, respectively; p < 0.001). Moreover, PwP presented less consistent (e.g. step time variability: 0.175 ± 0.156 s for control and 0.181 ± 0.179 for PD; p = 0.07) and asymmetric (e.g. step time asymmetry: 0.093 ± 0.086 for control and 0.098 ± 0.142 for PD; p = 0.116) walking patterns [44], with fluctuations in kinematics and frequency measures compared to healthy subjects [31, 33, 44].

Three studies also investigated turning [38, 40, 48] and confirmed that PwP take shorter turns (2.0 s and 2.2 s for PD and control, respectively; p = 0.001) with smaller angles (92.0° and 95.2° for PD and control, respectively; p = 0.001) [38]. In addition, PwP completed the turning movement at a slower pace than controls (turn mean velocity: 38 ± 5.7°/s and 43.3 ± 4.8°/s, respectively; p = 0.04) and with a greater number of steps (mean number of steps: 3.2 ± 0.8 and 1.7 ± 1.1, respectively; p = 0.04) [40].

One publication investigated the correlation of the monitored overall steps taken (3615/day) and time spent in moderate-to-vigorous-physical-activities (MVPA, 8.1 min/day) with the self-reported activity using the Physical Activity Scale in the Elderly–PASE; there was a moderate correlation for steps (r = 0.56, p = 0.003), but practically no correlation for MVPA (r = -0.003, p = 0.98) [53]. Finally, two works estimated that falls occurred most frequently in PwP with a more variable, less consistent walking pattern [39, 50]; furthermore frequency sensor-derived measures were successfully able to predict future falls even in patients with no previous fall history [39].

When assessing symptoms at-home, Pastorino et al. [32] classified bradykinesia with respect to the UPDRS outcome as measured by clinicians twice per day and achieved an accuracy of 68.3 ± 8.9% with the standard SVM and 74.4 ± 14.9% with a meta-analysis algorithm. Das et al. [34] obtained an accuracy versus symptom diaries of over 90% for both dyskinesia and tremor detection with a multiple instance learning ID-APR classifier. During a recording of ten days, a significant correlation (p < 0.0005) with an r = 0.64 between global median bradykinesia and UPDRS, and a correlation (p < 0.05) with a margin of error of 3.9 (over a range 0–8) between global median dyskinesia and UPDRS was found by Griffiths et al. [35]. Pérez-López et al. [41] developed an algorithm for the on/off state events recognition based on threshold detection and analysis of frequency patterns with a sensitivity of 99.9% and a specificity of 99.9% (compared to the symptom diary). Rodriguez-Molinero et al. [49] has built upon the previous study, increasing the sample size to 23 PwP and achieving an accuracy of 92.20%. Fisher et al. [45] built an ANN classifier that was validated from symptom diaries with a sensitivity ranging from 38% to 52% and specificity from 83% to 93% for the on/off states and for dyskinesia. The method implemented by Ossig et al. [46] had a moderate-to-strong correlation with subject diaries for on/off states and dyskinesia (p-values ranging from 0.404 to 0.658). For the tremor assessment, Battista and Romaniello et al. [47] accomplished a sensitivity of 99.3%, a specificity of 99.6%, and an accuracy of 98.9% as against the tremor diaries; Heijmans et al. [52] reported correlations of up to r = 0.43, when compared to diaries, while McNames et al. [54] detected tremor presence (incorrectly) just 1.1% of the time or less in healthy volunteers.

Characteristics of the studies

For the purpose of gathering large datasets from IMUs recordings lasting from one to 14 days, the most frequently used off-the-shelf devices were the DynaPort, Opal and AX3 (Fig 2), while five works used inertial non-commercial prototypes. The majority of the studies adopted off-the-shelf devices and off-line algorithm solutions. However, a potential implementation of ad-hoc hardware and on-board algorithms could enhance real-time feedbacks and ultimately have a meaningful impact in the life of patients living, for instance, in rural communities and remote areas [35, 46]. In both cases, the direct manipulation of raw data, gathered during the free-living acquisitions, avoids the use of aggregated data (i.e. step, distance) generated by “black box” software of commercial devices.

In the reviewed articles, diaries were completed by PwP or caregivers in order to track daily activities, medication intake, and symptom occurrences. However, the use of self-report for a complex task, such as the self-detection and recording of motor status over a prolonged period, may lead to misinterpretations and errors, particularly in PwP who have impaired cognition [58]. Patients may not always be able to correctly identify their own motor fluctuations and symptoms or they may log motor symptoms in incorrect time slots, or forget to update the records and then complete them many hours later from a recalled general state of function. Reportedly, diaries are not a reliable means of comparison; for example, Erb et al. [58] found that 38% of PwP in this study omitted approximately 25% of entries. However, developing digital versions, with alerts and prompts, may lessen the drawbacks typically associated with traditional paper-based diaries for PwP [59], while the involvement of caregivers trained in the data collection could benefit the quality of the reports.

The number of subjects involved in the data collections is another important aspect with an impact on the results. Sample sizes varied considerably among studies and ranged from one [33, 52] to 170 PwP, [50] from one [33] to 172 [50] controls, and from 1 [52] to 342 [50] volunteers in total (PwP and controls) in unsupervised environments. No pre-study calculation was reported in any of the papers to justify the sample size chosen. As a consequence, the small number of volunteers in certain experimental protocols generated less conclusive and decisive results in terms of statistical power.

Devices’ number and placement were various, depending on the outcomes measured. Concerning impaired locomotion, the center of mass was extensively used in literature to measure movement performance and level of stability [60–62]. Accordingly, to monitor activities such as walking and turning, most of the papers agreed to adopt a single sensor worn close to the waist [36, 37, 53] and lower back [33, 38, 39, 42–44, 48, 50, 51]. Besides, PwP may exhibit asymmetric walk due to the different level of impairment of the lower limbs, characterized by a reduction in walking speed, shuffling steps, and limited foot lifting [3]. Consequently, a sensor attached on the single limb would capture recordings with large variations in gait patterns and it would give just a partial overview of the patient’s status.

Sensor positioning and number is also crucial for the assessment of multiple symptoms on different subjects. In fact, tremor, dyskinesia, bradykinesia, and other PD related motor fluctuations affect upper and lower limbs differently depending on the manifestation and stage of the disease [3]. Thus, a combination of several devices might be more suitable for multiple and concurrent evaluations, however this would compromise the comfort of the system. Yet, given that fewer wearable devices enhance the acceptability, wearability and usability of the system, a sensor on the wrist may offer a good trade-off between applicability and end-user convenience.

Finally, given the potential continuous long-term adoption of wearable systems by PwP, aspects which were neglected in the identified papers, such as a system’s comfort of use, set-up process, instructions for use, support, aesthetics and display, should always be considered to guarantee long-term acceptability and efficacy of the system. For instance, the FDA-approved Parkinson’s Kinetigraph system (PKG), which provides continuous, objective, ambulatory assessments of PD symptoms, has been proved to show high patient acceptability, with 81% of the users reporting satisfactory outcomes [63]. These considerations are crucial if the final purpose is to gather large datasets and if PwP have to interact on a daily basis with the system.

Aim, outcome measures, type of analyses, and results

Kinematic parameters, such as duration of gait cycle, step length, and velocity, were clearly differentiated between the PD and healthy populations. In fact, PwP walked slower and with shorter steps [36, 37, 44]. Less consistent gait patterns with major fluctuations in kinematics and frequency measures were also observed [31, 33, 44]. Findings also underlined differences in turning [38, 40, 48], showing patients taking shorter turns with smaller angles and completing the turning movement slower and with a greater number of steps. Concerning the risk of falling, the relationship between the level of activity and impairments is still a matter of debate among the scientific community. On one side, more active patients could be more susceptible to falls since they are exposed to more unsafe situations, but on the other hand they could be at a lower risk of falling due to a better general health condition. Two reviewed articles estimated that falls occurred significantly more frequently in PwP with a less consistent walking pattern [39, 50], while fallers seemed to have a reduced capability to regulate gait due to a partial loss of postural stability [64]. Inertial wearable device can detect such impaired walking patterns and predict future falls even in patients with no previous fall history [39].

To evaluate tremor at-home, two papers reported an accuracy against the symptom diary higher than the 90% [34, 47]. In particular, Battista and Romaniello et al. [47] presented a promising method based on the spectral analysis of inertial data from a single wrist worn sensor, in conjunction with the detection of specific movement patterns generally related with Parkinsonism. To assess bradykinesia and dyskinesia, Griffiths et al. [35] implemented a fuzzy logic approach using data collected from an accelerometer on the most affected wrist; these algorithms are the core of the PKG, the first FDA-approved device for the continuous assessment of PD symptoms. In addition, regarding dyskinesia, Fisher et al. [45] developed an ANN classifier that was validated from symptom diaries obtaining a promising level of specificity (93%) but still with a low sensitivity level (38%). Finally, to detect on/off episodes, Pérez-López et al. [41] and Rodriguez-Molinero et al. [49] developed an algorithm based on the extraction of gait features from an accelerometer on the waist. The algorithm showed an accuracy of 92.2% when compared to the results of the diaries, however this approach relied upon gait parameters and required patient’s movement; therefore, it might not be suitable for the recognition during the advanced stage of the disease when PwP are mostly inactive.