Monitoring gait at home with radio waves in Parkinson's disease: A marker of severity, progression, and medication response.

We enrolled 50 individuals, including 34 with PD [age 69.4 (7.6) years, 26% women] and 16 without PD [age 66.4 (13.0) years, 50% women]. Among them, 40% were observed for 8 weeks, whereas the rest were enrolled in an ongoing longitudinal study and were monitored for up to 1 year. Table 1 summarizes the demographics and clinical characteristics of the participants, and fig. S1 provides more details about their MDS-UPDRS and ON-OFF behavior. In total, we collected over 200,000 gait speed measurements. The mean (SD) baseline in-home gait speed among those with PD was 0.70 (0.085) m/s, and that among those without PD was 0.91 (0.075) m/s. Thus, individuals with PD had about 23% lower in-home gait speed than the control cohort ( t = −8.10, P < 0.001). Any measurement incurs some variability that is independent from the disease and is due to noise in the measuring device or measurement condition. For example, the person may walk slower than usual if they are talking on the phone or may walk faster if they are trying to reach a ringing alarm. Further, the radio device may introduce some noise. Thus, it is important to average or aggregate individual gait measurements to obtain a consistent measurement that reflects the health status of the individual. To estimate the proper aggregation window, we computed the test-retest reliability ( 16 ). The results reported in Fig. 2A show that the test-retest reliability of gait speed increases with longer averaging windows, and both the average and confidence interval (CI) cross the 95% threshold ( n = 50) for a window size of only 14 days [intraclass correlation coefficient (ICC), 0.98; 95% CI, 0.97 to 0.99]. We also considered an alternative method for aggregating the measurements by computing the 75th percentile gait speed in each aggregation window. Such aggregation method focuses on high-speed gait measurements, and hence better reflects performance during ON times, defined as periods of improved mobility in response to levodopa. The results are similar; specifically, the 75th percentile gait speed and the associated CI cross the 95% threshold ( n = 50) with a window size of 14 days. Overall, our analysis shows that aggregating gait measurements over 2 weeks or longer leads to robust and repeatable results. Several prior studies have shown that gait speed is associated with important medical outcomes, such as health status, cognitive function, dementia, and mortality ( 17 – 20 ). In this study, we investigated whether gait speed measured passively at home provides a clinically meaningful assessment of PD severity. To do so, we compared in-home gait speed against common measures of PD severity such as the MDS-UPDRS part III subscore and total score. Since the MDS-UPDRS is assessed during ON time, we use the 75th percentile gait speed for this correlation to bias the gait speed toward the ON state. The results in Fig. 3 ( A and B ) show that among PD and control participants, in-home gait speed is strongly correlated with both the MDS-UPDRS total score ( n = 50, r = −0.83, P < 0.001) and the part III subscore ( n = 50, r = −0.86, P < 0.001). Further, a subanalysis focused on the axial symptoms as measured by the sum of items 1 and 9 to 13 in the MDS-UPDRS part III reveals that they, too, are strongly correlated with in-home gait speed ( n = 50, r = −0.83, P < 0.001). We also used a covariate-adjusted linear model to validate that the correlation between in-home gait speed and the MDS-U PDRS, both total and part III, is not affected by confounding factors such as age, gender, or comorbidities ( table S1 ). Given that MDS-UPDRS is the key metric for evaluating disease severity in PD clinical trials, this result demonstrates that in-home passive gait speed measurements provide a clinically meaningful assessment of PD severity. Figure 3C shows that the two common standards for assessing PD severity, namely, the MDS-UPDRS and Hoehn and Yahr (H&Y) stage ( 21 ), correlated strongly with in-home gait speed measurements but had a weak correlation with in-clinic measurements of Timed Up and Go (TUG) ( 22 ) and the Ten Meter Walk Test (TMWT). This result is consistent with past work, which shows weak correlation between in-clinic gait measures and MDS-UPDRS ( 11 , 23 – 25 ). The reason might be because gait speed measured in daily home activities reflects the habitual gait performance ( 26 , 27 ) and is less likely to be confounded by observer or Hawthorne effects ( 28 ). To further understand the relation between in-home and in-clinic gait assessment, we compared the TUG and TMWT against the median (50th percentile), 75th percentile, and 90th percentile in-home gait speed ( table S2 ). We found that the correlation between in-home gait speed and TUG/TMWT increased with the higher gait percentiles, indicating that in-clinic measurements focus on patient capacity, whereas in-home measurements focus on patient performance. Overall, the results provide evidence of the feasibility of in-home gait speed as a marker of PD severity. Next, we investigated whether the decline in gait speed over time differs between participants with and without PD and whether such a decline can capture disease progression among those with PD. This analysis is performed on the participants in the longitudinal study. To gain insight into how gait speed changes over time, we plot in Fig. 4A the change in gait speed over 1 year for two participants randomly selected from the PD and control cohorts who completed 12 months of monitoring. Regression analysis shows that the rate of decline for the participant with PD was −0.030 m/s per year (95% CI, −0.038 to −0.022; P < 0.001), which is faster than the gait decline of the control participant, which was −0.014 m/s per year (95% CI, −0.019 to −0.010; P < 0.001). Simply looking at gait decline in the two participants in Fig. 4A , one cannot judge whether the difference is due to the disease or other differences between the two individuals. Thus, we used linear mixed-effects models to analyze the rate of gait decline among the PD cohort and the control cohort. The models included random effects of slope and intercept for each participant and were adjusted for demographic covariates. There was a significant decline in gait speed over time of −0.026 m/s (95% CI, −0.034 to −0.019; P < 0.001) per year for participants with PD and −0.015 m/s (95% CI, −0.025 to −0.004; P = 0.007) per year for control participants as shown in Fig. 4B . These results show that older people, with and without PD, slow down as they age further ( 19 , 29 – 32 ). However, individuals with PD slow down much faster than those without PD. In other words, their gait speed declines almost twice as fast. To evaluate the statistical significance of the difference in gait speed decline between individuals with and without PD, a group and time interaction term was included in the model. We observed a statistically significant interaction between group and time with a greater rate of decline in the PD group ( P = 0.04). These results confirm previous evidence that changes in gait differ significantly between individuals with and without PD ( 33 ), and demonstrate that such difference can be captured at home using passive and touchless measurements. Another important finding is that PD progression was not captured reliably by changes in MDS-UPDRS at baseline, month 6, and month 12. We applied a similar linear mixed-effects model to the MDS-UPDRS values. The results show that the 1-year progression in MDS-UPDRS did not achieve statistical significance (part III subscore model, P = 0.67; total score model, P = 0.71). This indicates that in contrast to in-home gait speed, MDS-UPDRS part III and total assessments were unable to deliver sufficiently sensitive markers of 1-year disease progression in this population. This result is consistent with past work that reported that the 1-year progression in MDS-UPDRS part III was not statistically significant even with a population of over 200 individuals with PD ( 31 ). Table S7 shows that the lack of statistical significance can persist even in larger populations. Last, to control for the effects caused by changes in symptomatic therapy, we repeated the above analysis only for intervals without any changes in therapy. The resulting 1-year gait decline was −0.028 m/s (95% CI, −0.037 to −0.018; P < 0.001) and was significantly higher than the control group ( P < 0.034). We also applied the same statistical analysis to the MDS-UPDRS total score and part III subscore, for the same participants over the same period. The progression in MDS-UPDRS remained statistically not significant ( P = 0.82 for total score and P = 0.49 for part III subscore). Overall, the analysis above shows that in-home gait speed might be a sensitive marker of PD progression. Within 3 to 5 years of levodopa treatment, about 50% of individuals with PD begin to develop motor fluctuations ( 34 ), characterized by a decline in the duration of benefit of each dose with emergent motor fluctuations, a phenomenon known as “wearing-off” or “ON-OFF” effect ( 3 ). Now, clinicians and researchers may assess such ON-OFF symptoms using the Hauser diary ( 35 ), in which patients self-report the time they take their medications, and their functional status every 30 min as ON or OFF ( 35 ). Filling in the Hauser diary imposes a burden on participants and often results in low reliability due to poor adherence and recall bias ( 36 , 37 ). We examined the ability of our system to capture medication impact on motor fluctuations and ON-OFF symptoms. Toward this goal, we exploited the fact that patients usually have regular medication schedules, and hence, their motor fluctuations repeat around the same time across different days. Thus, for participants who were taking levodopa, we computed the intraday gait speed, which we define as the gait speed of an individual as a function of the hour in the day. We compared the intraday gait speed of each participant with their ON-OFF states through the day, as reported in their Hauser diaries. Figure 5 shows evidence that in-home gait speed can capture medication response and motor fluctuations. In particular, Fig. 5A plots the intraday gait speed of four participants with PD, along with the percentage of ON time, as reported in their Hauser diaries. The top graphs reveal that in-home gait speed oscillates in response to medication administration. One possible explanation for these oscillations is that they reflect the effect of medications in only controlling the symptoms; thus, when the patient takes their medications, their gait speed improves, but within 1 to 2 hours, the medication impact wears off and the gait speed declines again, until the patient takes the next dose. The figure also shows that the degree to which the oscillations are controlled differs across patients. For example, PD1’s motor symptoms are relatively well controlled; this is apparent in PD1’s intraday gait speed, which is fairly stable throughout the day and night, and is confirmed in his Hauser diary, which reports no OFF time (compare that to PD2, PD3, and PD4, who experience a large difference between day-time and night-time gait speed and wide fluctuations in their gait speed throughout the day). Also, note that the fluctuations are aligned with the medication times and the percentage of ON time in their Hauser diaries. In addition, we found that changes in intraday gait speed profiles reflected adjustments in medication plan. Figure 5B depicts the intraday gait speed of a participant with PD before and after his medication adjustment. Before the medication adjustment, the patient’s motor performance fluctuated widely throughout the day, showing clear peaks and valleys. After reducing the dose and increasing the frequency of medication administration, the motor fluctuation was mitigated, and the patient’s gait became more stable as shown in the figure. Another example of how the intraday gait profile captures the impact of medication adjustment is shown in fig. S2 . Next, we show that the variability in the intraday gait speed reflects patients’ assessment of the disease impact on their daily function. In particular, the MDS-UPDRS part IV asks participants to characterize the impact of motor fluctuations on their daily function and social interaction as one of the following: normal, slight, mild, moderate, and severe. We compute the variance in intraday gait speed for each individual with PD. Figure 5C shows a boxplot of the relationship between the variance in intraday gait speed and the disease impact on daily function as reported in the patient’s MDS-UPDRS functional impact of fluctuation subscore. Admittedly, the number of participants is relatively small ( n = 34), and only one of them reports a severe fluctuation impact; however, the results show that variances in intraday gait speed are highly correlated with patient’s perception of disease impact on their daily function ( r = 0.74, P < 0.001, Spearman’s rank-order correlation). To evaluate the system’s ability to detect additional health insights, we compared the system-derived clinical profiles against both general events experienced by the population and serious health events reported by the participants. We found patterns in the derived measurements that are associated with coronavirus disease 2019 (COVID-19) stay-at-home orders and a hospitalization due to atrial fibrillation. For individuals monitored both before and after the COVID-19 lockdown, we compared their time in bed per day and the number of walking events per day. Since the device can localize people from the radio signals reflected off their bodies, time in bed was computed by considering the total period when the person’s location was mapped to their sleeping area. After the COVID-19 lockdown, we observed an increase in the median [interquartile range (IQR)] time in bed from 7.81 (0.87) to 8.18 (1.30) hours per day ( P = 0.028) and an increase in the median (IQR) number of walking episodes from 12.68 (10.46) to 16.74 (18.80) per day ( P < 0.01) ( Fig. 6A ). This behavioral change can be partly explained by participants spending more time at home, and thus, our device captured more walking events. Studies have also indicated increasing sleep hours ( 38 , 39 ), which is consistent with our findings in Fig. 6A . Last, we found that the longitudinal gait speed trajectory was able to detect changes leading to a major medical event. One of the participants was hospitalized because of onset of new symptomatic atrial fibrillation and coronary artery disease. He underwent cardiac catheterization surgery and was discharged home after 11 days. Figure 6B shows the trajectory of his in-home gait speed before and after this hospitalization. The trajectory demonstrates a decline in gait speed starting about 6 days before hospitalization. On the day of hospitalization, his median (IQR) in-home gait speed reached its minimum value of 0.55 (0.14) m/s. After he was discharged from the hospital, his gait speed improved but remained lower than his baseline. Overall, the results in Fig. 6 show that passive gait monitoring might help in managing the overall health of patients with PD and can provide context for some changes in their motor symptoms. It also emphasizes the fact that gait speed is a general metric that is affected by PD and other health issues including cardiac problems.

Over a period of 3 years, we have recruited individuals with and without PD to have the wireless devices installed in their home as part of two observational research studies. The first study enrolled individuals with and without PD at two sites (Boston University, University of Rochester) with an observation period of 8 weeks. The second study is part of the National Institutes of Health–funded Morris K. Udall PD Research Center of Excellence at the University of Rochester, which aims to characterize individuals with and without PD at home using multimodal technologies. Those enrolled in our Udall substudy have the device installed in their home over a period of 2 years. By the time of writing, recruitment and study activities are still ongoing in the longitudinal 2-year study. Hence, in this article, we consider only data from the first year of monitoring. All participants with PD were screened and met the UK PD Society Brain Bank Clinical Diagnostic Criteria; control individuals had no evidence of primary or secondary Parkinsonism. After enrollment and screening, a study coordinator installed the wireless device in the participant’s home and connected the device to the patient’s home Wi-Fi network. Study coordinators also created a basic floor map of the monitored area to allow post hoc assessment of participant position in the home. Enrolled participants underwent clinical assessments at baseline, and in the longitudinal study, participants have visits at 6 months, 12 months, and 2 years after installation. During these visits, a movement disorder specialist conducted PD assessments including the MDS-UPDRS parts I to IV, H&Y stage, and TUG, among others. The longitudinal study was started before and has continued during the COVID-19 pandemic, resulting in a shift in clinical assessments, with some visits being conducted remotely. However, all baseline visits were conducted in person. During remote visits, measures that require in-person assessment, including TUG and TMWT, were not performed. The feasibility of remote performance of the MDS-UPDRS has been reported previously. In addition to the clinical assessments, participants with PD completed a home activity log (Hauser diary) for 14 days after each in-person visit documenting self-reported motor status (ON, ON with dyskinesia, OFF, or asleep) during each 30-min interval in the day. Participants additionally indicated whether they took a dose of dopaminergic therapy during each 30-min interval. The Hauser diary was self-reported on a paper form. All participants were trained on the use of the form during the visit, but we cannot determine that patients reported their motor status in a timely manner. Both the 8-week and the longitudinal studies collected MDS-UPDRS and H&Y stage at baseline for all individuals. In addition, the longitudinal study collected MDS-UPDRS assessments at month 6 and month 12. Also, the longitudinal study collected disease duration, whereas the 8-week study did not collect this information. All study procedures were approved by the Research Subjects Review Board (RSRB) at the University of Rochester (RSRB00001787 and RSRB00072169). Both the Massachusetts Institute of Technology (MIT) Institutional Review Board and the Boston University Charles River Campus Institutional Review Board ceded review to the University of Rochester RSRB. Our wireless device hangs on the wall and continuously emits and detects low-power wireless signals that reflect off nearby individuals, similar to radar ( Fig. 1 ). As individuals move, the reflected signals change, allowing the detection and localization of motion in an approximately 30-foot radius around the device ( 66 ). The movement trajectory is analyzed with respect to time to calculate the participant’s gait speed ( 67 ). This approach has been demonstrated to yield a highly accurate measure of gait speed where the relative error is between 1.9 and 4.2%, and the absolute error is between 0.015 and 0.023 m/s. For reference, we also provide a brief comparison of the accuracy of gait speed measurements using wearable sensors ( table S4 ). The device encrypts the data and uses the home wireless network to upload it to a cloud server hosted by Amazon Web Services, where it is processed to extract additional metrics of interest. We note that the device can monitor participants even in the presence of other individuals in the home. Past work has demonstrated methods that identify people using the radio signals that bounce off their bodies ( 68 – 70 ). In this paper, we leverage the method developed by Hsu et al. ( 67 ). Specifically, the device uses standard signal processing techniques such as Frequency-Modulated Continuous-Wave (FMCW) and antenna arrays to zoom in on radio signals from each voxel in space and separate signals based on spatial locations. To identify signals from the participant from others, participants are asked to wear a coin-size accelerometer for the first 2 weeks of the study. The accelerometer data are then used to train a machine learning classifier that identifies the radio signals that bounce off the participant from signals that bounce off other people and objects in the house. Once the classifier is trained, participants do not need to wear the accelerometer and can be identified solely from radio signals using the aforementioned classifier. As shown by Hsu et al. ( 67 ), this method is accurate and its average identification accuracy across different homes and individuals is 90%. The same radio device was used to detect errors in medication self-administration ( 71 ) and monitor patients with Alzheimer’s disease ( 72 ), endometriosis ( 73 ), COVID-19 ( 74 ), and facioscapulohumeral muscular dystrophy (FSHD) ( 75 ). Also, we have explored the use of the device with individuals with PD in a small pilot study focused on sleep and daily activities ( 76 ). At the time of the writing, the radio device used in this paper is incorporated in a large number of clinical studies in neurological diseases ( 77 , 78 ), immune diseases ( 79 , 80 ), and rare diseases ( 75 , 81 ). Some of these studies involve over 250 devices, and many of them run for several years. Even larger trials that involve 500 to 1000 devices, though will require more resources and planning, we believe are feasible. Operationally, the device is easy to deploy and configure using a phone app. The app allows the participant to have a remote call with a technician or engineer, who guides them through the deployment process. Once deployed, the device does not need input from the participant, who can go about their life with no overhead. The process of collecting data and uploading it to the cloud is fully automated and uses state-of-the-art security and networking protocols. Data are processed using a pipeline of signal processing and machine learning modules to output the desired physiological metrics for the trial (such as gait, sleep, breathing, and scratching) in accordance with the trial’s approved protocol. The resulting physiological signals are encrypted and stored in the cloud so that they can be accessed and analyzed remotely with proper security credentials. The device and cloud do not collect or store personally identifiable information, and all data are stored in a coded form. All statistical analysis was performed with Python version 3.8 (Python Software Foundation) and R version 3.4 (R Foundation). All boxplots are shown with a central mark at the median, bottom, and top edges of the boxes at 25th and 75th percentiles, respectively, and whiskers out to the most extreme points within 1.5 times the IQR. ICC was used to assess test-retest reliability. Nonoverlapping intervals with various sizes were used to average the gait speed measurements from the first 2 months after the baseline visit. We compared aggregate mean values among those with versus without PD using two-tailed independent sample t tests (α = 0.05). We assessed the correlation between in-home gait speed and clinical PD outcome measures [MDS-UPDRS part III subscore and total score ( 7 ), H&Y stage ( 21 ), TUG ( 22 ), and TMWT] at the baseline visit using Pearson correlation. Baseline in-home gait speed was measured from the first 14 days after the baseline visit. We used the 75th percentile in-home gait speed to reflect gait speed during ON time. These clinical PD outcomes were measured when participants were on medication. We have performed statistical analyses to ensure that the correlation between in-home gait speed and the MDS-UPDRS (total and part III scores) is not affected by confounding factors. In addition to demographics (age, gender, ethnicity, race, and education), we consider the following 15 comorbidities: cognitive performance; overall health; cardiac health; hypertension; vascular health; respiratory system health; eye, ear, nose, and throat (EENT) health; upper gastrointestinal (GI) health; lower gastrointestinal health; hepatic health; renal health; other genitourinary (GU) health; musculoskeletal and integumentary health; endocrine and metabolic health; and psychiatric and behavioral health. We score cognitive performance according to the Montreal Cognitive Assessment (MoCA), which was collected for all participants. Other comorbidities are scored according to the Modified Cumulative Illness Rating Scale (MCIRS), which was collected for the participants in the 2-month study and assessed retrospectively based on the medical history and standard guidelines ( 82 ) for the participants in the longitudinal study. We follow the standard covariate-adjusted linear model analysis ( 83 , 84 ). In the core model, we regard MDS-UPDRS total and part III as dependent variables and in-home gait speed as an independent variable with demographic covariates controlled (model A). We then augment this core model by further adjusting for potential comorbidities (models B to P). Table S1 summarizes the results. The first row in the table shows the linear relation between gait speed and the MDS-UPDRS. The other rows account for age, gender, and other demographic factors. The different columns account for the impact of various comorbidities. Note that there are two ways to show that a particular comorbidity is not a confounding factor: either the comorbidity’s relationship to the MDS-UPDRS is not significant ( P > 0.05) or the relationship of the comorbidity to the MDS-UPDRS may be significant but its presence does not change the linear relationship with gait speed. The table shows that demographics and comorbidity variables did not constitute confounding factors. We assessed per-participant changes in gait speed using linear regression. We assessed the longitudinal changes in gait speed for the PD and control cohorts using linear mixed-effect models. To compare the gait decline rate of the PD cohort with the control cohort, we used models that included a group × time interaction term (fixed effect) to determine whether the change in gait speed over time differed significantly. We included a random effect for intercept and slope for all participants. All models included demographic covariates, namely, gender, age, race, ethnicity, and education. Linear mixed-effects models for within-cohort analysis can be specified by the following formula. For the j th observed data point of subject i , let S P D i j represent gait speed and T i j represent time of observation with respect to (w.r.t.) baseline. Let S P D baseline , i represent baseline gait speed of subject i . Let v ∈ covariate . S P D i j - S P D baseline , i = u 0 i + β 1 + u 1 i × T i j + ε i j + Σ β v × v where u 0 ∼ N o r m a l 0 , σ u 0 , u 1 ∼ N o r m a l 0 , σ u 1 , E ∼ N o r m a l ( 0 , σ ) . Notably, we did not include a fixed-effect intercept in the model because we expect no gait speed decline at baseline. For combined PD and control cohort analysis, we included the group × time interaction term in the formula S P D i j - S P D baseline , i = u 0 i + β 1 + u 1 i × T i j + ε i j + β 2 × cohort i j × T i j + Σ β v × v where u 0 ∼ N o r m a l 0 , σ u 0 , u 1 ∼ N o r m a l 0 , σ u 1 , ε ∼ N o r m a l ( 0 , σ ) . Restricted maximum likelihood estimation was used for all models to estimate the parameters. As an exploratory endpoint, we collected from each participant information on events of interest during the monitoring period, including societal events, such as local lockdowns in the setting of the COVID-19 pandemic, and personal medical events, such as hospitalizations and medication adjustment. We assessed the impact of these events on system-derived measures. If an event was identified, we retrospectively reviewed relevant data, such as changes in schedule or time at home in response to lockdown orders, and used descriptive statistics or demonstrative figures to illustrate the changes. As this was an exploratory endpoint with numerous possible events, we did not prespecify what would qualify as an event or which parameters we would assess. In general, events of interest were identified through general knowledge of societal changes, such as lockdown dates in geographic area, recognition of a change in a parameter during routine data monitoring, such as participant out of range of the device for a prolonged period, or by participants’ reporting of an event during a routine study visit, such as hospitalization. If the event was noted during interim data analysis, further clinical information about the event of interest was obtained at the participant’s next scheduled study visit. To analyze the impact of the COVID-19 lockdown, we computed the number of walks per day and the time in bed per day for each participant. One-tailed Wilcoxon signed-rank test was used to assess whether there were significant increases or decreases in the measurements. Prelockdown measurements were computed from 1 January 2020 to 1 March 2020, and post-lockdown measurements were computed from 1 April 2020 to 1 June 2020. Participants whose devices were installed before 1 January 2020 were included in the analysis. To compute the number of walks per day, we used the method described in ( 67 ). To compute the time in bed per day, we used the method described in ( 85 ).

We demonstrated that a radio sensor can facilitate passive, touchless, and clinically meaningful assessment of PD at home. We presented evidence that the device can help in assessing PD disease severity, tracking PD progression, and objectively monitoring PD-specific disease features, including response to medication and fluctuations in motor functions. Further, it offers the ability to zoom in on each individual with PD and observe changes in physical functions due to events of clinical importance (including both short-term events such as medication adjustment and long-term events such as the COVID-19 lockdown), demonstrating the benefits of this technology for personalized care and telehealth. Our study has used passively measured in-home gait speed as the key metric to characterize and monitor PD. The sensitivity of in-home gait speed to PD severity and progression can be partly explained by the fact that gait requires a complex interplay between motor, cognitive, musculoskeletal, and cardiorespiratory systems, each of which can be affected by PD progression ( 40 ). The importance of assessing gait speed in the real-world environment is further demonstrated by our data, which show that comprehensive clinical tests such as the MDS-UPDRS correlate strongly with in-home gait speed but correlate only weakly with in-clinic measures of gait speed. These results are aligned with past observations that MDS-UPDRS total and part III scores are weakly correlated with in-clinic measures of gait speed ( 11 , 23 – 25 ). Our results also show that in-clinic gait measures such as TUG and TMWT correlate better with the faster instances of in-home gait speed, which is consistent with past reports ( 41 ) that in-clinic gait captures patient capacity, whereas in-home gait captures patient performance. This indicates that in-home and in-clinic gait measurements can play a complementary role in PD assessment. The results in this study might have important implications for clinical trials and the development of better and possibly personalized therapies for treating PD. A key challenge for PD clinical trials stems from the fact that the disease evolves slowly, yet MDS-UPDRS is not sufficiently sensitive to small changes in symptoms. As a result, PD clinical trials need to last for multiple years and enroll hundreds of participants before differences in MDS-UPDRS can be reported with sufficient statistical confidence ( 17 , 20 , 42 , 43 ). In contrast, our study shows that the decline in gait speed exhibits strong statistical significance in less than 1 year, even when the sample size is relatively small. This indicates that in-home passive gait measurements have the potential to enable clinical trials with a shorter time span and fewer participants, and ultimately accelerate the development of new therapies. Table S9 provides an example of a hypothetical 1-year clinical trial of a new PD treatment, demonstrating the potential for large reduction in the required sample size. Our approach has also the potential of facilitating recruitment and improving retention in clinical trials. This is because it facilitates virtual trials in the participants’ homes, reduces the need for clinic visits, and frees the patient and their caregiver from reporting and maintaining symptom diaries. The results in this study also have important implications for PD clinical care. First, the ability to evaluate patients’ response to medication is pivotal to PD medication titration ( 44 ). Our study provides important insight into medication response and the relationship between PD motor function status and in-home gait speed. Our approach could allow physicians to adjust the dose and then observe the impact of the adjustment on motor fluctuations. We also presented several cases where the system captured changes in physical function that reflect clinical and social events. These results suggest that in-home free-living characteristics could have broader implications, including being used as a tool to investigate behavioral symptoms, evaluate medication adherence, and predict risk of hospitalization. This technology might also have clinical utility for the assessment of individuals with PD who have traditionally been underserved, including those who live in rural areas and those with difficulty leaving the home due to limited mobility or cognitive impairment ( 45 ). The study has also some limitations. First, the studied population is relatively small. We note, however, that the power calculation ( 46 ) ensures that the number of participants is sufficient for achieving statistical significance and justifies the statistical validity of the results in the paper. Future studies with more participants are important to confirm the wide applicability of the results across diverse PD populations. Second, participants are monitored only at home and only within space covered by the radio device. Third, the evaluation of the system is limited by the precision of standard PD ratings and patients’ self-reported outcomes. In particular, the MDS-UPDRS ratings were acquired through episodic in-clinic assessments and could be affected by patients’ physical status on the day of in-clinic visit. Fourth, gait speed is a general metric not specific to PD and could be affected by other health issues. Our covariate analysis, however, provides evidence that none of the common comorbidities constitutes a confounding factor for our results. Last, PD exhibits several gait impairments that are not studied in this paper. Some impairments such as falls and freezing can potentially be captured by our radio device ( 47 , 48 ) and provide a natural topic for future work. Despite the above limitations, we believe that the study provides important insights and addresses key unmet needs in clinical care and drug development in PD. Our work is aligned with recent interest in leveraging digital technologies for assessing PD. Technologies that use cellphones or wearable devices are widely available and can be used inside and outside the home. However, they require patients to wear and charge devices ( 11 , 41 , 49 – 52 ), and unless the movements are scripted ( 11 , 49 ), their correlation with the MDS-UPDRS is inferior to our method ( 52 ). Technologies that use cameras can obtain more detailed motion information, but they are invasive of patients’ privacy and so far have been tested only in laboratory scenarios ( 53 ). By being passive, unscripted, and contactless, our approach satisfies an unmet need for assessing individuals with PD in their natural living environment for months and years, with minimal overhead. Last, we note that our findings open up new possibilities that could improve clinical practice across other illnesses. The medical literature shows that gait speed is a valuable metric for many diseases and health conditions including cardiovascular diseases, cancer, dementia, frailty, mortality, impaired mood, cognitive decline, ataxia, Huntington’s disease, progressive supranuclear palsy, and multiple system atrophy ( 14 , 17 – 20 , 54 – 65 ). Our study demonstrates the feasibility of continuously monitoring this metric at home in a passive manner. The monitoring device operates in the background, analyzing the reflection of radio signals. It tracks gait speed, its daily fluctuations, and its changes over time and provides this information remotely to the professional caregiver in a timely manner. Thus, without asking patients to leave home or actively measure themselves, doctors can achieve safe and frequent health monitoring and deliver more personalized care. This has important implications, particularly in light of COVID-19 social distancing measures and a growing population of individuals with chronic illnesses.

Parkinson’s disease (PD) affects about 1 to 2% of people aged 65 and older ( 1 ) and is the fastest growing neurological disorder in the world ( 2 ). It is the prototypic degenerative movement disorder, characterized by a combination of slowness, stiffness, tremor, and postural instability, that results in gait dysfunction. To date, there are no drugs that can prevent or halt PD. The most effective medication for improving symptoms is oral levodopa, a precursor of dopamine, which is reduced in the disease. Response to therapy is robust early in the course of PD, but as the disease progresses, individuals with PD develop motor fluctuations, characterized by periods of good medication effect (ON time) and periods of emergent PD symptoms as the medication wears off (OFF time) ( 3 ). Objective assessment of clinical progression and the impact of medications on symptoms is essential for both PD drug development and the delivery of medical care. Such assessment, however, is challenging for two reasons. First, PD specialists are concentrated in medical centers in urban areas, whereas individuals with PD are spread geographically and have problems traveling to PD centers due to old age, limited mobility, impaired cognition, and decreased driving ability ( 4 ). As a result, many patients with PD (40% of patients with PD in Medicare data) are not seen by a neurologist or PD specialist ( 5 ). Second, even when patients have access to a PD specialist, the resulting assessment is episodic and semisubjective. In particular, the Movement Disorder Society–Unified Parkinson’s Disease Rating Scale (MDS-UPDRS), widely used for assessing PD motor and nonmotor signs, is acquired through infrequent in-clinic assessments, and could be biased by both patient’s state on that particular day, such as being tired after a long drive to the medical center, and the subjectivity of patient or rater ( 6 , 7 ). These semisubjective episodic assessments are also widely used in clinical trials and, as such, affect the success of therapeutic development. Thus, there is an urgent need for objective, accurate, and continuous assessment of PD at home. Some researchers have looked into the possibility of using wearable and portable devices to monitor individuals with PD ( 8 – 11 ); however, solutions that require users to wear and manage devices can be problematic in this population, which, besides their motor difficulties, tends to be older, less technology savvy, and potentially cognitively impaired ( 12 ). Here, we present an approach for continuous assessment of PD at home without asking patients to wear sensors or actively measure themselves. We have developed a passive activity monitor that looks like a home Wi-Fi router. The device acts as low-power human radar; it sends out wireless signals (1000 times lower power than home Wi-Fi) and collects their reflections from the environment and nearby people. We used advanced signal processing and machine learning algorithms to analyze the reflected radio signals and extracted walking movements and trajectories through the home. We further analyzed these measurements to compute participants’ gait speed and assessed its correlation with PD severity, disease progression, and medication response, as illustrated in Fig. 1 . To evaluate our approach, we conducted two observational studies in which we monitored 50 participants (34 with PD and 16 age-matched control subjects) continuously in their homes. The first study enrolled 20 participants, who were monitored for 2 months. The remaining participants (60%) are enrolled in an ongoing 2-year observational study. The evaluation demonstrated that unscripted gait measurements collected using our radio device provide a valid marker of PD severity, progression, and medication response. Now, most of drug clinical trials and medical care are limited to subjective episodic PD measurements, typically conducted in a clinic. In contrast, the approach described here demonstrates the feasibility of assessing individual patients in their natural living environments to obtain objective, detailed, and clinically meaningful measurements of their disease. Furthermore, these measurements can be collected in a passive manner without asking patients to wear sensors or interfering with their lives. Our results also motivate the use of in-home radio-based gait monitoring to study other diseases associated with gait and motor dysfunction, such as Huntington’s disease ( 13 , 14 ) and multiple sclerosis ( 15 ).

SUPPLEMENTARY MATERIALS www.science.org/doi/scitranslmed.adc9669 Methods Figs. S1 to S3 Tables S1 to S9 Data file S1 References ( 86 – 92 ) View/request a protocol for this paper from Bio-protocol .