Background
Gait disturbances, such as reduced gait speed, shorter stride length, increased time of double support and slow turns, occur early in Parkinson’s disease (PD) and progress over time [
1,
2]
. It is also estimated that over 80% of people with PD eventually develop freezing of gait (FoG), an intermittent failure to initiate or maintain locomotion [
3,
4]. FoG and slow walking are the most significant factors affecting the quality of life in people with PD and are associated with an increased risk of falls [
5]. FoG episodes can be very short (< 1 s), short (2–5 s) or long (> 5 s) and are more common during walking conditions typical of daily life than during straight walking in a clinic or laboratory (i.e.; turning, gait initiation, when walking through doorways or when performing a concurrent dual-task when walking [
6,
7]).
Objectively assessing the severity of FoG is challenging from both a clinical and a research perspective [
7,
8]. In fact, as recently summarized in our previous work [
8], there still isn’t an optimal freezing score that clinicians can use. The ‘gold-standard’ to assess the presence of freezing (from actual video recordings [
9,
10] or computer-generated animations [
11]) is time consuming and does not represent daily fluctuations. Assessment of FoG in the clinic or laboratory is challenged by the fact that these assessment do not accurately represent severity or extent of FoG in daily life [
12,
13]. Increased attention, alertness, and effort to impress the examiner during testing may improve gait performance [
14‐
16]. This is particularly true for FoG, in fact FoG is difficult to elicit during a clinical visit or in the laboratory [
13,
17,
18] when participants focus attention on their walking. As walking and turning while dual‐tasking (DT) have been suggested to induce freezing, the addition of a DT is often used to elicit FoG in the laboratory environment [
12,
19].
Significant advancements in technology using wearable inertial sensors provides a new opportunity to objectively quantify subtle gait disturbances, such as FoG, in both clinical and laboratory settings [
20‐
22], and ultimately during daily life [
20,
21]. Objective measures of gait disturbances, such as FoG, have the potential to help inform effects of treatment, disease progression, and characterize fall risk.
Two recent reviews [
23,
24] have summarized different approaches to objectively measure FoG with wearable sensors. However, only three studies were performed in the home setting and the validity of the algorithms in the laboratory or home varied considerably (accuracy 79% to 96%) [
25‐
28]. While studies detecting FoG in a laboratory setting have been well-validated, studies focused on detecting FoG during daily life are relatively scarse [
23,
25,
26,
29‐
32]. In addition, the percentage of freezing during daily life, as well as the variability of it, have not yet been reported. In addition, the impact of FoG on mobility perception and other gait disturbances have not yet been investigated during daily life. Finally, open-source solutions to monitor FoG with wearable inertial sensors in free living conditions are not yet available. Common algorithms available to many investigators will improve reproducibility of results, external validation, algorithm improvements, and ultimately reducing barriers to applying digital health solutions for unsupervised FoG monitoring.
Here, we aimed to: (1) introduce a novel, objective algorithm to detect FoG episodes in the laboratory in a cohort of people with PD with and without freezing of gait compared to age-matched controls, and evaluate the performance of such algorithm with clinicians judgment of FoG; and (2) extend this approach to characterize FoG during daily life (7 days recording with inertial sensors on the feet) as well as investigate the association between subjects’ perception of freezing severity and other objective measures of walking and turning in a different cohort of people with PD with, and without, freezing of gait.
Discussion
This study introduces a novel, automated algorithm for detection and objective characterization of FoG episodes from inertial sensors on the feet. The proposed algorithm is simple and threshold-based, with one threshold based on angular velocity data and one on accelerometry data, to identify FoG episodes. Overall, we showed better agreement between clinical raters and the algorithm in detecting the number of FoG episodes in the laboratory (Study I) for long FoG episodes. In fact, for very-short and short FoG episodes, the ICCs during dual-task walking were lower compared to ICCs calculated for single-task walking.
Further, we explored this approach during unsupervised home monitoring (Study II) and found that the proposed FoG proxies, percent of walking time spent freezing and the variability of time spent freezing, were different between people with and without FoG. The percent of walking time spent freezing also was related to disease severity, measured with the MDS-UPDRS Part III and perception of mobility, measured with the mobility sub-score of the PDQ-39.
Here, we modified a threshold-based approach to detect FoG recently presented [
50]. Specifically, the proposed open-source algorithm first detects periods of walking and turning [
42,
50], then applies two thresholds, which need to be satisfied to label an episode as a “FoG episode”. Specifically, the first threshold is on the spectral power of the data coming from the accelerometers, a common way to identify FoG presence, based on the high-frequency components of the legs trembling [
51]. This method has well-known advantages and disadvantages [
24,
52], and can improve the detection of FoG episodes. To also detect FoG episodes not involving trembling of the knees, we added a threshold, based on the correlation between the right and left angular velocity of the feet. Usually, during regular walking, the correlation between right and left foot angular velocity is high, while it drops significantly prior to, and during, a FoG episode [
45,
46]. The same approach can be applied to wearable sensors placed on the feet or shins. In the present study, while evaluating the performance of this approach with the clinical raters, we found that: (1) the overall agreement between clinical and objective detection of FoG is strong when the sensors are placed on the feet, and moderate when the sensors are placed on the shins, (2) both the agreement between the two clinical raters, as well as the agreement between clinical raters and objective measures are better for the short (2–5 s) and long (> 5 s) FoG duration, whereas both agreements are poor for the very short FoG episodes (< 1 s); (3) in general the agreement in the dual-task walking condition seem lower compared to the single-task walking conditions for the short FoG episodes.
Recently, machine learning based methods [
24] (neural networks, decision trees, random forest, and support vector machine) have been proposed to surpass the FoG detection abilities of threshold-based methods. However, it is still unclear whether an algorithm that matches perfectly with clinical judgement is needed, even more so, when there is still discrepancy among clinical raters with more or less experience in detecting the same FoG episodes specifically for very short FoG episodes where the agreement between different raters is poor. Moreover, despite the higher sensitivity in detecting the occurrence of even shorter FOG episodes compared to the previous method [
24] (an accuracy above 90% was achieved), these approaches may require a higher computational cost, requiring up to several seconds from the occurrence of the episode to its detection, making those algorithms not suitable for real-time interventions, such as cueing. However, nowadays, the use of floating-point unit microcontrollers could overcome this limitation, in fact, such microcontrollers could compute advanced machine learning algorithms in real time with low power consumption.
Overall, the proposed approach reached AUC of 0.89 to 0.93 in discriminating people experiencing FoG or not, when FoG was classified by movement disorder neurologists. These AUCs and relative sensitivity, specificity and accuracy are similar to what reported in the literature using a variety of approaches [
23,
24]. Lastly, we observed lower agreement between raters and the algorithm for short episodes during dual-task walking. Although this should be verified in a separate cohort, a pontential explanation could be related to decrease smoothness of walking in people with PD and particularly in freezers [
53,
54]. It could be possible that dual-task further decreases smoothness of gait, such decrease may be picked up as freezing by the algorithm but not by the clinical raters.
After comparing the algorithm with clinical judgment in the laboratory, we extended our approach to unsupervised monitoring during daily life for 7 days in 48 people with PD, 23 of which reported having FoG according to the NFOGQ. The percentage of time spent freezing was significantly higher in those people who report themselves as freezers compared to non-freezers, while the variability of time spent freezing was lower in those reporting FoG. The lower variability found in freezers may indicate that a certain amount of FoG is present across the day and week. Instead, the higher variability found in the non-freezers may either indicate that we are picking up subtle hesitations that are not constantly present over the day, therefore increasing the variability, or the high variability could be due to a high false positive rate. Therefore, to confirm this finding, we would need to first validate the algorithm during daily life and follow longitudinally the same cohort of people with PD who does not report FoG at baseline.
Interestingly, we found that the short, 2–5 s duration, FoG episodes account for 69% of all episodes, while the rest are long (5–30 s) episodes, 1% of which have duration over 30 s. This suggests that short FoG episodes are the most common during the day. The percent of walking time spent freezing and its variability were related to disease severity, measured with the MDS-UPDRS Part III and to perceived mobility, measured with the PDQ-39, only in the freezers. The association between percent time spent freezing and the MDS-UPDRS Part III is not totally surprising, as freezing severity tends to increase with disease severity. The association between the variability of time spent freezing and mobility perception suggests that FoG may affect the perception of mobility in people with PD. Our objective measures of FoG during 7 days of continuous monitoring were not significantly associated with the NFOGQ performed at the beginning of the study, indicating that perception of FoG may be differ from the measured FoG over a week of monitoring. Although surprising, this may, in part, be explained by the items composing the NFOGQ. The questionnaire asks about the impact of freezing in daily life, in addition to the presence and severity of freezing. For some people, even mild freezing may significantly disturb walking,and cause fear of falling that may significantly impact or cause people to avoid activities of daily life.
We also characterize quantity and quality of walking and turning over 7 days of continuous monitoring. Our findings are in keeping with studies showing that
quantity of walking and turning is similar among people with and without FoG [
42,
55] while
quality of walking and turning may be more affected in people with FoG. Specifically, the average pitch angle at initial contact of foot with the ground was significantly smaller in freezers compared to non-freezers, consistent with more shuffling gait and more falls in freezers than non-freezers. The large variability of the pitch angle at initial contact and the high variability of the time spent freezing could potentially reflect fluctuations in number of freezing episodes due to periodic medication intake throughout the day.
In addition, the average turning angle was smaller in freezers compared to non-freezers, as previously reported in a larger cohort [
42]; and such difference could potentially be attributed to the fact that freezers may avoid larger turning angles, known to elicit more freezing, and explain that turning duration was significantly shorter in freezers. However, after correcting these outcomes for disease duration, only turning angle was still statistically significant, suggesting that gait disturbances such as shuffling are related to disease duration more than to freezing of gait. Turning angle was still significantly smaller in freezers compared to non-freezers after correcting for disease duration suggesting that freezers may modify their turning in order to avoid FoG. However, it is also possible that average turning angles were measured as small in freezers because they hesitated during a turn such that a large turn was detected as several small turns.
These findings, although promising, should be taken cautiously. Future work will need to validate the algorithm on a new dataset, increase the number of subjects at home and determine the validity of our objective freezing measures in daily life. Specifically, we plan to use either a mini-camera pointed at the feet or pressure insoles as a gold standard comparison for home recording of gait and turning, for comparison with the inertial sensor data. It is also possible that some participants with FoG show akinetic freezing, not involving trembling of the knees, so these events may not be identified with our threshold approach. At this time, we hope that by making this algorithm available to researchers, we could, together, further improve FoG detection.
Publisher's Note
Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.