Clinical and methodological challenges for assessing freezing of gait: future perspectives

INTRODUCTION

Freezing of gait (FOG) is among the most disabling symptoms in Parkinson’s disease (PD), as it reduces mobility and frequently results in falls and fall-related injuries¹. During an episode of FOG, patients experience a feeling as if their feet are being glued to the floor. These episodes are typically provoked during turning (particularly narrow turns in tight quarters), gait initiation, when walking through narrow passages such as doorways or when performing a concomitant task when walking.^{2, 3} More recently, it has become clear that anxiety is also a common provocative factor for FOG^{4, 5}.

Different behavioural phenotypes of FOG have been described. One phenotype classification is based on the type of leg movements: 1) FOG with alternating trembling of the legs with a frequency of 3–8 Hz, 2) FOG during which patients make very small shuffling steps, and 3) akinetic FOG when no movements of the legs are observed⁶. Another phenotype classification is based on the possible underlying neurobiological underpinnings and their related triggering events: 1) motor impairments (eg difficulties to turn on the spot), 2) increased levels of anxiety (eg more freezing when in a hurry) and 3) attentional impairments (eg difficulties performing a cognitive dual task when walking)⁷. It is yet unclear whether these different manifestations of FOG share the same underlying mechanism.

Despite its common and disabling features in everyday circumstances, FOG is surprisingly difficult to assess in daily clinical practice. This has several reasons. First, FOG disappears when patients shift from an automatic control of gait towards a more goal-directed one^{8, 9}. In other words, gait improves when patients consciously focus on walking, and this is what typically happens when the gait pattern is being examined by a clinician. Second, the typically well-lit and broad hospital corridors without obstacles – where patients are often assessed – are in many ways ideal for NOT provoking FOG, as the phenomenon is commonly provoked in narrow spaces and while turning, or in dimly lit environmnets where patients cannot use visual feedback to compensate¹⁰. Third, patients are typically ‘ON’ dopaminergic medication when they visit the clinician, while FOG occurs more often (and is more severe) when the dopaminergic medication has worn off⁶.

These difficulties to assess FOG are problematic as optimal management depends on valid and reliable assessment³. Assessment involves the very identification of FOG (i.e. the presence or absence of FOG), and also aims to map its’ severity. Reliable assessment of FOG is needed to initiate or change treatment (e.g. to start or increase the dosage of levodopa, or to initiate a referral to a specialised physiotherapist) and to evaluate its effect. Furthermore, the future evaluation of novel symptomatic therapies will demand reliable and objective outcomes.

In this narrative review, we will elaborate on the current state-of-the-art assessment of FOG for clinical and research purposes, as discussed in a specific session during the workshop. Moreover, we will highlight how wearable technology might improve the assessment in the near future, and what challenges need to be overcome to move these novel technologies from the research setting into daily clinical practice.

ASSESSING THE PRESENCE AND SEVERITY OF FOG IN CLINICAL PRACTICE

ASSESSING THE PRESENCE AND SEVERITY OF FOG WITH WEARABLE SENSORS

In research laboratory settings, multiple studies have objectively characterized FOG using instruments available in traditional gait laboratories, such as motion analysis to measure body kinematics²⁴, foot switches to measure foot contact, ^{25, 26} force plates²⁷ to measure forces under the feet or surface EMG²⁸ to characterize lower limb muscle activity. More recently, wearable inertial sensors have been used to characterize episodes of FOG^29–32. The possibility of detecting deterioration of gait pattern prior to FOG, and assessing FOG severity in daily life with wearable technology has emerged as a valid, potential tool^{33, 34}. Being able to predict an upcoming FOG episode – by analyzing changes in the quality of gait – is particularly interesting from possible therapeutics perspective, because this would potentially allow for a timely coupling to an intervention to maintain or improve the gait quality and thereby prevent the FOG episode from happening. An additional advantage is that the intervention – such as sensory cues – do not have to be presented continuously to the patient, which may go at the expense of compliance, and which carries the risk of habituation to the cues. Wearable inertial sensors, typically embedding a tri-axial accelerometer, gyroscope and magnetometers, have been used increasingly in the past decade to objectively measure static and dynamic balance in PD, walking and postural transitions both in the laboratory and only very recently in the home settings³⁵.

Regarding the algorithm used to characterize occurrence and severity of FOG, Hausdorff et al. and Moore et al. pioneered the use of leg movement frequency to quantify the typical trembling of the legs during FOG with wearable sensors (foot switches²⁶ or accelerometers³¹). This approach takes advantage of the fact that ‘high-frequency’ components of movement have been associated with the leg ‘trembling’ phenotype of FOG (at frequencies much higher than that associated with regular gait itself). Therefore, by placing wearable accelerometers on the shins, a frequency ratio can be calculated as the square of the total power of the signal in the 3–8 Hz band, defined as “freezing band” divided by the square of the total power in the 0.5–3 Hz band, defined as “movement band”, see Figure 1.

This algorithm, when applied in the laboratory, can reliably identify the occurrence (start and end) of freezing episodes and measure their duration under controlled conditions such as when initiating gait, turning, approaching an obstacle³¹ or during a Timed-Up and Go test²². It is based on the assumption that when the frequency ratio exceeds a certain threshold, a FOG episode is considered to have occured. Specifically, 78% of the freezing episodes were detected using a global “one size fits all” threshold for the frequency ratio, and the freezing detection reached 89% when the threshold was calibrated for each subject by accounting for the mean and standard deviation of the frequency of each subjects’ standing³¹. However, this approach was originally described for offline data analysis and not real-time FOG detection, and, even when correcting for each individual’s standing frequency, is prone to false positives when subjects are at rest or are performing other motor tasks. Furthermore, this approach does not classify any akinetic freezing, because none of the subjects experienced it in the studies reported so far. Akinetic freezing is likely be difficult to detect using the frequency ratio because the typical signature (trembling) is by definition missing. Algorithms can be taight to see that the patient is not moving, but whether this is a voluntary stop or an involuntary stop due to FOG may well be very difficult – if not impossible – to decide.

To overcome some of the latter limitations, different groups have modified the approach to either detect the occurrence and duration of FOG episodes^{32, 36–41} or to measure overall FOG severity^{42, 43}. Specifically, the introduction of a second threshold, detecting whether relevant motion was present when the frequency ratio was supposedly identifying a FOG episode, significantly improved the algorithm performance to detect FOG occurrence,³⁶ still keeping the computational cost of the algorithm low (low amount of resources necessary to run it) which is essential for online FOG detection. However, these methods seem to work better with relatively long temporal windows (optimal at 4.5s³⁶), and therefore, brief FOG episodes might be missed. More recently, other studies utilizing machine learning approaches have been applied to identify the best set of features to identify a FOG episode from sensors worn on the waist at home or on the shins in the laboratory environment ^44–47. Despite the higher sensitivity in detecting the occurrence of even shorter FOG episodes compared to the previous method (an accuracy above 90% was achieved), these approaches may require a higher computational cost (high amount of resources necessary to run the algorithm), requiring up to about a minute from the occurrence of the episode to its detection⁴⁴.

A recent review summarized the aforementioned studies and pointed out that the achieved sensitivity in detecting FOG occurrence, under controlled (laboratory settings) and uncontrolled (home setting) conditions ranges from 73 to 100%, the specificity ranged from 67 to 100%, and the accuracy ranged from 68% to 96%³⁴. Specifically, among the different sensors used, accelerometers were the most widely used, either as a single sensor or in combination with gyroscopes or magnetometers³⁴. The location used mostly was the shin in 66% of the studies, followed by the waist (33%; see Figure 2).

It is important to note that only a minority of studies were performed in the home setting (only two of the 18 papers in the review), and overall, in both controlled and uncontrolled settings, the validity of such approaches varied considerably (accuracy 68% to 96%). Of the two home-based studies, the most promising one used one sensor on the belt, achieving in the best case 91.7% sensitivity and 87.4% specificity using a machine learning approach trained on 21 patients with PD⁴⁶. However, this approach has not been validated in a new population, so the performance of this algorithm in another group of people with PD remains unclear. Evaluation of overfitting and validation on a different cohort is of paramount importance when using machine learning approaches that can fit a specific set of data very well, but may fail on a new set^{48, 49}. An additional short report measured the amount of FOG and its variability using a mixed approach (frequency of right and left foot accelerations and correlation of right and left foor angular velocities) with two wearable sensors on the feet worn for seven days at home⁵⁰. Results showed that quantitative measures of FOG were associated to the perceived FOG, assessed by the N-FOGQ, in a sample of fourteen subjects with FOG⁵⁰.

Figure 2 summarizes the main sources of variability on accuracy of FOG detection. Specifically, a large amount of variability can be explained by the differences in utilized protocol and sensors (see Figure 3 for sensors placement), with most consistent results when sensors are located on the shins. Another source is the difference in algorithm used, in fact there is clearly a variation across the original Moore Freezing ratio, a Freezing ratio adapted to an individual subject, or one derived from machine learning approaches, which whilst potentially offering better accuracy in detecting FOG may come with temporal delay because of its’ higher computational cost.

Medication status and heterogeneity of FOG in the measured group are poorly investigated and can represent a source of variability, in fact the majority of subjects tend to experience more freezing when OFF medication and among the subjects included in the same study there may be different subtypes of FOG patients⁷.

Lastly, an important limitation, common to all the cited studies, is the relatively small sample size, with N varying from 1 to 25, no approach developed in one cohort and tested in another one, and the fact that psychometric properties, apart from validity and reliability, are not fully established (for instance the clinical important change).

There is always a trade-off between optimal sensitivity and specificity on the one hand, and signal processing operations requiring significant time and computational resources on the other hand. When data analysis is done offline, computational resources do not respresent an obstacle. However, when portable devices must be able to detect FOG in real time (for example, because this could trigger a connected cueing device aimed at overcoming the freeze), then processing must be done immediately by high-accuracy algorithms. This processing time may well introduce significant delays in FOG detection. Moreover, the required computational resources could negatively affect power consumption and hence reduce the wear time. This coulbd be a real concern as adoption by patients will diminish if the detection device has to be recharged very frequently.

Conclusion

The summarized research findings on the measurement of FOG, both in the laboratory and more importantly at home, must be appropriately translated to both clinicians, therapists, caregivers, and patients, with the ultimate aim of improving the management of this debilitating symptom. However, the route to readily achieve this knowledge transfer is not fully clear at this point, although it should follow, where possible, the guidelines recently proposed by the MDS Task Force on Technology⁷³. Focused meetings aiming to determine which kind of information would be most beneficial for clinicians and therapists could be a start. Standardization procedures for protocol and sensors to use in the laboratory/clinic and sharing collected data available across centers, specifically for the data collected at home, would be of tremendous help to validate the most promising approaches to detect FOG.