Towards Real-Time Prediction of Freezing of Gait in Patients With Parkinson's Disease: Addressing the Class Imbalance Problem

Insights into Parkinson's Disease-Related Freezing of Gait Detection and Prediction Approaches: A Meta Analysis

Parkinson's Disease (PD) is a complex neurodegenerative disorder characterized by a spectrum of motor and non-motor symptoms, prominently featuring the freezing of gait (FOG), which significantly impairs patients' quality of life. Despite extensive research, the precise mechanisms underlying FOG remain elusive, posing challenges for effective management and treatment. This paper presents a comprehensive meta-analysis of FOG prediction and detection methodologies, with a focus on the integration of wearable sensor technology and machine learning (ML) approaches. Through an exhaustive review of the literature, this study identifies key trends, datasets, preprocessing techniques, feature extraction methods, evaluation metrics, and comparative analyses between ML and non-ML approaches. The analysis also explores the utilization of cueing devices. The limited adoption of explainable AI (XAI) approaches in FOG prediction research represents a significant gap. Improving user acceptance and comprehension requires an understanding of the logic underlying algorithm predictions. Current FOG detection and prediction research has a number of limitations, which are identified in the discussion. These include issues with cueing devices, dataset constraints, ethical and privacy concerns, financial and accessibility restrictions, and the requirement for multidisciplinary collaboration. Future research avenues center on refining explainability, expanding and diversifying datasets, adhering to user requirements, and increasing detection and prediction accuracy. The findings contribute to advancing the understanding of FOG and offer valuable guidance for the development of more effective detection and prediction methodologies, ultimately benefiting individuals affected by PD.

Freezing of gait assessment with inertial measurement units and deep learning: effect of tasks, medication states, and stops

BACKGROUND

Freezing of gait (FOG) is an episodic and highly disabling symptom of Parkinson's Disease (PD). Traditionally, FOG assessment relies on time-consuming visual inspection of camera footage. Therefore, previous studies have proposed portable and automated solutions to annotate FOG. However, automated FOG assessment is challenging due to gait variability caused by medication effects and varying FOG-provoking tasks. Moreover, whether automated approaches can differentiate FOG from typical everyday movements, such as volitional stops, remains to be determined. To address these questions, we evaluated an automated FOG assessment model with deep learning (DL) based on inertial measurement units (IMUs). We assessed its performance trained on all standardized FOG-provoking tasks and medication states, as well as on specific tasks and medication states. Furthermore, we examined the effect of adding stopping periods on FOG detection performance.

METHODS

Twelve PD patients with self-reported FOG (mean age 69.33 ± 6.02 years) completed a FOG-provoking protocol, including timed-up-and-go and 360-degree turning-in-place tasks in On/Off dopaminergic medication states with/without volitional stopping. IMUs were attached to the pelvis and both sides of the tibia and talus. A temporal convolutional network (TCN) was used to detect FOG episodes. FOG severity was quantified by the percentage of time frozen (%TF) and the number of freezing episodes (#FOG). The agreement between the model-generated outcomes and the gold standard experts' video annotation was assessed by the intra-class correlation coefficient (ICC).

RESULTS

For FOG assessment in trials without stopping, the agreement of our model was strong (ICC (%TF) = 0.92 [0.68, 0.98]; ICC(#FOG) = 0.95 [0.72, 0.99]). Models trained on a specific FOG-provoking task could not generalize to unseen tasks, while models trained on a specific medication state could generalize to unseen states. For assessment in trials with stopping, the agreement of our model was moderately strong (ICC (%TF) = 0.95 [0.73, 0.99]; ICC (#FOG) = 0.79 [0.46, 0.94]), but only when stopping was included in the training data.

CONCLUSION

A TCN trained on IMU signals allows valid FOG assessment in trials with/without stops containing different medication states and FOG-provoking tasks. These results are encouraging and enable future work investigating automated FOG assessment during everyday life.

Detection and prediction of freezing of gait with wearable sensors in Parkinson's disease

Freezing of gait (FoG) is one of the most distressing symptoms of Parkinson's Disease (PD), commonly occurring in patients at middle and late stages of the disease. Automatic and accurate FoG detection and prediction have emerged as a promising tool for long-term monitoring of PD and implementation of gait assistance systems. This paper reviews the recent development of FoG detection and prediction using wearable sensors, with attention on identifying knowledge gaps that need to be filled in future research. This review searched the PubMed and Web of Science databases to collect studies that detect or predict FoG with wearable sensors. After screening, 89 of 270 articles were included. The data description, extracted features, detection/prediction methods, and classification performance were extracted from the articles. As the number of papers of this area is increasing, the performance has been steadily improved. However, small datasets and inconsistent evaluation processes still hinder the application of FoG detection and prediction with wearable sensors in clinical practice.

Recent trends in wearable device used to detect freezing of gait and falls in people with Parkinson's disease: A systematic review

Background

The occurrence of freezing of gait (FOG) is often observed in moderate to last-stage Parkinson's disease (PD), leading to a high risk of falls. The emergence of the wearable device has offered the possibility of FOG detection and falls of patients with PD allowing high validation in a low-cost way.

Objective

This systematic review seeks to provide a comprehensive overview of existing literature to establish the forefront of sensors type, placement and algorithm to detect FOG and falls among patients with PD.

Methods

Two electronic databases were screened by title and abstract to summarize the state of art on FOG and fall detection with any wearable technology among patients with PD. To be eligible for inclusion, papers were required to be full-text articles published in English, and the last search was completed on September 26, 2022. Studies were excluded if they; (i) only examined cueing function for FOG, (ii) only used non-wearable devices to detect or predict FOG or falls, and (iii) did not provide sufficient details about the study design and results. A total of 1,748 articles were retrieved from two databases. However, only 75 articles were deemed to meet the inclusion criteria according to the title, abstract and full-text reviewed. Variable was extracted from chosen research, including authorship, details of the experimental object, type of sensor, device location, activities, year of publication, evaluation in real-time, the algorithm and detection performance.

Results

A total of 72 on FOG detection and 3 on fall detection were selected for data extraction. There were wide varieties of the studied population (from 1 to 131), type of sensor, placement and algorithm. The thigh and ankle were the most popular device location, and the combination of accelerometer and gyroscope was the most frequently used inertial measurement unit (IMU). Furthermore, 41.3% of the studies used the dataset as a resource to examine the validity of their algorithm. The results also showed that increasingly complex machine-learning algorithms had become the trend in FOG and fall detection.

Conclusion

These data support the application of the wearable device to access FOG and falls among patients with PD and controls. Machine learning algorithms and multiple types of sensors have become the recent trend in this field. Future work should consider an adequate sample size, and the experiment should be performed in a free-living environment. Moreover, a consensus on provoking FOG/fall, methods of assessing validity and algorithm are necessary.Systematic Review Registration: PROSPERO, identifier CRD42022370911.

Machine Learning and Wearable Sensors for the Early Detection of Balance Disorders in Parkinson's Disease

Dynamic posturography combined with wearable sensors has high sensitivity in recognizing subclinical balance abnormalities in patients with Parkinson's disease (PD). However, this approach is burdened by a high analytical load for motion analysis, potentially limiting a routine application in clinical practice. In this study, we used machine learning to distinguish PD patients from controls, as well as patients under and not under dopaminergic therapy (i.e., ON and OFF states), based on kinematic measures recorded during dynamic posturography through portable sensors. We compared 52 different classifiers derived from Decision Tree, K-Nearest Neighbor, Support Vector Machine and Artificial Neural Network with different kernel functions to automatically analyze reactive postural responses to yaw perturbations recorded through IMUs in 20 PD patients and 15 healthy subjects. To identify the most efficient machine learning algorithm, we applied three threshold-based selection criteria (i.e., accuracy, recall and precision) and one evaluation criterion (i.e., goodness index). Twenty-one out of 52 classifiers passed the three selection criteria based on a threshold of 80%. Among these, only nine classifiers were considered "optimum" in distinguishing PD patients from healthy subjects according to a goodness index ≤ 0.25. The Fine K-Nearest Neighbor was the best-performing algorithm in the automatic classification of PD patients and healthy subjects, irrespective of therapeutic condition. By contrast, none of the classifiers passed the three threshold-based selection criteria in the comparison of patients in ON and OFF states. Overall, machine learning is a suitable solution for the early identification of balance disorders in PD through the automatic analysis of kinematic data from dynamic posturography.

Prediction of Freezing of Gait in Parkinson's Disease Using Unilateral and Bilateral Plantar-Pressure Data

Background

Freezing of gait (FOG) is an intermittent walking disturbance experienced by people with Parkinson's disease (PD). FOG has been linked to falling, injury, and overall reduced mobility. Wearable sensor-based devices can detect freezes already in progress and provide a cue to help the person resume walking. While this is helpful, predicting FOG episodes before onset and providing a timely cue may prevent the freeze from occurring. Wearable sensors mounted on various body parts have been used to develop FOG prediction systems. Despite the known asymmetry of PD motor symptom manifestation, the difference between the most affected side (MAS) and least affected side (LAS) is rarely considered in FOG detection and prediction studies.

Methods

To examine the effect of using data from the MAS, LAS, or both limbs for FOG prediction, plantar pressure data were collected during a series of walking trials and used to extract time and frequency-based features. Three datasets were created using plantar pressure data from the MAS, LAS, and both sides together. ReliefF feature selection was performed. FOG prediction models were trained using the top 5, 10, 15, 20, 25, or 30 features for each dataset.

Results

The best models were the MAS model with 15 features and the LAS and bilateral models with 5 features. The LAS model had the highest sensitivity (79.5%) and identified the highest percentage of FOG episodes (94.9%). The MAS model achieved the highest specificity (84.9%) and lowest false positive rate (1.9 false positives/walking trial). Overall, the bilateral model was best with 77.3% sensitivity and 82.9% specificity. In addition, the bilateral model identified 94.2% of FOG episodes an average of 0.8 s before FOG onset. Compared to the bilateral model, the LAS model had a higher false positive rate; however, the bilateral and LAS models were similar in all the other evaluation metrics.

Conclusion

The LAS model would have similar FOG prediction performance to the bilateral model at the cost of slightly more false positives. Given the advantages of single sensor systems, the increased false positive rate may be acceptable to people with PD. Therefore, a single plantar pressure sensor placed on the LAS could be used to develop a FOG prediction system and produce performance similar to a bilateral system.

Modelling and identification of characteristic kinematic features preceding freezing of gait with convolutional neural networks and layer-wise relevance propagation

BACKGROUND

Although deep neural networks (DNNs) are showing state of the art performance in clinical gait analysis, they are considered to be black-box algorithms. In other words, there is a lack of direct understanding of a DNN's ability to identify relevant features, hindering clinical acceptance. Interpretability methods have been developed to ameliorate this concern by providing a way to explain DNN predictions.

METHODS

This paper proposes the use of an interpretability method to explain DNN decisions for classifying the movement that precedes freezing of gait (FOG), one of the most debilitating symptoms of Parkinson's disease (PD). The proposed two-stage pipeline consists of (1) a convolutional neural network (CNN) to model the reduction of movement present before a FOG episode, and (2) layer-wise relevance propagation (LRP) to visualize the underlying features that the CNN perceives as important to model the pathology. The CNN was trained with the sagittal plane kinematics from a motion capture dataset of fourteen PD patients with FOG. The robustness of the model predictions and learned features was further assessed on fourteen PD patients without FOG and fourteen age-matched healthy controls.

RESULTS

The CNN proved highly accurate in modelling the movement that precedes FOG, with 86.8% of the strides being correctly identified. However, the CNN model was unable to model the movement for one of the seven patients that froze during the protocol. The LRP interpretability case study shows that (1) the kinematic features perceived as most relevant by the CNN are the reduced peak knee flexion and the fixed ankle dorsiflexion during the swing phase, (2) very little relevance for FOG is observed in the PD patients without FOG and the healthy control subjects, and (3) the poor predictive performance of one subject is attributed to the patient's unique and severely flexed gait signature.

CONCLUSIONS

The proposed pipeline can aid clinicians in explaining DNN decisions in clinical gait analysis and aid machine learning practitioners in assessing the generalization of their models by ensuring that the predictions are based on meaningful kinematic features.

Real-Time Detection of Freezing Motions in Parkinson's Patients for Adaptive Gait Phase Synchronous Cueing

Parkinson's disease is the second most common neurodegenerative disease worldwide reducing cognitive and motoric abilities of affected persons. Freezing of Gait (FoG) is one of the severe symptoms that is observed in the late stages of the disease and considerably impairs the mobility of the person and raises the risk of falls. Due to the pathology and heterogeneity of the Parkinsonian gait cycle, especially in the case of freezing episodes, the detection of the gait phases with wearables is challenging in Parkinson's disease. This is addressed by introducing a state-automaton-based algorithm for the detection of the foot's motion phases using a shoe-placed inertial sensor. Machine-learning-based methods are investigated to classify the actual motion phase as normal or FoG-affected and to predict the outcome for the next motion phase. For this purpose, spatio-temporal gait and signal parameters are determined from the segmented movement phases. In this context, inertial sensor fusion is applied to the foot's 3D acceleration and rate of turn. Support Vector Machine (SVM) and AdaBoost classifiers have been trained on the data of 16 Parkinson's patients who had shown FoG episodes during a clinical freezing-provoking assessment course. Two clinical experts rated the video-recorded trials and marked episodes with festination, shank trembling, shuffling, or akinesia. Motion phases inside such episodes were labeled as FoG-affected. The classifiers were evaluated using leave-one-patient-out cross-validation. No statistically significant differences could be observed between the different classifiers for FoG detection (p>0.05). An SVM model with 10 features of the actual and two preceding motion phases achieved the highest average performance with 88.5 ± 5.8% sensitivity, 83.3 ± 17.1% specificity, and 92.8 ± 5.9% Area Under the Curve (AUC). The performance of predicting the behavior of the next motion phase was significantly lower compared to the detection classifiers. No statistically significant differences were found between all prediction models. An SVM-predictor with features from the two preceding motion phases had with 81.6 ± 7.7% sensitivity, 70.3 ± 18.4% specificity, and 82.8 ± 7.1% AUC the best average performance. The developed methods enable motion-phase-based FoG detection and prediction and can be utilized for closed-loop systems that provide on-demand gait-phase-synchronous cueing to mitigate FoG symptoms and to prevent complete motoric blockades.

Towards Real-time Prediction of Freezing of Gait in Patients with Parkinsons Disease: A Novel Deep One-class Classifier

Freezing of gait (FoG) is a common motor dysfunction in individuals with Parkinsons disease. FoG impairs walking and is associated with increased fall risk. On-demand external cueing systems can detect FoG and provide stimuli to help individuals overcome freezing. Predicting FoG before onset enables preemptive cueing and may prevent FoG. However, detection and prediction remain challenging. If FoG data are not available for an individual, patient-independent models have been used to detect FoG, which have shown great sensitivity and poor specificity, or vice versa. In this study, we introduce a Deep Gait Anomaly Detector (DGAD) using a transfer learning-based approach to improve FoG detection accuracy. We also evaluate the effect of data augmentation and additional pre-FoG segments on prediction rate. Seven individuals with PD performed a series of daily walking tasks wearing inertial measurement units on their ankles. The DGAD algorithm demonstrated average sensitivity and specificity of 63.0% and 98.6% (3.2% improvement compared with the highest specificity in the literature). The target models identified 87.4% of FoG onsets, with 21.9% predicted. This study demonstrates our algorithm's potential for accurate identification of FoG and delivery of cues for patients for whom no FoG data is available for model training.

Early Detection of Freezing of Gait during Walking Using Inertial Measurement Unit and Plantar Pressure Distribution Data

Freezing of gait (FOG) is a sudden and highly disruptive gait dysfunction that appears in mid to late-stage Parkinson’s disease (PD) and can lead to falling and injury. A system that predicts freezing before it occurs or detects freezing immediately after onset would generate an opportunity for FOG prevention or mitigation and thus enhance safe mobility and quality of life. This research used accelerometer, gyroscope, and plantar pressure sensors to extract 861 features from walking data collected from 11 people with FOG. Minimum-redundancy maximum-relevance and Relief-F feature selection were performed prior to training boosted ensembles of decision trees. The binary classification models identified Total-FOG or No FOG states, wherein the Total-FOG class included data windows from 2 s before the FOG onset until the end of the FOG episode. Three feature sets were compared: plantar pressure, inertial measurement unit (IMU), and both plantar pressure and IMU features. The plantar-pressure-only model had the greatest sensitivity and the IMU-only model had the greatest specificity. The best overall model used the combination of plantar pressure and IMU features, achieving 76.4% sensitivity and 86.2% specificity. Next, the Total-FOG class components were evaluated individually (i.e., Pre-FOG windows, Freeze windows, transition windows between Pre-FOG and Freeze). The best model detected windows that contained both Pre-FOG and FOG data with 85.2% sensitivity, which is equivalent to detecting FOG less than 1 s after the freeze began. Windows of FOG data were detected with 93.4% sensitivity. The IMU and plantar pressure feature-based model slightly outperformed models that used data from a single sensor type. The model achieved early detection by identifying the transition from Pre-FOG to FOG while maintaining excellent FOG detection performance (93.4% sensitivity). Therefore, if used as part of an intelligent, real-time FOG identification and cueing system, even if the Pre-FOG state were missed, the model would perform well as a freeze detection and cueing system that could improve the mobility and independence of people with PD during their daily activities.

Prediction of Freezing of Gait in Parkinson's Disease Using Wearables and Machine Learning

Freezing of gait (FOG) is one of the most troublesome symptoms of Parkinson’s disease, affecting more than 50% of patients in advanced stages of the disease. Wearable technology has been widely used for its automatic detection, and some papers have been recently published in the direction of its prediction. Such predictions may be used for the administration of cues, in order to prevent the occurrence of gait freezing. The aim of the present study was to propose a wearable system able to catch the typical degradation of the walking pattern preceding FOG episodes, to achieve reliable FOG prediction using machine learning algorithms and verify whether dopaminergic therapy affects the ability of our system to detect and predict FOG. Methods: A cohort of 11 Parkinson’s disease patients receiving (on) and not receiving (off) dopaminergic therapy was equipped with two inertial sensors placed on each shin, and asked to perform a timed up and go test. We performed a step-to-step segmentation of the angular velocity signals and subsequent feature extraction from both time and frequency domains. We employed a wrapper approach for feature selection and optimized different machine learning classifiers in order to catch FOG and pre-FOG episodes. Results: The implemented FOG detection algorithm achieved excellent performance in a leave-one-subject-out validation, in patients both on and off therapy. As for pre-FOG detection, the implemented classification algorithm achieved 84.1% (85.5%) sensitivity, 85.9% (86.3%) specificity and 85.5% (86.1%) accuracy in leave-one-subject-out validation, in patients on (off) therapy. When the classification model was trained with data from patients on (off) and tested on patients off (on), we found 84.0% (56.6%) sensitivity, 88.3% (92.5%) specificity and 87.4% (86.3%) accuracy. Conclusions: Machine learning models are capable of predicting FOG before its actual occurrence with adequate accuracy. The dopaminergic therapy affects pre-FOG gait patterns, thereby influencing the algorithm’s effectiveness.

A Supervised Machine Learning Approach to Detect the On/Off State in Parkinson's Disease Using Wearable Based Gait Signals

Fluctuations in motor symptoms are mostly observed in Parkinson's disease (PD) patients. This characteristic is inevitable, and can affect the quality of life of the patients. However, it is difficult to collect precise data on the fluctuation characteristics using self-reported data from PD patients. Therefore, it is necessary to develop a suitable technology that can detect the medication state, also termed the "On"/"Off" state, automatically using wearable devices; at the same time, this could be used in the home environment. Recently, wearable devices, in combination with powerful machine learning techniques, have shown the potential to be effectively used in critical healthcare applications. In this study, an algorithm is proposed that can detect the medication state automatically using wearable gait signals. A combination of features that include statistical features and spatiotemporal gait features are used as inputs to four different classifiers such as random forest, support vector machine, K nearest neighbour, and Naïve Bayes. In total, 20 PD subjects with definite motor fluctuations have been evaluated by comparing the performance of the proposed algorithm in association with the four aforementioned classifiers. It was found that random forest outperformed the other classifiers with an accuracy of 96.72%, a recall of 97.35%, and a precision of 96.92%.

Lower Limb Kinematics Trajectory Prediction Using Long Short-Term Memory Neural Networks

This study determined whether the kinematics of lower limb trajectories during walking could be extrapolated using long short-term memory (LSTM) neural networks. It was hypothesised that LSTM auto encoders could reliably forecast multiple time-step trajectories of the lower limb kinematics, specifically linear acceleration (LA) and angular velocity (AV). Using 3D motion capture, lower limb position-time coordinates were sampled (100 Hz) from six male participants (age 22 ± 2 years, height 1.77 ± 0.02 m, body mass 82 ± 4 kg) who walked for 10 min at 5 km/h on a 0% gradient motor-driven treadmill. These data were fed into an LSTM model with a sliding window of four kinematic variables with 25 samples or time steps: LA and AV for thigh and shank. The LSTM was tested to forecast five samples (i.e., time steps) of the four kinematic input variables. To attain generalisation, the model was trained on a dataset of 2,665 strides from five participants and evaluated on a test set of 1 stride from a sixth participant. The LSTM model learned the lower limb kinematic trajectories using the training samples and tested for generalisation across participants. The forecasting horizon suggested higher model reliability in predicting earlier future trajectories. The mean absolute error (MAE) was evaluated on each variable across the single tested stride, and for the five-sample forecast, it obtained 0.047 m/s2 thigh LA, 0.047 m/s2 shank LA, 0.028 deg/s thigh AV and 0.024 deg/s shank AV. All predicted trajectories were highly correlated with the measured trajectories, with correlation coefficients greater than 0.98. The motion prediction model may have a wide range of applications, such as mitigating the risk of falls or balance loss and improving the human-machine interface for wearable assistive devices.

Toward a Wearable System for Predicting Freezing of Gait in People Affected by Parkinson's Disease

Some wearable solutions exploiting on-body acceleration sensors have been proposed to recognize Freezing of Gait (FoG) in people affected by Parkinson Disease (PD). Once a FoG event is detected, these systems generate a sequence of rhythmic stimuli to allow the patient restarting the gait. While these solutions are effective in detecting FoG events, they are unable to predict FoG to prevent its occurrence. This paper fills in the gap by presenting a machine learning-based approach that classifies accelerometer data from PD patients, recognizing a pre-FOG phase to further anticipate FoG occurrence in advance. Gait was monitored by three tri-axial accelerometer sensors worn on the back, hip and ankle. Gait features were then extracted from the accelerometer's raw data through data windowing and non-linear dimensionality reduction. A k-nearest neighbor algorithm (k-NN) was used to classify gait in three classes of events: pre-FoG, no-FoG and FoG. The accuracy of the proposed solution was compared to state-of-the-art approaches. Our study showed that: (i) we achieved performances overcoming the state-of-the-art approaches in terms of FoG detection, (ii) we were able, for the very first time in the literature, to predict FoG by identifying the pre-FoG events with an average sensitivity and specificity of, respectively, 94.1% and 97.1%, and (iii) our algorithm can be executed on resource-constrained devices. Future applications include the implementation on a mobile device, and the administration of rhythmic stimuli by a wearable device to help the patient overcome the FoG.