Freezing of gait (FOG) is a gait impairment which occurs in Parkinson's disease (PD) patients. As PD progresses, the patient is unable to perform locomotion normally. This increases the risk of falls and adversely affects the patient's quality of life. In this article, a user-independent method has been proposed to detect FOG events in PD patients. The proposed method is divided into three phases. Phase-1 extracts the statistical features from a FOG dataset. Phase-2 divides the data into two clusters based on FOG events. Phase-3 selects significant factors, using a randomized block design with replication. A Random Forest model is built using a combination of significant factors obtained from the design of experiments. The proposed method classifies FOG events with an average sensitivity up to 94.33% and specificity up to 92.77%. This model can be integrated along with non-pharmaceutical treatments to generate sensory-motor feedback at the onset of a FOG event.
TopIntroduction
Parkinson's disease (PD) is a chronic neurological condition that affects a patient's ability to perform complex neuro-motor tasks such as walking, body balance, and writing (Morris, Iansek, & Churchyard, 1998). This arises due to the loss of neuromelanin containing dopamine neurons (Fahn, 2003). A PD patient shows four primary motor-based symptoms often grouped under the abbreviation TRAP, which stands for Tremor at rest, Rigidity, Akinesia, and Postural instability (Jankovic, 2008).
Freezing of gait (FOG) is a brief, and unpredictable period where a patient suffers from a sharp reduction that affects movement, despite the inclination to walk (Heremans, Nieuwboer, & Vercruysse, 2013; Nutt et al., 2011). Some features of FOG include a decrease in step length despite the willingness to move, not being able to clear the ground with one’s feet or toes, the sensation of not being able to move one’s feet along with freezing episodes, trembling in one or both legs (Nutt et al., 2011). FOG events can lead to frequent falling and injuries, which discourages a patient to walk and negatively impacts their quality of life.
Levodopa is the most effective drug to treat PD (Joseph Jankovic & Aguilar, 2008). Other available drugs are dopamine agonists (DA), catechol-o-methyl-transferase (COMT) inhibitors, and non-dopaminergic agents (Joseph Jankovic & Aguilar, 2008). Once the drug has been consumed, the period until another consumption can be divided into two stages: ‘ON' stage, when the drug is active, and ‘OFF’ stage, when the effect of the drug has faded (Bächlin et al., 2010). About 70% of FOG events are observed in advanced stages of PD; however, 26% of FOG events are reported in the early stages where patients have not undergone levodopa therapy (Heremans et al., 2013).
As the disease progresses, FOG becomes more resistant to pharmacologic treatments (Bloem, Hausdorff, Visser, & Giladi, 2004). In such cases, non-pharmaceutical treatments such as visual, auditory, tactile cueing (Boelen, 2007), physiotherapy (Nieuwboer, 2008), and deep brain stimulation (Ferraye, Debû, & Pollak, 2008) can act as alternative treatments.
Wearable sensors are placed on different parts of the body to understand the pattern of FOG events and continuously monitor the PD patient. Some commonly used sensors to collect FOG data are tri-axial accelerometers, gyroscopes, and magnetometers. The data collected can be further used to develop methods to detect and classify FOG events. Such methods can be incorporated along with non-pharmaceutical treatments to generate sensory-motor feedback such as the previously mentioned forms: visual, auditory, and tactile cueing. This feedback can be provided to PD patients at the onset of a FOG event. Thus, creating a closed loop sensory-motor feedback to the PD patients in order to overcome the FOG, which can prevent fall and injuries.
This paper aims to improve FOG detection and classification by developing a user-independent model to provide sensory-motor feedback. Misclassification during the prediction of FOG events as “no FOG” can be dangerous for the life of the patient. Conversely, the vice-versa scenario can generate false feedback, which gives misclassified no FOG but is harmless for the patients. Thus, the prediction of FOG events is more important than no FOG events. Hence, this paper focuses on enhancing the prediction of FOG events. As the patient may have a varying range of FOG events, a single classification model may not perform well over an entire group of patients, and training an individual classifier for each patient is time-consuming, costly, and infeasible. Therefore, this study presents a new approach based on clustering and classification of FOG events. To achieve this, patients are grouped into different clusters based on up to 99 features and a Random forest classifier is developed for each cluster.
This paper has organized an order: background is present in section 2, database and approach present in section 3, description of the methodology and results is in section 4, discussion is in section 5 and finally the conclusion in section 6.