Freezing of gait in Parkinson's disease: automatic early recognition of episodes from patients' inertial data

Freezing of gait (FoG) in Parkinson's disease is “a brief, episodic absence or marked reduction of forward progression of the feet despite the intention to walk”. It is a common symptom in 50-80 % of the patients, occurring most frequently during turning, when passing through narrow paths or overcoming an obstacle, and is strongly affected by cognitive aspects such as attention, stress or anxiety. Having a strong correlation with falls, FoG is considered as one of the most dangerous symptoms, with severe effects on autonomy and Quality of Life (QoL). Due to the great variability of FoG features , treatments have to be patient-specific. This aspect, together with the need of further investigating its causes, requires to develop objective methodologies to assess FoG. Laboratory assessment typically includes FoG eliciting in an artificial setting, with an altered emotional state of the patient due to examination anxiety. Thus many literature studies have been recently focused on home evaluation systems. In this regard, a popular option is based on a combination of data coming from wearable inertial sensors with machine learning algorithms, trained in order to detect FoG events. Although different, all these methodologies share a similar processing pipeline. The aim of this study was to explore a new approach to the problem, performing additional offline preprocessing to identify “region of interest" in the data, where there is a higher probability to identify FoG. Such step was introduced to increase time efficiency during the analysis and the precision of the final classifier predictions. In this study, tri-axial acceleration data coming from two previous experiments (defined as “Phase 1” and “Phase 2”), involving a total of 85 subjects and a commercial smartphone for data collection, were employed. Phase 1 data were used to design the algorithm for the selection of "region of interest", based on Continuous Wavelet Transform (CWT). On the selected pieces, a more traditional pipeline was implemented: the signals were segmented through a sliding window; from each segment a set of relevant spectral and temporal features was extracted and fed in input to several classifiers (K-NN, SVM, Random Forest) to compare their performances. Both multi-class and binary classification were explored. The best results were obtained using binary Support Vector Machine with RBF kernel, achieving accuracy, recall and precision of 95%, 84%, 90% respectively in a 10-fold stratified crossvalidation. Furthermore, the use of the implemented window of interest was found to increase specificity, in a false positive test over non-freezer patients, by 35 percentage points (from 62% to 97%). Even though the small amount of FoG data employed reduces the significance of the statistical results and some effort would be required to translate this offline approach to an online implementation, overall the importance of applying further preprocessing to the inertial data can be considered demonstrated.