Detecting Parkinsonian Tremor From IMU Data Collected in-the-Wild Using Deep Multiple-Instance Learning

Abstract

Parkinson's Disease (PD) is a slowly evolving neurological disease that affects about 1% of the population above 60 years old, causing symptoms that are subtle at first, but whose intensity increases as the disease progresses. Automated detection of these symptoms could offer clues as to the early onset of the disease, thus improving the expected clinical outcomes of the patients via appropriately targeted interventions. This potential has led many researchers to develop methods that use widely available sensors to measure and quantify the presence of PD symptoms such as tremor, rigidity and braykinesia. However, most of these approaches operate under controlled settings, such as in lab or at home, thus limiting their applicability under free-living conditions. In this work, we present a method for automatically identifying tremorous episodes related to PD, based on IMU signals captured via a smartphone device. We propose a Multiple-Instance Learning approach, wherein a subject is represented as an unordered bag of accelerometer signal segments and a single, expert-provided, tremor annotation. Our method combines deep feature learning with a learnable pooling stage that is able to identify key instances within the subject bag, while still being trainable end-to-end. We validate our algorithm on a newly introduced dataset of 45 subjects, containing accelerometer signals collected entirely in-the-wild. The good classification performance obtained in the conducted experiments suggests that the proposed method can efficiently navigate the noisy environment of in-the-wild recordings.

Unobtrusive detection of Parkinson’s disease from multi-modal and in-the-wild sensor data using deep learning techniques

“…The problem of automated PD detection has received much attention in recent years. A wide range of sensors have been suggested to capture specific aspects of the PD-related symptomatology, such as IMU sensors for detecting gait alterations [16][17][18][19] and hand tremor 10,[20][21][22][23] , microphones for identifying incidents of speech impairment [24][25][26][27] , keyboards (mechanical or virtual) 11,28,29 for detecting rigidity and bradykinesia, and writing equipment 9,16,30 for estimating the level of fine motor impairment.…”

“…ii) The subjects in almost all previous works had to follow scripted self-initiated data capture scenarios, such as performing specific actions, like writing or resting for some minutes 20,21,33 , pronouncing specific words and sentences 25,34 , drawing specific shapes 9,30 or transcribing given text excerpts 11,28 . Apart from our two previous works 23,29 , that collect uni-modal data in-the-wild, only one other work 35 considers the unscripted data collection problem. However, it does so limited to a home environment and involves very low scale experiments (as little as 2 PD patients), while also employing a custom-made body-worn system of accelerometers and motion sensors.…”

“…In the case of the acceleration modality, the data contribution is summarised as a set of fixed-length acceleration signal segments, namely the postural accelerations bag. Transforming the captured raw, variable-length accelerometer recordings to the PA bag involves a pre-processing and a selection step, similar to our earlier work 23 . First, problematic recordings are discarded.…”

“…On the other hand, subjects that end up with less than five segments in their bag (before the zero-padding step) are discarded altogether. Values for the aforementioned hyperparameters were selected based on previous work 23 .…”

Parkinson’s Disease Tremor Detection in the Wild Using Wearable Accelerometers

“…Papadopoulos et al [ 34 , 35 ] proposed a method for tremor detection using in-the-wild recordings from a smartphone. This method included a multiple-instance learning approach for training a deep neural network.…”