Knowledge representation and machine learning on wearable sensor data: a study on gait monitoring

Context of the research

The appearance of new wearable and non-invasive sensors is providing new and disruptive solutions for the monitoring of patients, disease management and follow-up of treatment adherence, and opens the door to the appearance of new and more personalized treatment methods.

A sensor is a device that measures a physical quantity and transforms it into a digital signal. It provides great amounts of continuous raw measurements that can be difficult to interpret by physicians or nurses. Hence, processing the raw measurements into potential clinical findings and biomarkers (towards the digital phenotyping) is becoming a key issue. In that regard, artificial intelligence techniques play an important role. Nevertheless, the huge amount of existing methods requires skilled people able to identify the suitable knowledge representation to be used, the most appropriate machine learning technique, which in turn depend on the kind of data available, the task to be performed (e.g. diagnosing or treatment), and how the quality of the learned findings will be measured.

Objectives

We aim to help engineers and researchers in selecting the most appropriate technique. To that end, we analyze different knowledge representation and machine learning methods on data gathered by a wearable sensor. In particular, the raw data is obtained from force sensors installed in a person shoes (sandals), and the data has been gathered when the patient walks. From the data, gait patterns can be obtained to support clinical decisions. In particular, the recovery period of a person which has suffered a hip surgery: adequate (short) or not-adequate (long); therefore, we are dealing with a classification task.

Description

Knowledge representation methods are analyzed according to two dimensions: abstraction (raw measurements, feature descriptions, and labels) and granularity (high and low).

On the other hand, we analyze machine learning methods according to the following dimensions: temporal (whether the methods are able to deal with chronological data or not); modelling facilities (availability of a generated model: eager methods generate an explicit model; lazy methods do not), and modelling complexity (linear versus non-linear). A total of 7 methods from different machine learning families are reviewed: k-nearest neighbor (KNN), Support Vector Machines (SVM), Artificial Neural Networks (ANN and recurrent ANN), Sequence-learning, and Stream-learning The methods selected are a representative sample of what is being used nowadays with wearable measurements.

Results

Methods are analyzed in terms of user interpretation, time, memory and adaptation (i.e. capacity of following physiological changes of the persons). High granularity representations that maintains the time order of the original data are the knowledge representation methods that exhibit the higher performance when combined with stream data learning. Other simple methods, as kNN have closer results to the best while offering excellent adaptation outcomes.

Conclusions

Wearable sensors provide from raw data to which several artificial intelligence methods can be applied to support either clinical or patients' decisions. In this work we present several features that should be taken into account when selecting the methods, and some recommendations when dealing with force sensors to conduct gait analysis.