Scholarlink

Human activity recognition (HAR) is a task which, if solved by machine intelligence, has a vast potential to improve technological advances in many different fields, medical engineering being one of the more prominent ones. Recent advances in the development of wearable sensor technologies, mobile computing, cloud computing and machine learning (ML) methods made cutting-edge technology broadly available to the scientific community. One such medical application is the therapy and monitoring optimization of subjects with Parkinson’s disease (PD). Their aim is to improve the objectiveness, quality, and quantity of monitoring the course of the disease through plantar pressure and movement data collection. This may, in turn, lead medical supervisors to better informed, AI-supported decisions and new insights about PD and its therapy. However, previous research has not sufficiently investigated algorithms to assist movement data analysis to lay the fundament, on which such goals can be accomplished. Thus, this paper presents a new methodology for machine learning based time-series classification using data collected by the ENVISIBLE ParKInSock system. ParKInSock is a multi-sensor system that collects plantar pressure data with 8 highly dynamic force sensing resistor (HD-FSR) cell sensors and linear acceleration and rotational velocity data with an inertial measurement unit (IMU) on each foot, generating a total of 28 data channels. These 28 signals are minimally processed, windowed, and used as features to train two machine learning algorithms: (1) the kernel-based RandOm Convolutional KErnel Transform (ROCKET) algorithm, and (2) a long short-term memory (LSTM) type recurrent neural network (RNN). These two models are trained to perform a supervised learning classification task with classes for activities of daily living (ADL). In a first trial, using cross-validation, the mean accuracy of the ROCKET and LSTM model on the test dataset were 99.61% ± 0.26% and 99.52% ± 0.22%, respectively. The dataset consisted of 4.4 h of five different ADL of one individual subject. In a second trial, six different ADL from three individual subjects created ~1 h of data. The average accuracy results on this dataset were 94.93% ± 1.49% and 95.71% ± 1.45% for the ROCKET and LTSM model, respectively.

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Deep Learning for Human Activity Recognition with Plantar Pressure and Movement Data from an Instrumented Smart Sock System

  • 章节

  • 出版模式：

    开放获取

  • 发表日期：

    2026年

查看全文
  • Noah Zuchna,
  • Bernd Resch,
  • Stephan Odenwald

摘要

Human activity recognition (HAR) is a task which, if solved by machine intelligence, has a vast potential to improve technological advances in many different fields, medical engineering being one of the more prominent ones. Recent advances in the development of wearable sensor technologies, mobile computing, cloud computing and machine learning (ML) methods made cutting-edge technology broadly available to the scientific community. One such medical application is the therapy and monitoring optimization of subjects with Parkinson’s disease (PD). Their aim is to improve the objectiveness, quality, and quantity of monitoring the course of the disease through plantar pressure and movement data collection. This may, in turn, lead medical supervisors to better informed, AI-supported decisions and new insights about PD and its therapy. However, previous research has not sufficiently investigated algorithms to assist movement data analysis to lay the fundament, on which such goals can be accomplished. Thus, this paper presents a new methodology for machine learning based time-series classification using data collected by the ENVISIBLE ParKInSock system. ParKInSock is a multi-sensor system that collects plantar pressure data with 8 highly dynamic force sensing resistor (HD-FSR) cell sensors and linear acceleration and rotational velocity data with an inertial measurement unit (IMU) on each foot, generating a total of 28 data channels. These 28 signals are minimally processed, windowed, and used as features to train two machine learning algorithms: (1) the kernel-based RandOm Convolutional KErnel Transform (ROCKET) algorithm, and (2) a long short-term memory (LSTM) type recurrent neural network (RNN). These two models are trained to perform a supervised learning classification task with classes for activities of daily living (ADL). In a first trial, using cross-validation, the mean accuracy of the ROCKET and LSTM model on the test dataset were 99.61% ± 0.26% and 99.52% ± 0.22%, respectively. The dataset consisted of 4.4 h of five different ADL of one individual subject. In a second trial, six different ADL from three individual subjects created ~1 h of data. The average accuracy results on this dataset were 94.93% ± 1.49% and 95.71% ± 1.45% for the ROCKET and LTSM model, respectively.