Enhanced human activity recognition in medical emergencies using a hybrid deep CNN and bi-directional LSTM model with wearable sensors

Abstract

Human activity recognition (HAR) is one of the most important segments of technology advancement in applications of smart devices, healthcare systems & fitness. HAR uses details from wearable sensors that capture the way human beings move or engage with their surrounding. Several researchers have thus presented different ways of modeling human motion, and some have been as follows: Many researchers have presented different methods of modeling human movements. Therefore, in this paper, we proposed the CNN BiLSTM model with undersampling to improve the recognition of human actions. The model is evaluated using state-of-the-art metrics, including accuracy, precision, recall, and F1-score, on two publicly available datasets: For instance, the MHEALTH and Actitracker. This will enable the team to attain test accuracies of up to 98.5% on the MHEALTH dataset. The proposed CNN-BiLSTM model outperforms the conventional deep learning methods, as reported in the Actitracker dataset, by about 5% improvement. HAR has many applications, one of which is used to keep vigil over elderly people who live alone to alert when one has fallen or when any strange movement is noticed which could be a sign that the individual is experiencing a medical Emergency. It can also be applied in physiotherapy, where the patient's development throughout rehabilitation exercises can be accessed.

Keywords: Bi-directional long short term memory; Convolutional neural networks; Human action recognition; Medical emergency; Wearable sensors.

Fig. 1

Architecture of an LSTM network.

Fig. 2

Architecture of a BiLSTM layer.

Fig. 3

Sample sensor data for the activity “Cycling” from the MHEALTH dataset.

Fig. 4

Sample sensor data for the activity “Walking” from the Actitracker dataset.

Fig. 5

Architecture of the proposed Deep CNN-BiLSTM.

Fig. 6

Accuracy and loss curves for (a) the Actitracker dataset and (b) the MHEALTH dataset.

Fig. 7

Actitracter sampling techniques.

Fig. 8

MHealth dataset for sampling techniques.

Fig. 9

ROC curves for (a) the Actitracker dataset and (b) the MHEALTH dataset.

Fig. 10

Confusion matrices (a) Actitracker dataset and (b) MHEALTH dataset.

Fig. 11

Effect of optimizers on validation loss (a) Actitracker dataset and (b) MHEALTH dataset.

Fig. 12

Mitigate the high computational requirements.