AboutDownloadsDocumentsForumsIssuesNews

This site serves for dissemination of upper body motion data from stroke patients during rehabilitation. The data is captured using wearable inertial measurement units and cameras. The data is fully labeled by trained annotators. A link to the code for the deep learning model is also provided.


We present PrimSeq, a pipeline to classify and count functional motions trained in stroke rehabilitation. PrimSeq encompasses three main steps: (1) capture of upper body motion during rehabilitation with wearable inertial measurement units (IMUs), (2) generation of primitive sequences from IMU data with the trained deep learning model, and (3) tallying of primitives with a counting algorithm.

To build this approach, we have collected a large realistic dataset of rehabilitating stroke patients, labeled it, and developed a new deep learning method to process it.

Using a previously established functional motion taxonomy (Schambra et al., 2019), we identified the five following classes of SOPs: reach, reposition, transport, stabilize, and idle. A reach is a UE motion to move into contact with a target object; a reposition is a UE motion to move proximate to a target object; a transport is a UE motion to convey a target object; a stabilize is a minimal-motion to keep a target object still; and an idle is a minimal-motion to stand at the ready near target object. We have labeled the motion data using these primitive definitions.

We invite you to download the data and the code.

References:
- Parnandi, A., Kaku, A., Venkatesan, A., Pandit, N., Wirtanen, A., Rajamohan, H., Venkataramanan, K., Nilsen, D., Fernandez-Granda, C. and Schambra, H., 2021. PrimSeq: a deep learning-based pipeline to quantitate rehabilitation training. arXiv preprint arXiv:2112.11330. (https://arxiv.org/abs/2112.11330)

- Kaku, A., Liu, K., Parnandi, A., Rajamohan, H.R., Venkataramanan, K., Venkatesan, A., Wirtanen, A., Pandit, N., Schambra, H. and Fernandez-Granda, C., 2021. Sequence-to-Sequence Modeling for Action Identification at High Temporal Resolution. arXiv preprint arXiv:2111.02521. (https://arxiv.org/abs/2111.02521)

Acknowledgement:
This work was funded by the American Heart Association/Amazon Web Service postdoctoral fellowship 19AMTG35210398 (A.P.), NIH R01 LM013316 (C.F.G., H.S.), NIH K02 NS104207 (H.S.), NIH NCATS UL1TR001445 (H.S.), and NSF NRT-HDR 1922658 (A.K., C.F.G.)

Feedback