A Lightweight Deep Residual Network for Rehabilitation Activity Recognition in Heterogeneous Pediatric Populations

Simone Costantini, Benedetta Giachetti, Fabio Alexander Storm, Emilia Biffi, Elena Mugellini, Anna Maria Bianchi, Giuseppe Andreoni
Proceedings of the Fourth Swiss AI Days, PMLR 309:13-26, 2026.

Abstract

Human Activity Recognition based on wearable Inertial Measurement Units (IMUs) has emerged as a promising technology for the automated quantification of rehabilitation dosage and rehabilitation activity identification. However, existing solutions rely on multi-sensor configurations limiting clinical usability or fail to generalize to populations with heterogeneous motor functions. This study developed a lightweight residual network with self-attention mechanisms for classifying different phases of the rehabilitation activities (Rest, Balance, Walk) using data from a single IMU placed at the lower back, and collected from a pediatric cohort including 10 neurotypical children (mean age: years, 5 females) and 8 patients with neuromotor disorders as a consequence of cerebral palsy or acquired brain injury (mean age: years, 4 females). A preliminary ablation study across different IMU channel combinations revealed that combining accelerometer, gyroscope, and magnetometer signals allowed the model to achieve the best performance, with the magnetometer providing a key contribution for better discriminating between low-dynamic activities (Rest and Balance). Based on the optimal channel configuration identified in the ablation study, a Leave-One-Subject-Out cross-validation framework proved the model generalization abilities across heterogeneous motor functional domains, achieving an average macro F1-score of 0.81. These results confirm that the proposed framework provides an ecological and reliable tool for the objective recognition and quantification of rehabilitation activity in a clinical context.

Cite this Paper

BibTeX
@InProceedings{pmlr-v309-costantini26a, title = {A Lightweight Deep Residual Network for Rehabilitation Activity Recognition in Heterogeneous Pediatric Populations}, author = {Costantini, Simone and Giachetti, Benedetta and Storm, Fabio Alexander and Biffi, Emilia and Mugellini, Elena and Bianchi, Anna Maria and Andreoni, Giuseppe}, booktitle = {Proceedings of the Fourth Swiss AI Days}, pages = {13--26}, year = {2026}, editor = {Kucharavy, Andrei and Delgado, Pamela and Schürch Todeschini, Valérie and Rumley, Sébastien}, volume = {309}, series = {Proceedings of Machine Learning Research}, month = {23--25 Mar}, publisher = {PMLR}, pdf = {https://raw.githubusercontent.com/mlresearch/v309/main/assets/costantini26a/costantini26a.pdf}, url = {https://proceedings.mlr.press/v309/costantini26a.html}, abstract = {Human Activity Recognition based on wearable Inertial Measurement Units (IMUs) has emerged as a promising technology for the automated quantification of rehabilitation dosage and rehabilitation activity identification. However, existing solutions rely on multi-sensor configurations limiting clinical usability or fail to generalize to populations with heterogeneous motor functions. This study developed a lightweight residual network with self-attention mechanisms for classifying different phases of the rehabilitation activities (Rest, Balance, Walk) using data from a single IMU placed at the lower back, and collected from a pediatric cohort including 10 neurotypical children (mean age: years, 5 females) and 8 patients with neuromotor disorders as a consequence of cerebral palsy or acquired brain injury (mean age: years, 4 females). A preliminary ablation study across different IMU channel combinations revealed that combining accelerometer, gyroscope, and magnetometer signals allowed the model to achieve the best performance, with the magnetometer providing a key contribution for better discriminating between low-dynamic activities (Rest and Balance). Based on the optimal channel configuration identified in the ablation study, a Leave-One-Subject-Out cross-validation framework proved the model generalization abilities across heterogeneous motor functional domains, achieving an average macro F1-score of 0.81. These results confirm that the proposed framework provides an ecological and reliable tool for the objective recognition and quantification of rehabilitation activity in a clinical context.} }
Endnote
%0 Conference Paper %T A Lightweight Deep Residual Network for Rehabilitation Activity Recognition in Heterogeneous Pediatric Populations %A Simone Costantini %A Benedetta Giachetti %A Fabio Alexander Storm %A Emilia Biffi %A Elena Mugellini %A Anna Maria Bianchi %A Giuseppe Andreoni %B Proceedings of the Fourth Swiss AI Days %C Proceedings of Machine Learning Research %D 2026 %E Andrei Kucharavy %E Pamela Delgado %E Valérie Schürch Todeschini %E Sébastien Rumley %F pmlr-v309-costantini26a %I PMLR %P 13--26 %U https://proceedings.mlr.press/v309/costantini26a.html %V 309 %X Human Activity Recognition based on wearable Inertial Measurement Units (IMUs) has emerged as a promising technology for the automated quantification of rehabilitation dosage and rehabilitation activity identification. However, existing solutions rely on multi-sensor configurations limiting clinical usability or fail to generalize to populations with heterogeneous motor functions. This study developed a lightweight residual network with self-attention mechanisms for classifying different phases of the rehabilitation activities (Rest, Balance, Walk) using data from a single IMU placed at the lower back, and collected from a pediatric cohort including 10 neurotypical children (mean age: years, 5 females) and 8 patients with neuromotor disorders as a consequence of cerebral palsy or acquired brain injury (mean age: years, 4 females). A preliminary ablation study across different IMU channel combinations revealed that combining accelerometer, gyroscope, and magnetometer signals allowed the model to achieve the best performance, with the magnetometer providing a key contribution for better discriminating between low-dynamic activities (Rest and Balance). Based on the optimal channel configuration identified in the ablation study, a Leave-One-Subject-Out cross-validation framework proved the model generalization abilities across heterogeneous motor functional domains, achieving an average macro F1-score of 0.81. These results confirm that the proposed framework provides an ecological and reliable tool for the objective recognition and quantification of rehabilitation activity in a clinical context.
APA
Costantini, S., Giachetti, B., Storm, F.A., Biffi, E., Mugellini, E., Bianchi, A.M. & Andreoni, G.. (2026). A Lightweight Deep Residual Network for Rehabilitation Activity Recognition in Heterogeneous Pediatric Populations. Proceedings of the Fourth Swiss AI Days, in Proceedings of Machine Learning Research 309:13-26 Available from https://proceedings.mlr.press/v309/costantini26a.html.

Related Material