.jpg)
In the field of physical rehabilitation, interpreting data from inertial sensors (IMU) is essential for specialists to understand a patient's progress. However, multivariate time series classification models often provide explanations that are difficult to translate into clinical practice, as they operate at the level of individual channels rather than grouping them according to biomechanical criteria such as muscle groups or joint segments. To overcome this limitation, approaches such as adaptive group-based counterfactual explanations have been developed, a methodology that allows generating modifications consistent with clinical reasoning.
The proposal is structured in two phases: first, a group ranking based on Shapley-Adaptive (SA) is used, which maintains the validity of the counterfactual but does not guarantee group sparsity. To solve this, learnable gate mechanisms (Learnable Gate, LG) are introduced, integrating trainable relevance weights together with perturbation masks. Experiments conducted with the KneE-PAD dataset, focused on knee rehabilitation exercises, demonstrate that this technique substantially improves modal group sparsity compared to the M-CELS baseline, while maintaining or even improving validity, temporal smoothness, and generation efficiency.
These advances are especially relevant for companies developing custom applications in the healthcare sector. The ability to provide interpretable explanations not only improves specialists' trust in AI systems but also allows for adjusting personalized treatments. In fact, the integration of AI for businesses in rehabilitation settings can be enhanced with techniques like these, where AI agents act as assistants in clinical decision-making.
From a technical perspective, implementing this type of solution requires robust infrastructures. Q2BSTUDIO, as a software development company, offers custom software that adapts to the specific needs of each project, whether in the field of rehabilitation or other sectors. Additionally, the use of AWS and Azure cloud services facilitates the processing of large volumes of sensor data, while business intelligence tools such as Power BI allow visualizing recovery patterns. Cybersecurity is another fundamental pillar when handling sensitive patient information, so audits and penetration testing are part of the development cycle.
In conclusion, group-based counterfactual explanations represent a step forward in the interpretability of machine learning models applied to movement time series. By aligning explanations with the language of rehabilitation professionals, a more direct link between technology and clinical practice is achieved. This opens the door to new forms of personalization and feedback that, supported by AI platforms for businesses and AI agents, can transform computer-assisted rehabilitation.
