TLDR: This research explores how social robots can autonomously guide physiotherapy exercises, adapting to individual patient needs. Facing a scarcity of patient data, the study used 33 expert healthcare practitioners as patient proxies to develop a behavior model. This model, combined with real-time feedback from wearable sensors (IMUs) for repetition counting and an AI language model (GPT-4) for perceived exertion from verbal responses, informed a reinforcement learning policy. The trained robot policy demonstrated the ability to personalize exercise instructions, adjusting repetition counts based on a patient’s exertion tolerance and fluctuating performance across different exercise plans in simulation, laying the groundwork for more effective robot-led rehabilitation.
In the evolving landscape of healthcare, social robots are emerging as a promising solution to address critical challenges in patient rehabilitation, particularly in physiotherapy. A recent research paper, “Policy Learning for Social Robot-Led Physiotherapy”, delves into how these intelligent machines can autonomously guide patients through exercise sessions, offering personalized support that adapts to individual needs and fluctuating performance.
Adherence to physical exercise programs is vital for individuals recovering from injuries like fractures or strokes. However, healthcare workforce shortages often mean patients face long waits for in-person assessments, relying instead on generic exercise guides. While technology has introduced new solutions, many still require constant human oversight, increasing staff workload. This highlights a clear need for autonomous, accessible, and personalized rehabilitation support.
Social robots have shown significant potential in this area, capable of delivering exercise sessions and adapting instructions based on real-time user feedback. Their physical presence can also enhance engagement beyond what screen-based agents offer. However, developing autonomous control policies for these robots is complex. Patients vary greatly in physical ability, and their performance can change daily, requiring the system to recognize and respond to both good and bad days.
Bridging the Data Gap with Expert Proxies
A major hurdle in developing robust policies for social robots in healthcare is the scarcity of patient behavior data. To overcome this, the researchers employed an innovative approach: they engaged 33 expert healthcare practitioners (physiotherapists, occupational therapists, etc.) as patient proxies. These experts interacted with the robot, role-playing various patient scenarios, providing invaluable data to inform a dynamic patient behavior model. This method, inspired by ‘patient simulation’ in medical training, allowed for the collection of diverse and representative behaviors without direct robot-patient interactions at an early stage, which could pose risks.
The robot prototype, built on the ARI platform by PAL Robotics, was designed to guide patients through standardized upper-limb exercises like arm abduction, flexion, and rotation. These exercises could be performed in three modes: assistive (with support), active (using only the affected limb), and resistive (with resistance bands), allowing for nine exercise variations to accommodate different recovery stages.
Sensing Performance and Perceived Exertion
To assess patient performance, two wearable Inertial Measurement Units (IMUs) were fitted to the affected arm to accurately count repetitions. For subjective feedback, the robot engaged patients in conversational dialogue, asking about their perceived exertion. OpenAI’s GPT-4 Large Language Model (LLM) was then used to process these verbal responses, translating them into numerical scores aligned with the Borg Rating of Perceived Exertion CR10 Scale. This LLM-based pipeline allowed the robot to estimate perceived exertion from natural language input, such as “That felt fine” or “That was a little difficult.”
Learning in Simulation: The Reinforcement Learning Approach
The collected expert data was crucial for developing a patient behavior model within a software simulation environment. This model could generate exercise performance metrics and subjective scores on perceived exertion. A reinforcement learning (RL) agent was then trained in this simulation. RL is well-suited for generating complex control policies through trial-and-error interactions, allowing the robot instructor agent to learn how to map session states to optimal instruction actions.
The RL agent’s goal was to adapt exercise instructions (specifically, the number of repetitions) in real-time based on patient-specific exertion thresholds, session-to-session variations in performance, and individual exercise plans. The reward function for the agent encouraged maximizing repetitions while ensuring the patient’s perceived exertion remained within a moderate, safe range, preventing overexertion or potential injury.
Adaptive and Personalized Guidance
The trained policy demonstrated promising results in simulation. It learned distinct approaches for different patient exertion tolerance groups (low, average, and high performers), pushing high performers with more repetitions while instructing low performers more conservatively. The policy also adapted to various performance patterns, such as a “linear increase” (patients improving) or a “struggling day” (patients performing poorly), quickly adjusting repetitions to match the patient’s current capability. Furthermore, it proved applicable to more customized exercise plans with varying repetition goals per exercise block.
This research introduces two significant advancements: an RL-based framework for adaptive instruction that dynamically adjusts to patient performance sensed through IMUs and naturalistic verbal feedback via an LLM, and the integration of expert knowledge into user behavior models, moving beyond generalized assumptions towards more effective real-world robot behavior.
Also Read:
- CoachMe: An AI System for Detailed Movement Correction in Sports
- Patient-Zero: Generating Realistic Virtual Patients Without Real Medical Records
Future Directions
While the findings are encouraging, the researchers acknowledge limitations. Perceived exertion alone can be subjective; future work will incorporate objective metrics like Range of Motion. The current policy also lacks awareness of patient performance prior to each session, suggesting the need for an initial assessment phase. Enhancements in LLM interpretation, generalizability of the patient model, and multi-session learning to track long-term patient progress are also planned. The ultimate goal is to validate this system through real-world patient evaluations, paving the way for truly autonomous and personalized physiotherapy support.
