Abstract
Personalized weight-bearing rehabilitation is crucial for orthopedic lower limb postoperative patients with different health backgrounds. However, physical therapists are in significant shortage, and the effectiveness of one-size-fits-all rehabilitation plans varies greatly. Here, using full-chain data from hospitalization to rehabilitation of over 1,200 patients across 25 medical institutions, a comprehensive AI-driven rehabilitation system was proposed as a Virtual Physical Therapist (VPT). The system achieves automatic formulation of overall rehabilitation plans at discharge (Multimodal Rehabilitation Schedule Network) and real-time dynamic interventions during the rehabilitation process (InsFormer). Multi-center validation showed that over 95% of patients achieved reasonable weight-bearing training in the appropriate rehabilitation period. Mean absolute error (MAE) of recovery time prediction is 0.8 week. Specifically, it showed superior rehabilitation outcomes for some patients with underlying diseases or the elderly, surpassing the diagnostic level of senior physicians. This generative AI approach adapts to changes in patients’ physical function and specific rehabilitation feedback, shortening recovery time, improving rehabilitation outcomes, and opening up new pathways for intelligent rehabilitation paradigms.
Introduction
Postoperative rehabilitation for lower limb orthopedic surgeries is a high-stakes balancing act. The timing and intensity of weight-bearing exercises are critical: excessive weight too early risks re-fracture or implant failure, while insufficient weight leads to poor bone healing and muscle atrophy. Despite its importance, rehabilitation often relies on "one-size-fits-all" protocols due to a global shortage of physical therapists and a lack of clinical consensus on how to tailor plans to individual factors like bone quality or comorbidities.
To address this, researchers have developed the "Virtual Physical Therapist" (VPT), an AI-driven system designed to provide personalized, data-backed guidance from the hospital bed to the home.
The Virtual Physical Therapist (VPT) Framework
The VPT system operates on a "full-chain" data model, integrating clinical records, medical imaging, and real-time sensor data to manage the entire recovery lifecycle.
The architecture is divided into two primary stages: initial planning at discharge and dynamic intervention during home recovery.
Methodology: A Two-Stage AI Approach
Stage 1: Multimodal Rehabilitation Schedule Network (MRSN)
At the point of discharge, the MRSN utilizes a TensorFusion Network (TFN) to synthesize three distinct data streams:
- Clinical Baseline Data: Patient demographics such as age and weight.
- Postoperative X-rays: Analyzed via a ResNet-18 architecture to assess surgical outcomes and bone alignment.
- Biochemical Indicators: Laboratory results, including hemoglobin levels.
By using TensorFusion rather than simple feature concatenation, the model captures complex interactions between these modalities. This stage predicts the patient's initial weight-bearing load and the total estimated recovery duration.
Stage 2: InsFormer (Dynamic Intervention)
Once the patient transitions to home rehabilitation, the system employs InsFormer, a Transformer-based multivariate time series model. This model processes daily data from a custom-built smart walking aid equipped with foot pressure sensors. By analyzing training scores, pain levels (via the VAS scale), and any adverse reactions, InsFormer "course-corrects" the rehabilitation plan weekly, mimicking the adaptive decision-making of a senior physician.
Multi-Center Validation and Results
The study was validated using a massive multi-center dataset from 25 medical institutions, involving over 1,200 patients across 14 different types of lower limb surgeries (e.g., hip osteotomies, ankle fractures, and total joint arthroplasties).
The system demonstrated exceptional predictive accuracy:
- MRSN Performance: Achieved a Pearson correlation of 0.944 for initial load and 0.957 for recovery time. The Mean Absolute Error (MAE) for recovery time was just 0.8 weeks.
- InsFormer Precision: Outperformed standard time-series models (like AutoFormer) with a weight-bearing prediction MAE of only 0.8 kg.
- Clinical Success: Over 95% of patients in the validation cohort maintained weight-bearing training within the medically appropriate window.
Key Insights: The Role of Biochemical Markers and Complex Cases
A significant finding was the predictive power of biochemical markers. Specifically, hemoglobin levels emerged as the strongest predictor of recovery time. Lower postoperative hemoglobin often indicates greater surgical trauma and blood loss, signaling the AI to recommend a more conservative weight-bearing progression.
Furthermore, the VPT system proved particularly effective for "complex" patients, such as the elderly or those with underlying health conditions. In these instances, the AI's ability to detect subtle physiological signals allowed it to outperform the diagnostic accuracy of senior physicians, who may be more inclined to follow generalized, less-optimized protocols.
Conclusion and Future Directions
The Virtual Physical Therapist represents a shift toward intelligent, personalized rehabilitation paradigms. By shortening recovery times and improving outcomes through data-driven precision, it offers a scalable solution to the therapist shortage.
Future research will focus on expanding the dataset to include rarer orthopedic conditions and exploring the underlying biological mechanisms of recovery. Additionally, the researchers aim to generalize the VPT framework for neurorehabilitation, such as stroke recovery, using incremental pre-training methods for specialized wearable devices.
Summary updated