💻 computer science

Research on intelligent assessment and exercise risk prediction model of knee joint function in elderly patients

This paper proposes a multi-modal computational framework that fuses gait video, inertial signals, plantar pressure, and knee flexion data to achieve high-accuracy intelligent assessment of knee function and exercise risk prediction for elderly patients, demonstrating superior stability and robustness compared to existing baseline models.

Original authors: Xinghai Yang, Xiaoyan Liu, Ye Li, Xiaolu Zhang

Published 2026-06-30

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Xinghai Yang, Xiaoyan Liu, Ye Li, Xiaolu Zhang

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine your knee is like a complex, high-performance car engine. For years, doctors have checked this engine by looking at it once, asking the driver how it feels, and maybe taking a quick snapshot. But what if the engine only sputters when you turn a corner or go up a hill? A single snapshot misses the whole story.

This paper introduces a new "super mechanic" for elderly patients' knees. Instead of just looking at one thing, this system watches the knee from four different angles at the exact same time, creating a complete, 3D movie of how the joint actually works.

Here is how the system works, broken down into simple parts:

1. The Four "Eyes" of the System

Most old methods relied on just one source of information, like a single security camera. This new model uses four different "sensors" that work together like a team of detectives:

The Video Camera: It watches the patient walk, tracking the path of their legs and where their weight shifts (like watching a dancer's footwork).
The Pressure Mat: It acts like a sensitive floor that feels exactly how hard the patient is pressing down on their toes versus their heels.
The Motion Sensors (IMU): These are like tiny accelerometers (similar to what's in your phone) that feel the shake, speed, and rotation of the leg bones.
The Angle Tracker: It measures exactly how much the knee bends and straightens, like a protractor for the joint.

2. The "Translator" (Multimodal Fusion)

The problem with having four different sensors is that they all speak different languages. The video speaks in "pictures," the pressure mat in "numbers," and the sensors in "vibrations."

The paper's model has a special Translator in the middle. It takes all these different signals and forces them to speak the same language. It also acts like a smart editor: if the camera is blurry or the sensor slips, the Translator knows to trust the other sensors more and ignore the bad data. This ensures the final picture isn't ruined by one glitchy sensor.

3. The "Time-Traveler" (Temporal Modeling)

Knee problems often happen in a sequence, not just at one frozen moment. A person might walk fine for three steps, then stumble on the fourth.

Old models looked at the knee like a series of still photos. This new model looks at it like a movie. It watches the "story" of the movement over time. It can spot patterns like, "Oh, the patient always leans to the left right after they stand up." By understanding the flow of time, it can predict if a movement is about to go wrong before the patient even realizes it.

4. The "Personalized Risk Meter"

Finally, the system doesn't just say "Safe" or "Danger." It calculates a Risk Score.

Think of this like a weather forecast. A "storm" (high risk) for a young, athletic person might be a "light drizzle" for an elderly person with a weak knee. This model creates a personalized baseline for each patient. It learns what is "normal" for that specific person and then alerts them if their knee starts acting up compared to their own usual pattern, rather than comparing them to a generic standard.

The Results: How Good is it?

The researchers tested this "super mechanic" on 186 elderly patients with 1,240 different movement samples. Here is what they found:

Accuracy: It predicted movement risks with 91.3% accuracy.
Precision: When it said a risk was present, it was right 92.1% of the time.
Comparison: It performed significantly better than older methods (like standard computer vision or simple machine learning), which often get confused by the messy, real-world movements of elderly people.

The Bottom Line

This paper doesn't claim to cure knees or replace doctors. Instead, it offers a digital tool that combines video, pressure, motion, and angles into one smart system. It helps doctors see the "whole movie" of a patient's knee function rather than just a single frame, allowing them to spot instability and predict risks much earlier and more accurately than before.

The authors note that while the system works great in controlled hospital settings, it still needs more testing in messy, real-world home environments to ensure it works perfectly everywhere.

Problem Statement

As the global population ages, knee joint functional degeneration significantly impacts the independent mobility of the elderly. While traditional clinical assessments rely on scale scoring, manual observation, and local index comparisons, these methods struggle to stably describe joint stability, load transfer characteristics, and risk changes during continuous movements. Existing research has developed parallel paths for visual analysis, wearable sensing, and machine learning, but they often suffer from inconsistent sampling granularity, insufficient temporal correlation, and weak unified representation. Consequently, knee function assessment and risk prediction are frequently conducted separately, failing to form an integrated computational chain that can handle the dynamic evolution of knee functional status from heterogeneous data sources.

Methodology

The authors propose a unified computational framework that integrates multi-modal feature fusion, temporal state representation, and probabilistic risk inference. The system is built upon a dataset of 1,240 evaluation samples from 186 elderly patients, utilizing four synchronized data modalities: gait video, inertial measurement unit (IMU) signals, plantar pressure sequences, and knee flexion/extension angle measurements.

The framework consists of three core components:

Intelligent Assessment Model (Dual-Task Structure):
- Input Organization: Data is sliced into fixed-length windows (50 frames) with overlapping steps.
- Feature Extraction: Four parallel branches extract specific features: visual topology and trajectory (video), load distribution (pressure), angular velocity and rotation (IMU), and range of motion (angles).
- Dual-Task Output: The model simultaneously outputs a continuous functional score (for tracking subtle rehabilitation changes) and discrete functional status grades (for screening and stratification).
- Architecture: A backbone network using 3D convolutional layers and bidirectional Gated Recurrent Units (GRU) aggregates local dynamics and temporal trends into a shared embedding space.
Multimodal Feature Fusion:
- Normalization & Projection: Single-modal features undergo parametric normalization and linear projection to align dimensions.
- Reliability Weighting: A modal reliability mechanism calculates weights ( $\alpha_m$ ) based on both feature content and data quality (sampling completeness, smoothness, missing rate). This prevents noisy modalities (e.g., occluded video or drifting sensors) from dominating the fusion.
- Structural Priors: The fusion layer incorporates structural priors related to knee mechanics (e.g., load redistribution, lateral displacement) to explicitly strengthen key patterns.
- Consistency Loss: A cross-modal consistency loss ensures that sample relationships (similarity matrices) remain coordinated across different modalities.
Temporal Representation & Risk Prediction:
- State Trajectory: Overlapping windows are processed through a gated state update mechanism to balance short-time fluctuations with historical state retention.
- Stage-Aware Attention: An attention mechanism weights specific action phases (e.g., support transition, peak flexion) more heavily than others.
- Stability Index: A unified stability index ( $\kappa_t$ ) is derived from angle, pressure, and inertia changes to quantify joint instability.
- Probabilistic Risk Inference: The system employs a Bayesian risk modification framework. It calculates a base risk probability and then applies an individualized risk threshold ( $\tau_i$ ) based on the patient's specific baseline (average stability, range of motion, and gait speed). This corrects for individual differences, preventing low-mobility patients from being misclassified as high-risk due to uniform boundaries.
- Loss Function: Training utilizes a comprehensive loss combining cross-entropy, ranking loss, and a margin constraint to ensure clear separation between high-risk and low-risk samples.

Key Contributions

Unified Framework: The study constructs an integrated model that fuses video, pressure, inertia, and angle data into a single state space, moving beyond single-source observations.
Dynamic Temporal Modeling: Unlike static classification, the model emphasizes the dynamic evolution of knee function, capturing continuous action clips and identifying potential dangerous patterns through time-series state modeling.
Individualized Risk Calibration: The introduction of an individual threshold correction mechanism allows the model to adapt to patients with varying baseline mobility and rehabilitation stages, reducing false positives in low-mobility populations.
Dual-Task Output: The model simultaneously provides continuous functional scores for tracking progress and discrete risk grades for clinical decision-making.

Experimental Results

The model was evaluated against CNN-LSTM, Random Forest, and unimodal baseline models using a test set derived from 186 patients.

Functional Assessment: The proposed model achieved a Mean Absolute Error (MAE) of 4.9% for functional score estimation, significantly outperforming the CNN-LSTM (7.6%), Random Forest (9.1%), and video-only (10.4%) models.
Risk Prediction: The model attained an accuracy of 91.3%, with a precision of 92.1%, recall of 89.4%, and an F1-score of 90.7%.
Ablation Studies: Removing the multimodal fusion component increased the functional score MAE to 6.3% and reduced risk accuracy to 86.9%. Removing temporal representation decreased the F1-score by 4.4 percentage points.
Robustness: The model demonstrated superior stability across different cross-validation folds compared to control groups, with tighter clustering in error/accuracy space.
Statistical Significance: Paired t-tests confirmed that the performance improvements over all control methods were statistically significant ( $p < 0.01$ ).

Significance and Claims

The paper claims that this framework provides a feasible technical support system for digital rehabilitation assessment and risk early warning in intelligent medical scenarios. By integrating multi-modal data and temporal modeling, the system can:

Provide stable, continuous, and interpretable computational outputs for knee function assessment.
Support three distinct application scenarios: outpatient screening (rapid identification), rehabilitation follow-up (tracking change trends), and home monitoring (low-cost real-time inference).
Offer a deployable basis for assistant decision-making in intelligent medical environments.

The authors acknowledge limitations, noting that the current data is from a single center and focuses on standard assessment tasks, which may limit generalization to complex home environments. They suggest future work should focus on multi-center data expansion, edge computing deployment, and self-supervised pre-training for long-term individual trajectory modeling.