The research enlisted 92 adults diagnosed clinically as patients with KOA, exhibiting varying degrees of severity as evaluated by the Kellgren and Lawrence classification. These participants demonstrated independent walking capability for a duration of 2 minutes. All participants were sourced from Zhujiang Hospital of Southern Medical University, and the tests were administered under the guidance of proficiently trained medical personnel.

Participants were thoroughly briefed on the procedures and paradigm of the functional tests before they underwent the assessments. Adequate intervals were implemented between each test session to avoid the impact of fatigue. Successively, the participants underwent the TUGT and 40mFPWT to respectively evaluate the patients’ overall functional mobility, balance capacity, short-distance walking performance, and gait speed. The TUGT and 40mFPWT scores indicate the time taken to complete the test, with higher scores representing worse patient function. Table 2 summarizes the participants’ demographic characteristics and physical function tests. The results of the Mann-Whitney U tests revealed no statistical difference between male and female groups in the 40mFPWT and TUGT outcomes (P=.48 and P=.50, respectively).

The footwear system used for the collection of plantar pressure data has been detailed in prior studies [30]. Each shoe is equipped with eight integrated pressure sensors capable of detecting vertical ground reaction forces during walking. The sensor position distributions and corresponding pressure-sensing regions are illustrated in Figure 1A. Before formal data collection, participants were required to wear suitable socks and shoes. Ample time was provided for participants to adjust and ensure proper fitting of the shoes before engaging in natural walking. The participants were then asked to walk independently back and forth in a 20 m corridor at a freely walking speed for 2 minutes to simulate everyday locomotor activities. The shoe system collected plantar pressure signal data during walking at a frequency of 20 Hz and transmitted it to a mobile phone in real time via Bluetooth (Bluetooth Special Interest Group; Figure 1B).

After stepwise feature selection, a subset of features most relevant to the current problem is identified for developing machine learning models. The models considered include linear regression (LR), support vector machine (SVM), random forest (RF), Adaptive Boosting (AdaBoost), Extreme Gradient Boosting (XGBoost), and Light Gradient Boosting Machine (LGBM). Hyperparametric tuning of each model was performed using Optuna (Preferred Networks, Inc) [44] and 5-fold cross-validation was used to evaluate the results.

Based on the training results of each model and referring to previous studies [45], the 2-level ensemble learning model was constructed using the stacking regression method [46], as shown in Figure 4A. The training data were randomly partitioned into 5 mutually exclusive subsets, 4 of which were used as 5-fold cross-validation of the inner loop to train the models. At the first level, 4 decision-tree-based regressors including RF, AdaBoost, XGBoost, and LGBM were used to fit the training folds, respectively, and then predict the validation fold. The predictions from these 5 rounds were stacked to form the input features for the second-level regressor, which uses a simple model such as LR or SVM. The validation set of the external loop is used to evaluate the performance of the stacking model and hyperparameter tuning. Additionally, for comparison purposes, simple average ensemble (SAE) and weighted average ensemble (WAE) were also adopted to construct 2-level models for these 4 tree-based models. The final prediction of the SAE model is obtained by taking the average of the predictions from all individual models, while WAE assigns different weights to the predictions of each model according to their performance, allowing models with better performance to have a greater influence on the final prediction, as shown in Figure 4B. The ensemble model can reduce variance, enhance robustness, and improve generalizability by combining the prediction results from multiple models.

Mean absolute error (MAE), mean absolute percentage error (MAPE), and root mean squared error (RMSE) were used as regression performance metrics to evaluate the results of different models [47,48]. To avoid deceiving performance caused by data bias, the average of the known training set labels is used as the prediction of the unknown test set to calculate these metrics as the performance of the baseline model. The improvement of each metric for each model relative to the baseline model is calculated and normalized into a regress relative index (RI) to comprehensively evaluate the model performance [49,50], as shown in equation 11.

Where model_i and baseline_i represent the performance metrics corresponding to each model and the baseline model, respectively.

Models with high RI values will be considered as candidate regressors for 2-level ensemble learning models. The contribution of each model in WAE prediction was weighted according to the proportion of their RI value. The final predicted value ŷ_WAE of the WAE model can be calculated by equation 12.

All experimental procedures were approved by the institutional review board of Zhujiang Hospital of Southern Medical University (IRB 2019-KY-016-02). This study ensures informed consent with the right to withdraw. Participants’ privacy is safeguarded through data anonymization. Compensation for human subjects involved a payment of CN ¥200 (CN ¥1=US $0.14) per individual.

Table 4 shows the benchmark experimental results of feature selection, comparing the performance of various algorithms using the RF regressor as the base model, and it can be seen that the method using stepwise feature selection performs better than the principal component analysis algorithm in both tasks.

After applying the Spearman correlation coefficient to filter out low-correlation noise features, the 40mFPWT task retained 161 features and the TUGT task retained 131 features. Then select an optimal subset of 61 features with mRMR, which was set to not exceed the number of training samples. For 40mFPWT, the subset of 11 features identified by the Sequential Backward Floating Selection method yields the best performance. For TUGT, the optimal feature subset determined by the Sequential Forward Floating Selection method consists of 10 features. The optimized feature subsets and Spearman correlation coefficients for both tasks are depicted in Figure 5.

After feature selection, six regression models were trained and their hyperparameters were tuned using Optuna with 5-fold cross-validation for both the 40mFPWT and TUGT tasks, respectively. RF, AdaBoost, XGBoost, and LGBM were used as first-level regressors for the stacked model, and the performance of LR and SVM as second-level regressors were compared, respectively. SAE combines the predictions from 4 trained tree-based models equally, while WAE weighs the contribution of each ensemble member proportionally based on RI value. The performance of each model compared with the baseline model is shown in Table 5 for 40mFPWT and Table 6 for TUGT. Among individual models, the LGBM model achieved the best predictive performance in the 40mFPWT task, while the XGBoost model performed best in the TUGT task. The performance of the stacking model using SVM as the second-level regressor is better than that using LR in both tasks. The average integration model using the WAE strategy has better performance than that using SAE in both tasks.

The best-performing trained models obtained from optimizing hyperparameters on the training set were used to generate predictions for the holdout test set, including the individual model, stacked model, and average ensemble model. The results in the holdout test set for the 40mFPWT are shown in Table 7 and for the TUGT are shown in Table 8. Among all models, the WAE model using 4 tree-based models has the best generalization performance in the test set, with MAE of 2.686 seconds, MAPE of 9.602%, and RMSE of 3.316 seconds for the prediction of 40mFPWT, and for TUGT with MAE of 1.280 seconds, MAPE of 12.389%, and RMSE of 1.905 seconds.

This study developed a functional evaluation model using multidimensional plantar pressure features to predict the functional performance of patients with KOA. The plantar pressure data collected by the shoe system were preprocessed through feature engineering. These features were then input into the trained model to enable prediction and thereby realize functional assessment and monitoring of patients with KOA.

The results of feature selection indicated that age was the most relevant predictor of functional performance on the 2 tasks. It is reasonable that higher age correlated with a longer duration to finish the function tests, and therefore poorer function (Figure 5). Notably, 4 features showed negative correlations with 40mFPWT outcomes (Figure 5A). Specifically, higher values of the L_7-8PPP feature were associated with better functionality, aligned with previous findings that individuals with KOA tend to exhibit diminished plantar pressure in the heel region [25,26]. Four SI features were selected for the TUGT task. This is likely because diseases or impairments that impact proprioception or postural stability could thereby influence balance performance on the TUGT by altering one’s symmetry [51]. The results showed that plantar pressure–derived features in the forefoot and rearfoot regions, specifically PG and PPP values, exhibited relatively strong correlations with functional test outcomes. In addition, COP-derived features, SI features, and weak foot features were selected and exhibited relatively strong correlations with functional test outcomes. These observed relationships are biologically plausible and concordant with existing understandings of pathological gait patterns in populations with KOA.

The performance of the baseline model, which uses the mean of known labels in the training set as predictions for the test set, provides an unbiased evaluation metric without issues of overfitting or overoptimism. The model-free nature of this design ensures a fair assessment of predictive gains attributable to model architecture rather than data traits. The RI value of each model was obtained by calculating the sum of the improvement of each performance metric relative to the baseline. A higher RI value represents better overall model performance. The results demonstrate that tree-based models outperform LR and SVM models significantly (Tables 3 and 4). Using SVM as the second-level regressor in the stacked model yields better results than LR, primarily attributable to SVM’s ability to handle nonlinear relationships and demonstrate robustness against outliers in the data. Within the averaged ensemble framework, the WAE model outperformed the SAE model, possibly due to its weighted aggregation mechanism. The WAE model produced superior outcomes by assigning higher weights to individual models with higher predictive power, thereby generating more refined ensemble predictions.

Due to feature selection being performed on the training set, the models may still overfit the training set even with the use of cross-validation techniques, resulting in overly optimistic performance estimates. Therefore, in general, the model performs better on a training set than on a holdout test set containing unknown samples. The performance on the test set provides a more robust evaluation of the models’ generalization capability when encountering new data. The results indicate that the WAE model demonstrated the best generalization performance in both tasks, rather than the stacked model that used SVM as the second-level regressor (Tables 5 and 6). This aligns with the principle of Occam razor in model selection which is to prefer the more parsimonious model when performance is otherwise comparable [52].

Based on the prediction results from the 2 functional tests, the features we extracted appear to have a reasonably close correlation with functional performance. The models demonstrated good generalization for predicting traditional clinical function tests. In the future, the model could be integrated into a terminal application to longitudinally monitor patients’ functional status. The identified plantar pressure features could serve as an evaluation tool to guide the rehabilitation and assessment of progress for patients with KOA, offering clear advantages per time efficiency, longitudinal documentation, and accuracy compared to conventional functional tests.

The findings of this study have significant clinical implications for the management and rehabilitation of patients with KOA. The proposed techniques enable continuous, real-time monitoring of patients’ functional status beyond clinical settings. This capability facilitates more personalized and timely interventions. By accurately assessing functional performance, patients gain greater insight into their condition thereby improving overall management. Additionally, the system’s ability to decrease the frequency of hospital visits and extensive clinical assessments contributes to cost-effectiveness, alleviating the burden on health care systems.

Several limitations of this study should be noted. First, this study’s cohort consisted exclusively of patients diagnosed with KOA without the inclusion of data from healthy control participants for comparative analysis. The validity of the model prediction is limited to the patients’ population with KOA. The participant data of this study are mostly female, and the validity of the model may be biased toward female patients. Second, this was a cross-sectional study without long-term longitudinal monitoring of patients in their daily living environments. Third, it only conducted a feasibility study on functional estimation based on wearable plantar pressure features for 2 clinical functional tests recommended by OARSI for KOA. Hence, the proposed techniques require further validation in larger prospective cohorts and preferably multicenter trials to corroborate generalizability.

Despite its limitations, we believe that this wearable plantar pressure technique captures objective quantitative indicators of functional status and has great application value. The preliminary findings indicate this methodology holds promise for enabling remote, quantitative monitoring of rehabilitation progress over time. Further work will refine the system for broader clinical application and validation.

This study aims to develop a lower extremity motor function evaluation model for patients with KOA based on multidimensional gait features, which was suitable for a wearable plantar pressure measurement system. The average performance and variability of left and right foot features were extracted from the raw plantar pressure data. An extensive feature database including 380 gait features and 4 physical characteristics was established by mathematical methodologies. Optimal feature subsets for both tasks are selected after stepwise feature selection including Spearman correlation coefficient, mRMR, and wrapper methods. Individual regression models and a 2-level ensemble learning model were trained for the 40mFPWT and TUGT tasks, respectively. The WAE model that weighs the contribution of each ensemble member proportionally based on RI value has the best performance in the testing set, with an MAE of 2.686 seconds, MAPE of 9.602%, and RMSE of 3.316 seconds for the 40mFPWT and an MAE of 1.280 seconds, MAPE of 12.389%, and RMSE of 1.905 seconds for the TUGT. The proposed technique has the potential to be a novel approach for objectively quantifying the functionally dependent gait features, which could be developed as a tool for the rehabilitation evaluation of motor function in individuals with KOA. This study fills the vacancy in dynamic functional assessment for patients with KOA based on wearable devices. In future work, a variety of sensing technologies will be integrated to evaluate and predict more functional tests, providing more accurate and scientific support in fields such as sports medicine and rehabilitation therapy.