Abstract
The study developed a robust and interpretable ML model for the early prediction of PSD following TKA. This model can aid in preoperative risk stratification, facilitating personalized management strategies to improve postoperative outcomes. Further validation in larger cohorts and diverse settings is necessary to enhance its broader clinical applicability.
Introduction
Total knee arthroplasty (TKA) is a cornerstone treatment for end-stage osteoarthritis, significantly improving quality of life for many. However, approximately 20% of patients report dissatisfaction post-surgery. A critical, yet frequently overlooked, factor in this dissatisfaction is postoperative sleep disturbance (PSD), which affects nearly half of the TKA population. PSD is not merely a symptom of fatigue; it exacerbates pain perception, impairs wound healing, increases the risk of depression, and extends hospital stays, leading to higher healthcare costs.
Traditional statistical approaches, such as linear logistic regression, often struggle to model the complex, non-linear interactions between clinical, psychological, and socio-environmental factors that influence sleep. This study leverages advanced machine learning (ML) to develop a predictive tool capable of identifying at-risk patients early, enabling personalized perioperative interventions.
Methods
This retrospective study analyzed data from 505 patients who underwent TKA at Tianjin Hospital between May 2024 and March 2025. The study workflow included patient screening, feature selection, model training, and performance evaluation.
Feature Selection and Model Training
To identify the most significant predictors from an initial pool of 38 variables, the researchers employed Least Absolute Shrinkage and Selection Operator (LASSO) regression. This method effectively reduced dimensionality while maintaining model interpretability.
Ten different ML algorithms were compared, including Logistic Regression, Support Vector Machines (SVM), Random Forest, XGBoost, and Gradient Boosting Machines (GBM). Hyperparameters were optimized using Bayesian optimization, and model robustness was ensured through 10-fold cross-validation.
Results
The Gradient Boosting Machine (GBM) emerged as the superior model, achieving an Area Under the Curve (AUC) of 0.906, an accuracy of 0.834, and a sensitivity of 0.879. In contrast, traditional Logistic Regression performed poorly with an AUC of only 0.666, highlighting the advantages of ensemble tree-based methods in capturing complex patient data.
The LASSO analysis identified six key predictors for PSD:
- Postoperative pain (VAS at 1 month)
- Age
- Postoperative anxiety (at 1 month)
- Living alone
- Urban residence
- Smoking history
Discussion
A significant finding of this study is the impact of socio-environmental factors. Patients living alone and those residing in urban areas were at higher risk for PSD. Living alone likely reflects a lack of immediate social support for pain management, while urban environments introduce stressors like noise and light pollution that disrupt recovery.
Furthermore, the identification of postoperative pain and anxiety as top predictors suggests that PSD is a modifiable complication. Enhanced perioperative pain management and psychological support could significantly mitigate sleep disturbances. To ensure the model remains a "white box" rather than a "black box," the researchers utilized SHAP (SHapley Additive Explanations) analysis, allowing clinicians to visualize the specific contribution of each risk factor for individual patients.
Limitations and Future Work
While promising, the study is limited by its single-center design and a relatively short follow-up period of one month. Future research should focus on multi-center validation and the inclusion of additional variables, such as genetic markers and long-term sleep outcomes. Integrating this predictive model into real-time clinical workflows will be essential to evaluate its impact on patient recovery in practice.
Summary updated