Classification and stratification of patient pain archetypes following total knee arthroplasty: a machine learning approach.

BACKGROUND

Up to 20% of total knee arthroplasty (TKA) patients experience significant postoperative pain that delays recovery and increases risk of chronic pain. Early identification of high-risk patients may allow for timely targeted interventions. Prior studies on postoperative pain trajectories have been limited by small cohorts and restricted methodology. This exploratory study characterized postoperative pain archetypes using machine learning and developed a perioperative predictive model to identify patients at higher risk for postoperative pain.

METHODS

This single-center retrospective study analyzed 17,200 primary unilateral TKAs (2021-2024), randomly divided into 80% training (n=13,760) and 20% testing (n=3,440) sets. Pain scores (Numeric Rating Scale (NRS) 0-10) collected 0-72 hours postoperatively were modeled using long short-term memory with contrastive learning and K-means clustering. Least Absolute Shrinkage and Selection Operator regression was applied to 107 preoperative and intraoperative variables for feature selection, and selected variables were used to train an eXtreme Gradient Boosting (XGBoost) model. Model performance was assessed using accuracy and area under the receiver operating characteristic curve (ROC-AUC).

RESULTS

Two distinct pain archetypes were identified on the training set: a "low pain" (n=7,082) and "high pain" cluster (n=6,678). The high pain cluster had higher mean postoperative NRS scores, greater cumulative pain burden, higher opioid consumption, and more chronic pain consultations. XGBoost predicted high-pain cluster membership with 64% accuracy and ROC-AUC of 0.68, with key predictors included younger age, ambulatory surgery, higher mode of tolerable NRS score, genicular block, and higher Patient-Reported Outcomes Measurement Information System-10 pain scores. Clustering using NRS data from only the first 12 or 24 postoperative hours showed moderate concordance with 72-hour results (61.3% and 62.4%, respectively).

CONCLUSIONS

Machine learning identified distinct postoperative pain trajectories after TKA, and early pain data predicted later pain patterns. Incorporating early pain profiles into perioperative care may support proactive, individualized pain management.