Machine learning predicts chronic hip surgery pain from preoperative recovery factors

A machine-learning model combining nutritional status, inflammation, psychological factors, and muscle measurements from routine preoperative CT scans predicted chronic postsurgical pain after total hip arthroplasty in adults with hip fracture. In an independent temporal test cohort of 352 patients, the support vector machine model achieved an area under the receiver operating characteristic curve of 0.854, while its accuracy was 0.770, sensitivity 0.789, specificity 0.763, negative predictive value 0.913, and Brier score 0.125. The findings point to a multifactorial risk profile in which physiological reserve, muscle quality, inflammation, age, and preoperative anxiety collectively help identify patients whose surgical pain may persist for at least three months.

The retrospective, single-center cohort study included 881 consecutive patients aged 50 or older who underwent primary unilateral total hip arthroplasty for hip fracture at the Affiliated Hospital of Qingdao University between October 2022 and March 2026. Patients enrolled through September 2024 formed the development cohort of 529, with 370 assigned to model training and 159 to internal validation. A second group of 352 patients enrolled from October 2024 through March 2026 served as an independent temporal test set. Chronic postsurgical pain was defined as new or increased surgery-related pain around the operated hip that persisted for at least three months and reached at least 4 on the 0-to-10 numeric rating scale at the three-month assessment, after other identifiable causes had been excluded.

Before modeling, investigators assembled a broad set of candidate predictors covering demographics, comorbidities, perioperative factors, preoperative and early postoperative pain, long-term analgesic use, anxiety, depression, pain catastrophizing, laboratory-based nutrition and inflammation measures, and CT-derived muscle characteristics. The laboratory indicators included the Geriatric Nutritional Risk Index, Prognostic Nutritional Index, neutrophil-to-lymphocyte ratio, C-reactive protein-to-albumin ratio, and Systemic Immune-Inflammation Index. Two complementary feature-selection methods, least absolute shrinkage and selection operator regression and the Boruta algorithm, were used within the training data. Their intersection retained seven predictors: GNRI, gluteus medius/minimus muscle density, gluteus medius/minimus cross-sectional area, neutrophil-to-lymphocyte ratio, age, preoperative anxiety, and PNI.

The researchers then compared six algorithms: logistic regression, random forest, support vector machine, extreme gradient boosting, categorical boosting, and k-nearest neighbors. The support vector machine was selected after considering discrimination, sensitivity, and interpretability. Its area under the curve was 0.974 in the training set and 0.944 in the internal validation set, where sensitivity reached 0.886 and negative predictive value 0.949. Performance declined in the later temporal test set but remained above the level of random classification. The test-set precision-recall area was 0.754, calibration was described as showing agreement between predicted probabilities and observed outcomes, and decision-curve analysis indicated greater net benefit than treating everyone or no one across certain risk thresholds. In the temporal test set, positive predictive value was 0.534 and the F1 score was 0.637, indicating that the model’s positive predictions remained substantially less certain than its negative predictions.

Shapley additive explanations were used to examine how the final support vector machine assigned importance to its inputs. GNRI made the largest contribution, accounting for 35.4% of the model’s attribution, followed by gluteus medius/minimus muscle density at 13.6% and neutrophil-to-lymphocyte ratio at 12.5%. Age contributed 10.8%, preoperative anxiety 9.7%, PNI 9.5%, and gluteus medius/minimus cross-sectional area 8.4%. The direction of the model’s associations was also informative: lower GNRI, lower gluteus medius/minimus muscle density, higher neutrophil-to-lymphocyte ratio, older age, and preoperative anxiety were associated with increased predicted risk. Lower PNI and lower muscle-related parameters showed an overall trend toward greater predicted risk. These are model-derived associations, not evidence that changing any one factor would cause pain to fall or rise. The analysis explains how the algorithm uses the selected variables; it does not establish a causal mechanism for chronic postoperative pain.

The CT component introduced a two-dimensional view of the muscles around the hip. Cross-sectional area was treated as a measure of muscle quantity, while mean attenuation in Hounsfield units served as a measure of muscle quality; lower density generally indicates greater intramuscular fat infiltration. Measurements were taken from the side opposite the fracture because pain, immobilization, and local injury could alter muscle appearance around the injured hip. The gluteus maximus was assessed at the acetabular roof, the gluteus medius and minimus below the sacroiliac joint, and the rectus femoris below the lesser trochanter. Trained assessors blinded to postoperative pain outcomes performed the measurements, and repeated assessments in 20 patients showed excellent intra-observer and inter-observer reliability. Intraclass correlation coefficients ranged from 0.865 to 0.983 within observers and from 0.850 to 0.966 between observers.

Patients who developed chronic postsurgical pain in the development cohort were older and more likely to report preoperative long-term analgesic use, smoking, diabetes, anxiety, and pain catastrophizing. They also had lower gluteus maximus density, smaller gluteus medius/minimus cross-sectional area, lower gluteus medius/minimus density, lower GNRI and PNI, and higher neutrophil-to-lymphocyte ratio, C-reactive protein-to-albumin ratio, and Systemic Immune-Inflammation Index. Chronic pain occurred in 137 of 529 patients in the development cohort, an incidence of 25.9%, and in 90 of 352 patients in the temporal test cohort. The researchers did not report a statistically significant difference in event frequency between the two cohorts, with P equal to 0.913. These group differences are observational and cannot establish that malnutrition, muscle deterioration, inflammation, or anxiety caused persistent pain.

The model’s principal implication is that a single preoperative pain score may capture only part of a patient’s vulnerability to chronic postoperative pain. The strongest signal, GNRI, reflects nutritional risk using serum albumin and body-weight information, while PNI combines albumin and lymphocyte measurements and the neutrophil-to-lymphocyte ratio reflects inflammatory balance. The muscle measures add information about both the amount and quality of peri-hip tissue, and anxiety captures a psychological factor that may influence pain processing and participation in rehabilitation. The study authors describe these factors as components of preoperative recovery capacity, but the model itself does not demonstrate a biological pathway in which they interact causally. Its temporal validation and interpretability analysis strengthen the model as a risk-estimation tool without making it a treatment test. The study also did not report a direct comparison showing that this combined feature set performed better than a model limited to surgery-related information alone.

Several limits temper the findings. The study came from one hospital, and the independent test cohort was later in time but remained part of the same center, so multicenter external validation and prospective testing are still needed. All participants had complete data for the included variables, and no missing-data imputation was performed, which may restrict how the approach performs when information is incomplete. The final model also included only seven features after selection, and feature selection was not repeated within each cross-validation fold, a design choice the authors acknowledged could leave some risk of overfitting despite 19.57 outcome events per selected variable. Contralateral muscle measurements reduced local injury effects but might not represent pre-fracture muscle reserve in patients with bilateral sarcopenia, frailty, or chronic disuse. Pain is subjective, while rehabilitation adherence, treatment preferences, lifestyle, and other perioperative influences were not fully captured.

To make the model accessible, the research team built a web-based calculator that accepts the seven selected patient and imaging values and returns an individualized probability of chronic postsurgical pain. Such a tool could help clinicians identify higher-risk patients before discharge and prompt closer attention to recovery planning, but its practical value has not yet been established in a clinical trial. The study did not test whether nutritional optimization, psychological support, rehabilitation changes, or other interventions reduce pain, and it cannot show that acting on a model prediction improves outcomes. The results instead provide a prediction framework for a defined population of adults with hip fracture undergoing primary unilateral arthroplasty at one hospital. Larger prospective multicenter cohorts, objective functional outcomes, longitudinal pain trajectories, and intervention studies will be needed before that framework can guide patient care.

Subject of Research: Machine-learning prediction of chronic postsurgical pain after total hip arthroplasty in adults with hip fracture

Article Title: Nutritional status, inflammatory burden, and CT-derived muscle phenotypes for predicting chronic postsurgical pain after total hip arthroplasty: an explainable machine learning study

Article References: Jia, X., Han, X., Cui, X., Zhao, Y., Zhang, Y., & Xu, K. (2026). Nutritional status, inflammatory burden, and CT-derived muscle phenotypes for predicting chronic postsurgical pain after total hip arthroplasty: an explainable machine learning study. Frontiers in Nutrition, 13, Article 1958865. https://doi.org/10.3389/fnut.2026.1958865

Image Credits: AI Generated

DOI: 10.3389/fnut.2026.1958865

Keywords: chronic postsurgical pain, total hip arthroplasty, hip fracture, machine learning, nutritional risk, inflammation, skeletal muscle, preoperative anxiety, CT imaging, clinical prediction, Nutritional, status