Mind Matters: Pain-Associated Psychological Distress Phenotypes and Machine Learning-Based Predictive Models in Osteoarthritis and Total Joint Arthroplasty

Abstract

Data from the Centers for Disease Control and Prevention projects that the number of adults with physician-diagnosed osteoarthritis (OA) is expected to increase from 67.2 million in 2025 to 78.4 million by 2040.1 OA contributed to 9.5% (95% confidence interval (CI): 8.6, 10.1) more years lived with disability in 2020 compared to rates in 1990, and ranked seventh among the leading causes of years lived with disability.1 Although total joint arthroplasty (TJA) is the gold standard treatment for OA and is generally effective, approximately 15-30% of patients continue to experience postoperative pain and disability.2 Given the high prevalence of negative mood, maladaptive coping, and low self-efficacy among patients with OA, the leading indication for TJA, psychological distress may be a key contributor to suboptimal recovery and an important target for improving outcomes. A key aspect of OA disease and management is the variability in patient characteristics, including clinical, structural, and psychosocial factors.3 While much research has focused on patients under established care, less is understood about the demographic, health, and psychological profiles of patients seeking care for a new episode of OA, or whether differences exist in these factors between OA joint sites. Identifying these potential differences at the baseline encounter may offer valuable opportunities for intervention, guide phenotyping efforts, and help slow the progression of functional impairments. As patients progress through the OA disease continuum and potentially undergo TJA, they often encounter increasing behavioral and psychological burdens related to their condition.4 Identifying and grouping patients based on psychological distress is essential for providing comprehensive, person-centered care and improving overall quality of life. Psychological distress can intensify pain perception, increase disability, reduce patients' quality of life, hinder participation in social activities, impair the ability to perform activities of daily living, and affect work capacity.4–10 Recognizing the psychological aspects of pain allows clinicians to adopt a biopsychosocial approach that addresses both physical and behavioral health needs, ultimately leading to more effective and sustainable pain management strategies.11–13 The results of integrating psychologically informed treatment and addressing this distress have demonstrated promise in patients with OA and those undergoing TJA.14–22 Targeted interventions may include coping skills training, positive reinforcement, motivational interviewing, and cognitive behavioral strategies.13 However, significant gaps remain in understanding whether baseline demographic, health, and psychological profiles differ across OA joint sites, how psychological distress evolves following TJA, and which patient subgroups may benefit most from behavioral health referral or targeted psychologically informed treatment strategies. Therefore, this dissertation seeks to address key gaps in the literature through three specific aims, with the overall goal of informing phenotyping efforts, enhancing risk stratification, advancing precision-based care, and improving outcomes for individuals with OA and those undergoing TJA. The three aims of this dissertation are outlined below to describe the patient population and methods used. Aim I: To describe baseline demographic, health, and psychological profiles in patients seeking care for a new episode of hip, knee, or shoulder osteoarthritis. We leveraged data from the Duke University Total Joint Arthroplasty Learning Health Unit (LHU) in patients with a new episode of hip, knee, or shoulder OA. Various patient demographics and the Charlson Comorbidity Index score (CCI) were assessed. Patient-reported outcome measures (PROMs) included the Patient-Reported Outcomes Measurement Information System-Physical Function (PROMIS-PF), PROMIS Pain Interference (PI), and the Optimal Screening for Prediction of Referral and Outcome-Yellow Flag (OSPRO-YF) measure. All continuous variables were assessed for normality and found to be non-normally distributed; therefore, the Kruskal-Wallis test was used to compare patient demographics, psychological profiles, and patient-reported outcomes (PROs) across OA joint sites. Post hoc pairwise comparisons were conducted using Wilcoxon rank-sum tests with Bonferroni correction for multiple testing. Pearson’s chi-square and Fisher’s exact tests were used to compare categorical variables, as appropriate based on cell counts. A total of 3,228 patients were included in the analysis. On average, patients were 62.5 years old (standard deviation (SD) 12.0), predominantly female (64.1%), and Caucasian (68.4%). Most patients in the cohort had a diagnosis of knee OA (61.7%) compared to hip (26.0%) and shoulder OA (12.3%). The average body mass index (BMI) was 32.5 kg/m² (SD 7.9), and 64.3% of participants were never tobacco users. We observed that despite having lower comorbidity burden, patients with hip OA demonstrated worse physical function (mean PROMIS-PF T-score = 39.4, SD 8.00; p < 0.001), higher pain interference (mean PROMIS-PI T-score = 62.3, SD 7.19; p < 0.001), and greater psychological distress (mean OSPRO-YF count = 5.51, SD 3.54; p < 0.001) compared to patients seeking care for knee or shoulder OA. Psychological domains such as negative mood, negative coping, and low self-efficacy/acceptance were most prevalent in patients with hip OA compared to those with knee or shoulder OA. While the magnitude of differences in physical function, pain interference, and OSPRO-YF counts was modest, our results suggest that considering joint-specific differences when assessing patient needs and initial intervention strategies may be important for improving long-term OA care management. Future research should investigate the longitudinal stability of clinical and psychological profiles across OA joint sites to determine whether these profiles remain consistent or change over time in relation to disease progression or treatment. Aim II: To describe the transition of psychological distress phenotypes among patients undergoing total joint arthroplasty. Our objectives were to 1) characterize and describe the transition of psychological phenotypes from pre- to post-TJA and 2) evaluate patient characteristics across phenotypes. To address our first objective, latent transition analysis (LTA) was used to identify and describe transition probabilities of psychological distress phenotypes preoperatively and postoperatively using the OSPRO-YF tool. To address our second objective, demographics, PROMIS-PI, PROMIS-PF, pain scores (numeric pain rating scale), and High Impact Chronic Pain (HICP) were compared across phenotypes using both parametric and non-parametric tests, as appropriate. Post hoc analyses involved pairwise comparisons with Bonferroni correction. The study included 494 patients who underwent primary hip (43%) or knee (57%) arthroplasty at Duke University Health System (2018-2024). Considering various fit and performance indices, LTA identified five distinct phenotypes pre- and postoperatively: 1) low distress, 2) low self-efficacy and acceptance, 3) low self-efficacy and poor pain coping, 4) poor pain coping, and 5) high distress. Patients’ age, history of anxiety, depression, Elixhauser comorbidity score, and length of stay were associated with phenotype membership (all p < 0.05). Age was lowest in the poor pain coping phenotype (61.5 years old) and highest in the low self-efficacy and acceptance phenotype (67.9 years old). Anxiety, depression, comorbidity score, and length of stay were more frequent and highest in the high distress phenotype. About half of the patients (n=223, 45.1%) remained in the same class from pre- to post-TJA. Among patients characterized as high distress preoperatively (n=108), 63 (58.3%) remained in the high distress category postoperatively. The incidence of new-onset high distress postoperatively was 6.0% (23 of 386). Aside from those with high distress before surgery, most patients who transitioned to high distress after surgery were classified as the “low self-efficacy and poor pain coping” phenotype preoperatively. Statistically significant differences were observed across the phenotypes in both preoperative and postoperative PROs. High distress correlated with lower PROMIS-PF scores, higher PROMIS-PI scores, and greater prevalence of HICP (p < 0.001). Patients in the low distress phenotype reported higher PROMIS-PF, lower PROMIS-PI, less pain, and lower prevalence of HICP (p < 0.001). We found that some phenotypes demonstrated dynamic, state-like characteristics, while others remained more stable (i.e., low distress, high distress). These transitions likely reflect the dynamic nature of certain psychological and behavioral factors, emphasizing the importance of a personalized approach to intervention. Such insights could guide targeted screening and the development of adaptive, phenotype-specific interventions aimed at optimizing both psychological well-being and postsurgical outcomes after arthroplasty. Aim III: To evaluate machine learning algorithms and develop a risk calculator predicting high psychological distress after total joint arthroplasty. For Aim III, we used a cohort of patients (n = 553) undergoing primary hip, knee, or shoulder arthroplasty to develop and evaluate machine learning models that predict membership in the high psychological distress phenotype after TJA. Four machine learning (ML) models, including XGBoost, random forest, elastic-net logistic regression, and Super Learner, were trained to predict the high distress phenotype using only preoperative demographic, clinical, and patient-reported data. The dataset was split into training and testing cohorts (70:30). Model performance was assessed through area under the curve (AUC), accuracy, Brier score, sensitivity, specificity, calibration, and decision curve analysis. A web-based risk calculator was developed from the best-performing model. Postoperatively, 19% (n=106) of patients were identified as high distress. The elastic-net model showed strong discriminative ability (AUC=0.82) and superior interpretability compared to more complex models. Key preoperative predictors of high distress included OSPRO-YF count and summary score, comorbidity burden, depression, PROMIS-PF, and female sex. The elastic-net model was used to develop an interactive web-based Shiny app calculator to facilitate clinical translation. The model may enhance preoperative strategies by identifying patients who could benefit from targeted behavioral health support before surgery. Future research should focus on external validation to confirm the model’s generalizability and utility across various clinical settings.