1 Introduction

Total hip arthroplasty (THA) is increasingly recognized as a transformative surgical procedure for patients with advanced hip osteoarthritis, significantly restoring mobility and enhancing quality of life. Driven by an aging population and the high prevalence of degenerative joint diseases, demand for THA procedures has markedly increased, reflected by a 69.5% rise between 2006 and 2014 alone.¹ This substantial growth underscores the procedure's success, while simultaneously highlighting the necessity for all stakeholders to adapt to the increasing volume and complexity of cases effectively.

Beyond the rise in procedural demand, THA poses unique financial and clinical challenges for Medicare and Medicaid, social programs responsible for covering a considerable proportion of these surgeries. The global period is inclusive of the first 90 days post operatively, which may capture complications including readmissions and emergency department (ED) visits. Medicaid patients incur significantly greater total costs within the 90-day postoperative period compared to their non-Medicaid counterparts including extended hospitalizations, higher ED utilization, and increased readmission rates owing to a frequently higher comorbidity burden.² A recent analysis further quantified this disparity, demonstrating that Medicaid coverage independently predicts elevated risk for 90-day postoperative complications, with an odds ratio of 3.15 for ED visits and 2.46 for readmissions after adjusting for variables such as age, body mass index (BMI), and overall health status.³ Such evidence underscores the urgent need for targeted strategies that mitigate these postoperative complications, enhance patient care, and optimize healthcare resource allocation.

Recently, machine learning (ML) has emerged as a powerful analytic approach within orthopedic research, particularly for risk assessment and prediction of surgical outcomes. A study by Clement et al. demonstrated the capability of ML algorithms to accurately predict functional outcomes following THA, highlighting their distinct advantage over traditional statistical methods in managing complex and nonlinear clinical data.⁴ By incorporating a broad array of patient-specific information—including demographics, clinical comorbidities, and perioperative variables—ML can effectively stratify patients based on their individual risk for postoperative complications, thereby enabling clinicians to target interventions more precisely. Leveraging this potential, the present study aims to utilize data from the Michigan Arthroplasty Registry Collaborative Quality Initiative (MARCQI) to develop and validate ML models specifically for predicting 90-day postoperative readmissions and ED visits. By doing so, we intend to enhance clinical decision-making, improve patient outcomes, and promote cost-effective management strategies for THA patients.

2 Methods

2.2

2.2 Predictive modeling

For the predictive factor analysis, multilayer perceptron neural network (MPNN) machine learning method (Fig. 1) was applied to predict the factors contributing to readmission in 90 days and their individual normalized importance. The collected data was split into a training set (70% of datasets) and testing set (30%) of datasets to investigate the ability of the MPNN machine learning method to determine the outcome. 70% of our datasets were randomly selected and utilized to train ML algorithms and 30% of datasets were utilized for internal validation and testing. The MPNN architecture included one hidden layer and 50 units in the hidden layer. The loss function was cross-entropy. The Bayesian method was applied for regularization. A batch with the automatic method was selected for the training type. The scaled conjugate gradient was used for the optimization algorithm. The initial learning rate was 0.4. Hyperbolic tangent was applied for the activation function in the hidden layer and Softmax was used for the activation function in the output layer. Predicting outcome was summarized as readmission in 90 days (Yes vs. No). Independent variable importance analysis was then performed. The accuracy of machine learning MPNN model discrimination was determined by using the area under the curve (AUC) of receiver operating curve (ROC). Excellent candidate models was determined with an AUC exceeding 0.8. SPSS software (Ver. 28, IBM, Armonk, NY) was used for traditional statistical analysis and MPNN machine learning analysis.

Close

Fig. 1 The multilayer perceptron neural network ML architecture used for prediction of factors contributing to readmission in 90 days.

Export to PPT

3 Results

3.3

3.3 MPNN analysis

A multilayer perceptron network (MPNN) model was developed to predict 90-day readmissions using demographic and clinical variables. The model included 12 input variables and a single hidden layer with six neurons. The training accuracy of the model was 94.7%, and the testing accuracy was 95.2%. The area under the curve (AUC) for readmission classification was 0.710, indicating moderate predictive performance (Fig. 2) (Fig. 3).

Close

Fig. 2 Accuracy of MPNN ML model in predicting readmission in 90 days among THA patients.

Export to PPT

Close

Fig. 3 The normalized importance of independent factors related to readmission in 90 days among THA patients.

Export to PPT

MPNN analysis ranked the most influential predictors of readmission as follows: LOS (importance = 100.0%), ASA score (62.9%), bleeding disorder (51.7%), alcohol use (22.3%), preoperative BMI (18.1%), and age at case (18.1%). LOS was the strongest predictor, with extended hospital stays significantly increasing readmission risk.

The model used a hyperbolic tangent activation function in the hidden layer and a softmax function in the output layer. The cross-entropy error for training was 181.5, while the testing error was 72.2. The overall percentage of incorrect predictions was 5.3% for training and 4.8% for testing.

4 Discussion

This study identified key predictors for 90-day readmissions following THA, utilizing demographic, clinical, and perioperative variables combined with a ML-based predictive model. The major findings include: (1) a 90-day readmission rate of 5.1%, with the strongest predictors being length of hospital stay, ASA score, discharge disposition, and current smokers (2) the multilayer perceptron neural network (MPNN) achieved moderate predictive accuracy (AUC = 0.710), highlighting its potential clinical utility for individualized risk stratification.

The significance of these findings lies primarily in their practical implications for clinical management and healthcare resource utilization. Length of hospital stay emerged as the strongest predictor, with longer hospitalization markedly increasing readmission risk. This aligns with prior research indicating that extended hospital stays are associated with higher odds of 90-day readmission after THA. For instance, a study by Benito et al. found that patients with a LOS greater than one day had significantly higher odds of 90-day readmission compared to those discharged within a day.⁵ Specifically in our model, THA patients with an LOS of 2 days (odds ratio (OR) of 2.89 for 90-day readmission), 3 days (OR of 2.80), and 4 days (OR of 2.83) had greater 90-day readmission rates compared with LOS of 1 day (p < .05). Another ML study by Chen et al. found that, in revision THA cases, ML models accurately predicted prolonged length of stay, with an AUC of 0.82. Key predictors included infection as the indication for revision, preoperative labs, transfusion, operation time, and age.⁶ Similarly, Phruetthiphat et al. found that in a retrospective analysis of THA and TKA patients, infection was the leading cause of 30-day readmission, with comorbidities, BMI, and lower preoperative functional capacity significantly associated with higher readmission risk.⁷ Longer hospital stay was also identified as an independent predictor of readmission.⁷ These findings underscore the importance of efficient progression to early discharge, tailored to patient-specific risk factors, to prevent readmission.

Our findings corroborate previous literature that identifies ASA score as a significant predictor of postoperative complications.⁸ Consistent with existing studies, patients with higher ASA scores (ASA ≥3) experienced higher readmission rates, indicative of poorer baseline health and greater perioperative risk.⁹ For example, a study by Schaeffer et al. demonstrated that patients with an ASA score of ≥3 had a 2.9 times greater risk of readmission following total joint arthroplasty.

Moreover, discharge disposition significantly influenced readmission rates, with patients discharged to skilled nursing or rehabilitation facilities exhibiting markedly higher readmission rates compared to those discharged home. These findings parallels those of other orthopedic studies, suggesting that institutional discharge may reflect patient frailty and complex medical needs rather than merely influencing outcomes independently.¹⁰ A study by Cole et al. found that patients discharged to skilled nursing or rehabilitation facilities after THA had significantly higher 30-day readmission and complication rates compared to those discharged home, even after matching.¹⁰ Factors like older age, ASA IV status, COPD, and steroid use were also associated with increased readmission risk. Interestingly, Rajesh et al. found that in a large matched cohort, robotic-assisted THA was associated with a shorter hospital stay and fewer discharges to skilled nursing facilities compared to manual THA, while 90-day readmission rates were similar between groups.¹¹

The application of ML provides valuable insights beyond traditional statistical analysis, capturing complex interactions among variables that conventional methods might overlook. Our model achieved a moderate level of predictive accuracy, demonstrating potential for further refinement. Similar studies have shown that ML models, such as random forests and gradient boosting machines, can effectively predict 90-day readmissions after THA. Shaikh et al., using statewide data, was able to develop ML models—particularly random forest and gradient boosting—which predicted 30-day and 90-day outcomes after total joint arthroplasty with moderate accuracy (AUCs up to 0.73), and performed similarly when externally validated at a single center.¹² Another study by Howard et al. found that in a tertiary care center, ML models using electronic health records and patient-reported outcomes predicted 90-day TJA readmissions with strong performance (AUC >0.8). Key predictors included diabetes, certain medications, discharge to skilled nursing, and low confidence in social activity participation.¹³ These findings support the integration of ML approaches in clinical settings to enhance risk stratification and inform targeted interventions.

Alternate explanations for our findings must also be considered. For example, while prolonged hospital stay was identified as a major predictor, it is plausible that extended hospitalizations reflect pre-existing patient complexity rather than a causality of increased readmission risk. Additionally, the role of insurance type may partially reflect differences in social determinants of health, access to care, health literacy, or social support networks rather than the insurance status alone. Future studies should explore these underlying factors more deeply to clarify causal relationships.

Clinically, these findings highlight actionable areas for intervention, particularly improving preoperative patient optimization, careful postoperative discharge planning, and focused follow-up care for high-risk patients identified via predictive modeling. For instance, early identification of patients with a history of higher ASA scores could prompt targeted interventions such as enhanced perioperative monitoring, anticoagulation management, or structured postoperative rehabilitation programs.

Limitations of this study include its retrospective nature and reliance on retrospective cohort data, which inherently carry risks of recording errors and incomplete capture of relevant clinical nuances. Additionally, our ML model demonstrated moderate predictive ability, suggesting that additional predictive variables, larger sample sizes, or more sophisticated ML techniques may improve future models' performance. Lastly, generalizability might be limited by regional and institutional differences specific to the MARCQI.

Future research should explore prospective validation of ML-driven predictive models, assessing real-world effectiveness in reducing postoperative complications when integrated into clinical workflows. Additionally, examining social determinants of health more comprehensively could refine the predictive accuracy and clinical applicability of risk stratification models. Further studies might also assess patient-centered outcomes and quality-of-life metrics alongside readmission and cost metrics to fully understand intervention impacts.

In conclusion, this study demonstrates that length of hospital stay, ASA score, discharge disposition, and current smoking significantly predict 90-day readmissions following THA. ML models, while promising, require further refinement for optimal clinical implementation. By integrating these predictive insights into clinical practice, healthcare providers may effectively target interventions, reduce postoperative complications, and enhance quality of care for THA patients.

CRediT authorship contribution statement

Zachary Crespi: Conceptualization, Data curation, Writing – review & editing, Formal analysis. Usher Khan: Writing – original draft. Abdul-Lateef Shafau: Writing – original draft. Fong Nham: Writing – review & editing. Chaoyang Chen: Software, Investigation, Methodology. Bryan Little: Supervision. Hussein Darwiche: Supervision.

Patient consent

Statement: No guardian or patient consent form was needed for this study as this is non-applicable to the design of the study.

Ethics in publishing statement

I testify on behalf of all co-authors that our article submitted followed ethical principles in publishing.

All authors agree that:

This research presents an accurate account of the work performed, all data presented are accurate and methodologies detailed enough to permit others to replicate the work.

This manuscript represents entirely original works and or if work and/or words of others have been used, that this has been appropriately cited or quoted and permission has been obtained where necessary.

This material has not been published in whole or in part elsewhere.

The manuscript is not currently being considered for publication in another journal.

That generative AI and AI-assisted technologies have not been utilized in the writing process or if used, disclosed in the manuscript the use of AI and AI-assisted technologies and a statement will appear in the published work.

That generative AI and AI-assisted technologies have not been used to create or alter images unless specifically used as part of the research design where such use must be described in a reproducible manner in the methods section.

All authors have been personally and actively involved in substantive work leading to the manuscript and will hold themselves jointly and individually responsible for its content.

Corresponding author's name: Usher Khan, BS.

Date: June 06, 2025.

Funding statement

This research did not receive any specific grant from funding agencies in the public, commercial, or not-for-profit sectors.