Abstract
Aims
To identify variables independently associated with same-day discharge (SDD) of patients following revision total knee arthroplasty (rTKA) and to develop machine learning algorithms to predict suitable candidates for outpatient rTKA.
Methods
Data were obtained from the American College of Surgeons National Quality Improvement Programme (ACS-NSQIP) database from the years 2018 to 2020. Patients with elective, unilateral rTKA procedures and a total hospital length of stay between zero and four days were included. Demographic, preoperative, and intraoperative variables were analyzed. A multivariable logistic regression (MLR) model and various machine learning techniques were compared using area under the curve (AUC), calibration, and decision curve analysis. Important and significant variables were identified from the models.
Results
Of the 5,600 patients included in this study, 342 (6.1%) underwent SDD. The random forest (RF) model performed the best overall, with an internally validated AUC of 0.810. The ten crucial factors favoring SDD in the RF model include operating time, anaesthesia type, age, BMI, American Society of Anesthesiologists grade, race, history of diabetes, rTKA type, sex, and smoking status. Eight of these variables were also found to be significant in the MLR model.
Conclusion
The RF model displayed excellent accuracy and identified clinically important variables for determining candidates for SDD following rTKA. Machine learning techniques such as RF will allow clinicians to accurately risk-stratify their patients preoperatively, in order to optimize resources and improve patient outcomes.
Cite this article: Bone Jt Open 2023;4(6):399–407.
Take home message
This study identified important perioperative factors integral in determining same-day discharge of patients following revision total knee arthroplasty (rTKA).
Machine learning techniques can accurately predict patients suitable for outpatient rTKA.
Using these techniques, surgeons can engage in shared decision-making with patients to develop an effective strategy for management.
Introduction
Total knee arthroplasty (TKA), an effective intervention for patients with end-stage knee osteoarthritis, is among the fastest-growing procedures performed worldwide due to the increased longevity of the population and the burden of osteoarthritis.1 While postoperative outcomes following TKA are generally positive, revision surgery can be required for various indications, including aseptic loosening, instability, and infection.2 As the frequency of TKA procedures increases, the incidence of revision total knee arthroplasty (rTKA) is also rising.3-5 Since rTKA is a more complex procedure than primary TKA, it results in relatively higher complication rates, extended hospitalization, and less satisfactory functional outcomes.6
Revision TKA has a median hospital length of stay (LOS) of three to five days.7 Recent advancements in multimodal pain control, blood management techniques, and rapid recovery protocols have allowed for a shorter postoperative LOS.8,9 This is exemplified by the recent (2021) removal of rTKA procedures from the Inpatient Only (IPO) list by the Centres for Medicare and Medicaid Services (CMS) in the USA.10 Furthermore, recent literature has demonstrated that outpatient rTKA with same-day discharge (SDD) is safe in carefully selected patients and case scenarios, and with appropriate patient selection, early discharge does not increase 90-day readmissions or emergency department visits.11,12 While cardiopulmonary disease, bleeding disorders, chronic kidney disease, obesity, uncontrolled diabetes (type I or II), and dependent functional status have been shown to be predictive of SDD in primary arthroplasty, there is very limited data identifying factors that preclude SDD following rTKA.13 Given the differences between primary and revision TKA, these factors cannot be extrapolated, and as such, accurately determining factors associated with SDD is vital to determine appropriate outpatient selection for rTKA.
Machine learning is a branch of artificial intelligence that creates complex models to iteratively improve its predictive capacity based on the quantity of data input. Through its ability to learn complicated non-linear or linear relationships, machine learning decreases bias and can provide more accurate results when compared to conventionally used logistic regression.14 Machine learning has previously been used to predict LOS for various non-orthopaedic surgeries, but has only recently started to gain traction in orthopaedics, with one recent study using artificial neural networks to predict SDD following primary TKA.15 However, this is the first study using machine learning techniques to predict SDD following rTKA.
The purpose of this study is to develop trained machine learning models, cross-referenced with traditional multivariable logistic regression (MLR), to determine the most important pre- and perioperative variables that may predict SDD in patients undergoing rTKA. We hypothesized that the trained models would have better accuracy when compared to MLR and will identify variables associated with increased overall health to be predictive of SDD.
Methods
Data source
Data were obtained from the American College of Surgeons National Quality Improvement Programme (ACS-NSQIP) database from the years 2018 to 2020. ACS-NSQIP is a large clinical database that collects over 150 pre-, peri-, and postoperative variables up to 30 days following surgery at over 680 USA hospitals. Rigorous data collection and auditing by the American College of Surgeons has allowed for high-quality data with inter-reviewer reliability greater than 98%.16
Study population
Patients with elective, unilateral rTKA procedures and a total LOS between zero and four days were isolated in the ACS-NSQIP database using Current Procedural Terminology codes 27486 and 27487,17 corresponding to rTKA with or without allograft one component only and revision of the femoral and entire tibial component, respectively. Patients were then categorized into two cohorts: LOS of zero days (SDD), and LOS of two to four days.
Variable selection
Baseline patient demographic data, including sex, race, age, and BMI, were collected. Patient comorbidities and preoperative variables that were collected include diabetes mellitus I and II requiring medication, dyspnea status, smoking, chronic obstructive pulmonary disease (COPD), hypertension requiring medication, chronic steroid use, > 10% weight loss in the six months preceding surgery, current need for dialysis, history of disseminated cancer, current open/ infected wound, congestive heart failure within 30 days prior to surgery, ascites within 30 days prior to surgery, history of bleeding disorder, and red blood cell (RBC) transfusions within 72 hours prior to surgery. Additionally, patient functional status (classified as either independent or partially/totally dependent), preoperative anaemia status (classified using World Health Organization (WHO) guidelines into normal, mildly anaemic, and moderate-severely anaemic), and American Society of Anesthesiologists (ASA)18 grade (classified as either ASA grade I to II or ASA grade III to V) were collected.19,20 Indication for rTKA was collected, and included aseptic mechanical loosening, instability, pain, and periprosthetic joint infection (PJI). Intraoperative variables were collected, including type of rTKA (single component vs all components), primary anesthetic type, and operating time. All variables are defined in the ACS-NSQIP user guide.21
Exclusion criteria and missing data
Patients who had missing data for LOS, who underwent concurrent/additional surgical procedures, or who were readmitted within 30 days of discharge, were excluded from this study. Patients with LOS of one day were excluded from our dataset, as prior literature reports that one-day LOS in the ACS-NSQIP database includes a combination of outpatient and inpatient cases, likely as a result of discrepancies at the regional and hospital levels regarding postoperative observational care.15,22 Other variables with missing values were imputed using multiple imputation with the missForest methodology.23 The number of patients with missing data was as follows: LOS (n = 102); BMI (n = 584); ASA grade (n = 23); functional status (n = 47); and operating time (n = 71).
Statistical analysis
Prior to developing the machine learning models, the data were randomly divided into training (75%) and internal validation (25%) datasets. Five popular machine learning techniques, random forest (RF), artificial neural network (ANN), gradient boosted tree (GBT), Naïve Bayes (NB), and support vector machine (SVM), were then applied to the training and validation datasets to predict SDD following rTKA. RF is a regression-based classification algorithm that aggregates many decision trees trained on randomly sampled subsets of a complex dataset.24,25 When developing the RF model, a grid search was used to determine the best combination of tuning parameters, including number of trees and number of features at each split. These models were chosen based on previous machine learning studies that focused on binary classifications.15,26 Logistic regression was also performed using independent-samples t-tests and Pearson’s chi-squared tests to evaluate continuous and categorical variables, respectively. Statistically significant variables were then examined using MLR. All analyses were completed using Stata v. 16.1 (Stata Corp, USA). An α of 0.05 was used for statistical significance.
The predictive capacity of each model was assessed and compared using the area under the receiver operating characteristics curve (AUC), which is a measure of discrimination and is the gold-standard metric of machine-learning assessment. The receiver operating characteristic curve plots the false-positive rate on the x-axis and the true-positive rate on the y-axis. The AUC, or c-statistic, ranges from 0.5 to 1, with an AUC of 0.50 indicating that the model being studied has a 50% chance of predicting the outcome, and thus cannot distinguish between patients who did and did not undergo SDD.27 Model performance was also measured by visually assessing calibration curves. Calibration refers to the agreement between predicted outcomes and the observed frequency of the outcome.28 The model performs well on calibration when the curve is close to the bisector. Decision curves were further constructed in order to evaluate the models. The y-axis of the decision curve represents the net benefit, which judges whether clinical decisions have more benefit than harm, and each point on the x-axis represents a threshold probability that differentiates between patients who underwent SDD and those who did not. The model with the greatest AUC was further analyzed to determine the ten most important variables for predicting SDD, rated based on their contribution to the model.29
Results
Factors associated with same-day discharge following revision TKA
In total, 5,600 patients were included in the final analyses. Of these patients, 342 (6.1%) were discharged on the same day of surgery. Comparisons between patients who underwent SDD and patients who did not with regard to demographics and perioperative variables are displayed in Table I. Variables significantly associated with SDD in the MLR analysis are also displayed in Table I.
Table I.
Logistic regression analysis comparing demographic data, comorbidities, and preoperative and intraoperative variables between same-day discharge patients (zero-day length of stay) and normal discharge (two- to four-day length of stay) patients following revision total knee arthroplasty.
| Characteristic | SDD (n = 342) | Normal discharge (n = 5,258) | p-value | Multivariable p-value | OR (95% CI) |
|---|---|---|---|---|---|
| Demographics | |||||
| Mean age, yrs (SD) | 64.43 (9.87) | 65.94 (9.94) | 0.006 | 0.0024 | 1.02 (1.01 to 1.03) |
| Mean BMI, kg/m2 (SD) | 32.23 (6.27) | 34.07 (7.18) | < 0.001 | 0.009 | 1.03 (1.01 to 1.05) |
| Sex, n (%) | 0.009 | 0.010 | |||
| Male | 160 (46.78) | 2,089 (39.73) | 1.51 (1.17 to 1.94) | ||
| Female | 182 (53.22) | 3,169 (60.27) | 1 | ||
| Race, n (%) | 0.005 | ||||
| White | 296 (86.55) | 4,180 (79.50) | |||
| Black | 36 (10.53) | 972 (18.49) | |||
| Asian | 7 (2.05) | 59 (1.12) | |||
| Other | 3 (0.87) | 47 (0.89) | |||
| Comorbidities, n (%) | |||||
| Diabetes mellitus | 44 (12.87) | 1,156 (21.99) | < 0.001 | 0.015 | 0.66 (0.47 to 0.93) |
| Smoking history | 35 (10.23) | 489 (9.30) | 0.566 | ||
| CHF | 0 (0.00) | 27 (0.51) | 0.184 | ||
| Hypertension | 211 (61.70) | 3,627 (68.89) | 0.005 | ||
| Steroid use | 13 (3.80) | 264 (5.02) | 0.313 | ||
| Bleeding disorders | 11 (3.22) | 173 (3.29) | 0.941 | ||
| Dyspnea | 12 (3.51) | 351 (6.68) | 0.021 | ||
| COPD | 12 (3.51) | 295 (5.61) | 0.098 | ||
| Weight loss | 0 (0.00) | 22 (0.42) | 0.231 | ||
| Dialysis | 1 (0.29) | 24 (0.46) | 0.659 | ||
| Disseminated cancer | 3 (0.88) | 10 (0.19) | 0.011 | ||
| Ascites | 0 (0.00) | 1 (0.02) | 0.799 | ||
| Open wound/wound infection | 1 (0.29) | 56 (1.07) | 0.168 | ||
| RBC transfusions | 1 (0.29) | 10 (0.19) | 0.679 | ||
| Dependent functional status | 1 (0.29) | 178 (3.39) | 0.002 | 0.002 | 0.12 (0.02 to 0.85) |
| Anaemia WHO class, n (%) | < 0.001 | ||||
| Normal | 276 (80.70) | 3,400 (64.66) | 1 | ||
| Mild | 63 (18.42) | 1,565 (29.76) | 0.580 | 0.58 (0.42 to 0.79) | |
| Moderate/severe | 3 (0.88) | 293 (5.57) | < 0.001 | 0.20 (0.05 to 0.66) | |
| ASA grade III to V, n (%) | 168 (49.12) | 3,501 (66.58) | < 0.001 | 0.0037 | 0.69 (0.54 to 0.89) |
| Indication, n (%) | < 0.001 | < 0.001 | |||
| PJI | 6 (1.75) | 712 (13.54) | 1 | ||
| Noninfectious | 17.48 (2.42 to 126.18) | ||||
| Loosening | 94 (27.49) | 1,496 (28.45) | |||
| Instability | 81 (23.69) | 630 (11.98) | |||
| Pain | 44 (12.87) | 418 (7.95) | |||
| Other | 117 (34.21) | 1,496 (28.45) | |||
| Intraoperative variables | |||||
| One-component rTKA | 188 (54.97) | 1,394 (26.51) | < 0.001 | < 0.001 | 1.61 (1.24 to 2.08) |
| Mean operating time, mins (SD) | 91.30 (47.80) | 144.96 (64.03) | < 0.001 | < 0.001 | 1.02 (1.02 to 1.02) |
| Anaesthesia type | < 0.001 | ||||
| General | 161 (47.08) | 3,208 (61.01) | 1 | ||
| Neuraxial | 139 (40.64) | 1,028 (19.55) | < 0.001 | 2.42 (1.88 to 3.14) | |
| MAC/IV | 38 (11.11) | 931 (17.71) | 0.014 | 0.62 (0.43 to 0.91) | |
| Regional | 2 (0.58) | 85 (1.62) | 0.498 | 0.61 (0.14 to 2.54) | |
| None/Other | 2 (0.58) | 6 (0.11) | 0.0024 | 15.44 (2.64 to 90.34) |
-
ASA, American Society of Anesthesiologists; CHF, congestive heart failure; CI, confidence interval; COPD, chronic obstructive pulmonary disease; MAC/IV, monitored anesthetic care/intravenous; OR, odds ratio; PJI, periprosthetic joint infection; RBC, red blood cell; rTKA, revision total knee arthroplasty; SDD, same-day discharge; WHO, World Health Organization.
Machine learning algorithm development
The RF model had the highest AUC upon internal validation (0.810), outperforming the other models (Table II) (Figure 1). The RF model also demonstrated superior performance upon visual inspection of the calibration curves (Figure 2) as well as the decision curve analysis (Figure 3), in which it demonstrated a higher net benefit across the threshold values. The RF model identified operating time < 95 minutes as the most important variable for determining SDD, followed by anaesthesia type, age < 73, BMI, ASA grade, race, history of diabetes, rTKA type, sex, and smoking status (Figure 4).
Table II.
Summary of model training and validation results.
| Model | Training AUC | Validation AUC |
|---|---|---|
| MLR | 0.808 | 0.772 |
| RF | 0.831 | 0.810 |
| ANN | 0.824 | 0.789 |
| GBT | 0.828 | 0.802 |
| NB | 0.819 | 0.792 |
| SVM | 0.811 | 0.778 |
-
ANN, artificial neural network; AUC, area under the curve; GBT, gradient-boosted tree; MLR, multivariable logistic regression; NB, Naïve Bayes; RF, random forest; SVM, support vector machine.
Fig. 1
Receiver operating curve of random forest training model for patients with same-day discharge (black) and without same-day discharge (gray) following revision total knee arthroplasty.
Fig. 2
Calibration curves with 95% confidence intervals for prediction of same-day discharge using a) multivariable logistic regression (LG), b) random forest (RF) regression, c) artificial neural network (NN) regression, d) gradient-boosted tree (GBT) regression, e) Naïve Bayes (NB) regression, and f) support vector machine (SVM) regression.
Fig. 3
Decision curve analyses showing the net benefit against the threshold probabilities based on decisions from the model outputs. The curve titled “all” represents the prediction that all patients would undergo same-day discharge (SDD). The curve titled “none” represents the prediction that all patients would not undergo SDD. SVM, support vector machine.
Fig. 4
Normalized importance of pre- and intraoperative factors for same-day discharge based on random forest model. Importance is the degree to which the model is dependent on the factor. ASA, American Society of Anesthesiologists; rTKA, revision total knee arthroplasty.
Discussion
Outpatient rTKA with SDD has become more prevalent, in part due to a goal of decreasing the overall cost to the healthcare system.11 This study applied various machine learning strategies to predict SDD following rTKA in order to identify important characteristics for appropriate outpatient selection. Among all six tested models, the RF model yielded the highest AUC in both the training (AUC 0.831) and validation (AUC 0.810) datasets, demonstrating excellent accuracy and predictability. Prior literature has found RF to be superior to other machine learning techniques in handling many factors and highly non-linear data. As such, RF appears to be the machine learning algorithm of choice in many clinical studies which use analytical databases such as ACS-NSQIP.30,31 The results from our RF suggest that shorter operating time, anaesthesia type, age, BMI, ASA grade, race, history of diabetes, rTKA type, sex, and smoking status are most predictive of SDD following rTKA; all but race and smoking status were significant in the MLR analysis.
Operating time was a variable significantly highlighted in both RF and MLR, independent of whether the procedure was a single- or both-component revision. This finding has been previously reported by Garbarino et al,32 who used ACS-NSQIP from 2008 to 2016 to assess the effect of operating time on in-hospital LOS in patients undergoing rTKA. They found that as operating time increases, postoperative LOS also increases, thus concluding that appropriate staffing, communication, and coordination are vital to decreasing LOS and consequently, hospital costs. This finding was also recently reported by Buller et al,12 who conducted a matched cohort study at a single tertiary care referral centre. Furthermore, studies have found that increased operating time results in not only an increased LOS, but also increased postoperative complications such as surgical site infection, sepsis, and renal insufficiency.33,34 As such, taking appropriate measures to increase the efficiency and limit the duration of the rTKA operation is vital to decrease hospital costs and improve patient outcomes. Additionally, our MLR analysis found indication, specifically aseptic failure, to be associated with SDD, likely due to the association of PJI with more adverse perioperative outcomes.35 Based on our study, if the orthopaedic surgeon anticipates that the procedure will last more than 95 minutes, if the procedure is a both-component rTKA, or if the indication is PJI, they may want to consider admitting the patient to the hospital.
We found that the type of anaesthesia, specifically neuraxial anaesthesia, was also significant in both RF and MLR. Neuraxial anaesthesia results in decreased operating time, LOS, and postoperative complications such as surgical site infection in both primary and revision TKA.15,36,37 This is likely due to sympathetic blockade and consequent reduced blood loss.38,39 Therefore, increased utilization of neuraxial anaesthesia may improve patient outcomes and decrease LOS associated with rTKA.
BMI, age, anaemia status, ASA grade, sex, and history of diabetes were also identified from the RF model and considered significant predictors of SDD following rTKA by MLR analysis. These variables have been previously reported on in the context of primary TKA. Specifically, Wei et al15 found that upon using ANN to predict SDD following primary TKA, preoperative sodium, BMI, age, anaesthesia type, operating time, race, and anaemia status were key factors. Thus, accounting for these characteristics and delaying surgery until modifiable risk factors such as BMI have been addressed may greatly benefit the patient.
Worldwide, the incidence of rTKAs is projected to increase rapidly.5 Consequently, selecting patients suitable for outpatient rTKA has become increasingly relevant. Our study used many variables, including demographics, comorbidities, and perioperative factors, to determine the ideal candidate for outpatient rTKA. Several studies have previously shown the advantages of machine learning techniques such as RF and ANN to study risk assessment for various orthopaedic procedures, such as delirium in geriatric patients following hip fracture fixation and postoperative complications following lumbar spine fusion.26,40 Moreover, these techniques have also shown promise in predicting SDD following procedures such as laminectomy surgery and primary total knee and total hip arthroplasties.15,41,42 However, to the best of our knowledge, this is the first study applying RF and ANN to predict SDD following rTKA. This study adds to the existing literature regarding the utilization of various machine learning techniques to predict complications and SDD following common orthopaedic procedures. Physicians can apply these machine learning models to aid in appropriate patient selection for outpatient rTKA and engage in shared decision-making with patients. Furthermore, implementation of the important variables identified in this study may be used to optimize patients perioperatively.
Our study has limitations. First, our analyses may be biased by the retrospective nature of this study; however, this approach allowed for a large sample size, greatly improving the accuracy of the models tested.43 Second, as with any database study, we are limited by the number of variables included in the dataset and the quality of the dataset. However, ACS-NSQIP is a large dataset with many variables, and previous studies have found it to be of high quality, with high inter-reviewer reliability.16 This minimizes bias, allowing us to accurately distinguish between patients with SDD and patients with inpatient hospitalizations. Third, the application of machine learning to medicine is still relatively novel, and although several features were selected as important by the RF model, the exact nature of their importance is difficult to interpret. However, in this study, we have also provided the analyses from traditionally used MLR for cross-reference, allowing for greater clarification as to the importance of certain variables. It is recommended that surgeons use all published information on rTKA for appropriate decision-making. Finally, external validation of these models in a dataset that is not obtained through NSQIP is warranted.
In summary, the machine learning techniques developed in this study, especially RF, displayed excellent accuracy for the prediction of SDD following rTKA. Once cross-referenced with conventional MLR, the most predictive variables included operating time, BMI, age, and type of anaesthesia. With the rising emphasis on value-based care, outpatient surgical practice is rapidly expanding across all healthcare subspecialties.44,45 Incorporating models such as these can allow orthopaedic surgeons to accurately risk-stratify their patients preoperatively to achieve the most optimal outcomes.
References
1. Hamilton DF , Howie CR , Burnett R , Simpson A , Patton JT . Dealing with the predicted increase in demand for revision total knee arthroplasty: challenges, risks and opportunities . Bone Joint J . 2015 ; 97-B ( 6 ): 723 – 728 . Crossref , PubMed Google Scholar
2. Kulshrestha V , Datta B , Mittal G , Kumar S . Epidemiology of revision total knee arthroplasty: A single center’s experience . Indian J Orthop . 2019 ; 53 ( 2 ): 282 – 288 . Crossref , PubMed Google Scholar
3. Kurtz SM , Ong KL , Lau E , et al. International survey of primary and revision total knee replacement . Int Orthop . 2011 ; 35 ( 12 ): 1783 – 1789 . Crossref , PubMed Google Scholar
4. Cram P , Lu X , Kates SL , Singh JA , Li Y , Wolf BR . Total knee arthroplasty volume, utilization, and outcomes among Medicare beneficiaries, 1991-2010 . JAMA . 2012 ; 308 ( 12 ): 1227 – 1236 . Crossref , PubMed Google Scholar
5. Patel A , Pavlou G , Mújica-Mota RE , Toms AD . The epidemiology of revision total knee and hip arthroplasty in England and Wales: A comparative analysis with projections for the United States. A study using the National Joint Registry dataset . Bone Joint J . 2015 ; 97-B ( 8 ): 1076 – 1081 . Crossref , PubMed Google Scholar
6. Stirling P , Middleton SD , Brenkel IJ , Walmsley PJ . Revision total knee arthroplasty versus primary total knee arthroplasty: a matched cohort study . Bone Jt Open . 2020 ; 1 ( 3 ): 29 – 34 . Crossref , PubMed Google Scholar
7. Lindberg-Larsen M , Jørgensen CC , Bæk Hansen T , Solgaard S , Odgaard A , Kehlet H . Re-admissions, re-operations and length of stay in hospital after aseptic revision knee replacement in Denmark: a two-year nationwide study . Bone Joint J . 2014 ; 96-B ( 12 ): 1649 – 1656 . Crossref , PubMed Google Scholar
8. Berend ME , Berend KR , Lombardi AV . Advances in pain management: game changers in knee arthroplasty . Bone Joint J . 2014 ; 96-B ( 11 Supple A ): 7 – 9 . Crossref , PubMed Google Scholar
9. Krauss ES , Cronin M , Suratwala SJ , Enker P , Rosen L , Segal A . Use of intravenous tranexamic acid improves early ambulation after total knee arthroplasty and anterior and posterior total hip arthroplasty . Am J Orthop (Belle Mead NJ) . 2017 ; 46 ( 5 ): E314 – E319 . PubMed Google Scholar
10. Sutton R , Chisari E , Scaramella A , Krueger CA , Courtney PM . Total hip and knee revisions are really outpatient procedures? Implications of the removal from the inpatient only list . J Arthroplasty . 2022 ; 37 ( 8S ): S732 – S737 . Crossref , PubMed Google Scholar
11. Law JI , Adams JB , Berend KR , Lombardi AV , Crawford DA . The feasibility of outpatient revision total knee arthroplasty in selected case scenarios . J Arthroplasty . 2020 ; 35 ( 6S ): S92 – S96 . Crossref , PubMed Google Scholar
12. Buller LT , Hubbard TA , Ziemba-Davis M , Deckard ER , Meneghini RM . Safety of same and next day discharge following revision hip and knee arthroplasty using modern perioperative protocols . J Arthroplasty . 2021 ; 36 ( 1 ): 30 – 36 . Crossref , PubMed Google Scholar
13. Kort NP , Bemelmans YFL , van der Kuy PHM , Jansen J , Schotanus MGM . Patient selection criteria for outpatient joint arthroplasty . Knee Surg Sports Traumatol Arthrosc . 2017 ; 25 ( 9 ): 2668 – 2675 . Crossref , PubMed Google Scholar
14. Linardatos P , Papastefanopoulos V , Kotsiantis S . Explainable AI: A review of machine learning interpretability methods . Entropy (Basel) . 2020 ; 23 ( 1 ): 18 . Crossref , PubMed Google Scholar
15. Wei C , Quan T , Wang KY , et al. Artificial neural network prediction of same-day discharge following primary total knee arthroplasty based on preoperative and intraoperative variables . Bone Joint J . 2021 ; 103-B ( 8 ): 1358 – 1366 . Crossref , PubMed Google Scholar
16. Ingraham AM , Cohen ME , Bilimoria KY , et al. Association of surgical care improvement project infection-related process measure compliance with risk-adjusted outcomes: implications for quality measurement . J Am Coll Surg . 2010 ; 211 ( 6 ): 705 – 714 . Crossref , PubMed Google Scholar
17. Malik AT , Scharschmidt TJ , Li M , Jain N , Khan SN . Are Joint Surgeons Being Adequately Compensated for Single-Component versus Double-Component Revision TKA? An Analysis of Relative Value Units . J Knee Surg . 2020 ; 33 ( 6 ): 593 – 596 . Crossref , PubMed Google Scholar
18. Saklad M . Grading of patients for surgical procedures . Anesthesiol . 1941 ; 2 ( 3 ): 281 – 284 . Google Scholar
19. Gu A , Malahias M-A , Selemon NA , et al. Increased severity of anaemia is associated with 30-day complications following total joint replacement . Bone Joint J . 2020 ; 102-B ( 4 ): 485 – 494 . Crossref , PubMed Google Scholar
20. Hart A , Khalil JA , Carli A , Huk O , Zukor D , Antoniou J . Blood transfusion in primary total hip and knee arthroplasty. Incidence, risk factors, and thirty-day complication rates . J Bone Joint Surg Am . 2014 ; 96-A ( 23 ): 1945 – 1951 . Crossref , PubMed Google Scholar
21. No authors listed . National Surgical Quality Improvement Program . American College of Surgeons . https://www.facs.org/quality-programs/data-and-registries/acs-nsqip/ ( date last accessed 2 May 2023 ). Google Scholar
22. Bovonratwet P , Webb ML , Ondeck NT , et al. Definitional differences of “outpatient” versus “inpatient” THA and TKA can affect study outcomes . Clin Orthop Relat Res . 2017 ; 475 ( 12 ): 2917 – 2925 . Crossref , PubMed Google Scholar
23. Stekhoven DJ , Bühlmann P . MissForest--non-parametric missing value imputation for mixed-type data . Bioinformatics . 2012 ; 28 ( 1 ): 112 – 118 . Crossref , PubMed Google Scholar
24. Boulesteix A-L , Janitza S , Hornung R , Probst P , Busen H , Hapfelmeier A . Making complex prediction rules applicable for readers: Current practice in random forest literature and recommendations . Biom J . 2019 ; 61 ( 5 ): 1314 – 1328 . Crossref , PubMed Google Scholar
25. Couronné R , Probst P , Boulesteix A-L . Random forest versus logistic regression: a large-scale benchmark experiment . BMC Bioinformatics . 2018 ; 19 ( 1 ): 270 . Crossref , PubMed Google Scholar
26. Oosterhoff JHF , Karhade AV , Oberai T , Franco-Garcia E , Doornberg JN , Schwab JH . Prediction of postoperative delirium in geriatric hip fracture patients: A clinical prediction model using machine learning algorithms . Geriatr Orthop Surg Rehabil . 2021 ; 12 : 21514593211062276 . Crossref , PubMed Google Scholar
27. Mandrekar JN . Receiver operating characteristic curve in diagnostic test assessment . J Thorac Oncol . 2010 ; 5 ( 9 ): 1315 – 1316 . Crossref , PubMed Google Scholar
28. Nattino G , Finazzi S , Bertolini G . A new calibration test and A reappraisal of the calibration belt for the assessment of prediction models based on dichotomous outcomes . Stat Med . 2014 ; 33 ( 14 ): 2390 – 2407 . Crossref , PubMed Google Scholar
29. Altmann A , Toloşi L , Sander O , Lengauer T . Permutation importance: a corrected feature importance measure . Bioinformatics . 2010 ; 26 ( 10 ): 1340 – 1347 . Crossref , PubMed Google Scholar
30. Mogensen UB , Ishwaran H , Gerds TA . Evaluating random forests for survival analysis using prediction error curves . J Stat Softw . 2012 ; 50 ( 11 ): 1 – 23 . Crossref , PubMed Google Scholar
31. Lebedev AV , Westman E , Van Westen GJP , et al. Random Forest ensembles for detection and prediction of Alzheimer’s disease with a good between-cohort robustness . Neuroimage Clin . 2014 ; 6 : 115 – 125 . Crossref , PubMed Google Scholar
32. Garbarino LJ , Gold PA , Sodhi N , et al. The effect of operative time on in-hospital length of stay in revision total knee arthroplasty . Ann Transl Med . 2019 ; 7 ( 4 ): 66 . Crossref , PubMed Google Scholar
33. Belmont PJ , Goodman GP , Waterman BR , Bader JO , Schoenfeld AJ . Thirty-day postoperative complications and mortality following total knee arthroplasty: incidence and risk factors among a national sample of 15,321 patients . J Bone Joint Surg Am . 2014 ; 96-A ( 1 ): 20 – 26 . Crossref , PubMed Google Scholar
34. Bohl DD , Ondeck NT , Darrith B , Hannon CP , Fillingham YA , Della Valle CJ . Impact of operative time on adverse events following primary total joint arthroplasty . J Arthroplasty . 2018 ; 33 ( 7 ): 2256 – 2262 . Crossref , PubMed Google Scholar
35. Dai WL , Lin ZM , Shi ZJ , Wang J . Outcomes following revision total knee arthroplasty septic versus aseptic failure: A national propensity-score-matched comparison . J Knee Surg . 2021 ; 34 ( 11 ): 1227 – 1236 . Crossref , PubMed Google Scholar
36. Wilson JM , Farley KX , Erens GA , Guild GN . General vs spinal anesthesia for revision total knee arthroplasty: Do complication rates differ? J Arthroplasty . 2019 ; 34 ( 7 ): 1417 – 1422 . Crossref , PubMed Google Scholar
37. Pugely AJ , Martin CT , Gao Y , Mendoza-Lattes S , Callaghan JJ . Differences in short-term complications between spinal and general anesthesia for primary total knee arthroplasty . J Bone Joint Surg Am . 2013 ; 95-A ( 3 ): 193 – 199 . Crossref , PubMed Google Scholar
38. Holte K , Foss NB , Svensén C , Lund C , Madsen JL , Kehlet H . Epidural anesthesia, hypotension, and changes in intravascular volume . Anesthesiology . 2004 ; 100 ( 2 ): 281 – 286 . Crossref , PubMed Google Scholar
39. Carpenter RL , Caplan RA , Brown DL , Stephenson C , Wu R . Incidence and risk factors for side effects of spinal anesthesia . Anesthesiology . 1992 ; 76 ( 6 ): 906 – 916 . Crossref , PubMed Google Scholar
40. Kim JS , Merrill RK , Arvind V , et al. Examining the ability of artificial neural networks machine learning models to accurately predict complications following posterior lumbar spine fusion . Spine (Phila Pa 1976) . 2018 ; 43 ( 12 ): 853 – 860 . Crossref , PubMed Google Scholar
41. Li Q , Zhong H , Girardi FP , et al. Machine learning approaches to define candidates for ambulatory single level laminectomy surgery . Global Spine J . 2022 ; 12 ( 7 ): 1363 – 1368 . Crossref , PubMed Google Scholar
42. Zhong H , Poeran J , Gu A , et al. Machine learning approaches in predicting ambulatory same day discharge patients after total hip arthroplasty . Reg Anesth Pain Med . 2021 ; 46 ( 9 ): 779 – 783 . Crossref , PubMed Google Scholar
43. Deng F , Huang J , Yuan X , Cheng C , Zhang L . Performance and efficiency of machine learning algorithms for analyzing rectangular biomedical data . Lab Invest . 2021 ; 101 ( 4 ): 430 – 441 . Crossref , PubMed Google Scholar
44. DeCook CA . Outpatient joint arthroplasty: Transitioning to the ambulatory surgery center . J Arthroplasty . 2019 ; 34 ( 7S ): S48 – S50 . Crossref , PubMed Google Scholar
45. Courtney PM , West ME , Hozack WJ . Maximizing physician-hospital alignment: Lessons learned from effective models of joint arthroplasty care . J Arthroplasty . 2018 ; 33 ( 6 ): 1641 – 1646 . Crossref , PubMed Google Scholar
Author contributions
T. Yeramosu: Conceptualization, Data curation, Formal analysis, Investigation, Methodology, Validation, Writing – review & editing.
W. Ahmad: Conceptualization, Data curation, Formal analysis, Investigation, Methodology, Writing – review & editing.
J. Satpathy: Conceptualization, Investigation, Project administration, Resources, Supervision, Validation, Writing –review & editing.
J. M. Farrar: Conceptualization, Investigation, Resources, Supervision, Writing – review & editing.
G. J. Golladay: Conceptualization, Investigation, Project administration, Resources, Supervision, Validation, Writing – review & editing.
N. K. Patel: Conceptualization, Investigation, Project administration, Resources, Supervision, Validation, Writing – review & editing.
Funding statement
The authors received no financial or material support for the research, authorship, and/or publication of this article.
ICMJE COI statement
T. Yeramosu, W. Ahmad, J. Satpathy, and J. M. Farrar have no conflicts of interest to report. G. J. Golladay reports royalties from Stryker, and speaker payments or honoraria from Cerus, unrelated to this study. G. J. Golladay is also on the editorial/governing board of Arthroplasty Today and Journal of Arthroplasty, and a Board member of American Association of Hip and Knee Surgeons publications committee and Virginia Orthopaedic Society. N. K. Patel is on the editorial board of Journal of Arthroplasty, and is Program Co-Chair of the Virginia Orthopaedic Society.
Data sharing
The datasets generated and analyzed in the current study are not publicly available due to data protection regulations. Access to data is limited to the researchers who have obtained permission for data processing. Further inquiries can be made to the corresponding author.
Ethical review statement
This study did not undergo IRB review due to the public, de-identified nature of the database.
Open access funding
The authors confirm that the open access fee for this study was funded by the Virginia Commonwealth University Department of Orthopaedic Surgery.
© 2023 Author(s) et al. This is an open-access article distributed under the terms of the Creative Commons Attribution Non-Commercial No Derivatives (CC BY-NC-ND 4.0) licence, which permits the copying and redistribution of the work only, and provided the original author and source are credited. See https://creativecommons.org/licenses/by-nc-nd/4.0/
