Machine-learning Models Predict 30-Day Mortality, Cardiovascular Complications, and Respiratory Complications After Aseptic Revision Total Joint Arthroplasty.

BACKGROUND: Aseptic revision THA and TKA are associated with an increased risk of adverse outcomes compared with primary THA and TKA. Understanding the risk profiles for patients undergoing aseptic revision THA or TKA may provide an opportunity to decrease the risk of postsurgical complications. There are risk stratification tools for postoperative complications after aseptic revision TKA or THA; however, current tools only include nonmodifiable risk factors, such as medical comorbidities, and do not include modifiable risk factors. QUESTIONS/PURPOSES: (1) Can machine learning predict 30-day mortality and complications for patients undergoing aseptic revision THA or TKA using a cohort from the American College of Surgeons National Surgical Quality Improvement Program database? (2) Which patient variables are the most relevant in predicting complications? METHODS: This was a temporally validated, retrospective study analyzing the 2014 to 2019 National Surgical Quality Improvement Program database, as this database captures a large cohort of aseptic revision THA and TKA patients across a broad range of clinical settings and includes preoperative laboratory values. The training data set was 2014 to 2018, and 2019 was the validation data set. Given that predictive models learn expected prevalence of outcomes, this split allows assessment of model performance in contemporary patients. Between 2014 and 2019, a total of 24,682 patients underwent aseptic revision TKA and 17,871 patients underwent aseptic revision THA. Of those, patients with CPT codes corresponding to aseptic revision TKA or THA were considered as potentially eligible. Based on excluding procedures involving unclean wounds, 78% (19,345 of 24,682) of aseptic revision TKA procedures and 82% (14,711 of 17,871) of aseptic revision THA procedures were eligible. Ten percent of patients in each of the training and validation cohorts had missing predictor variables. Most of these missing data were preoperative sodium or hematocrit (8% in both the training and validation cohorts). No patients had missing outcome data. No patients were excluded due to missing data. The mean patient was age 66 ± 12 years, the mean BMI was 32 ± 7 kg/m 2 , and the mean American Society of Anesthesiologists (ASA) Physical Score was 3 (56%). XGBoost was then used to create a scoring tool for 30-day adverse outcomes. XGBoost was chosen because it can handle missing data, it is nonlinear, it can assess nuanced relationships between variables, it incorporates techniques to reduce model complexity, and it has a demonstrated record of producing highly accurate machine-learning models. Performance metrics included discrimination and calibration. Discrimination was assessed by c-statistics, which describe the area under the receiver operating characteristic curve. This quantifies how well a predictive model discriminates between patients who have the outcome of interest versus those who do not. Relevant ranges for c-statistics include good (0.70 to 0.79), excellent (0.80 to 0.89), and outstanding (> 0.90). We estimated 95% confidence intervals (CIs) for c-statistics by 500-sample bootstrapping. Calibration curves quantify reliability of model predictions. Reliable models produce prediction probabilities for outcomes that are similar to observed probabilities of those outcomes, so a well-calibrated model should demonstrate a calibration curve that does not deviate substantially from a line of slope 1 and intercept 0. Calibration curves were generated on the 2019 validation data. Shapley Additive Explanations (SHAP) visualizations were used to investigate feature importance to gain insight into how models made predictions. The models were built into an online calculator for ongoing testing and validation. The risk calculator, which is freely available ( http://nb-group.org/rev2/ ), allows a user to input patient data to calculate postoperative risk of 30-day mortality, cardiac, and respiratory complications after aseptic revision TKA or THA. A post hoc analysis was performed to assess whether using data from 2020 would improve calibration on 2019 data. RESULTS: The model accurately predicted mortality, cardiac complications, and respiratory complications after aseptic revision THA or TKA, with c-statistics of 0.88 (95% CI 0.83 to 0.93), 0.80 (95% CI 0.75 to 0.84), and 0.78 (95% CI 0.74 to 0.82), respectively, on internal validation and 0.87 (95% CI 0.77 to 0.96), 0.70 (95% CI 0.61 to 0.78), and 0.82 (95% CI 0.75 to 0.88), respectively, on temporal validation. Calibration curves demonstrated slight over-confidence in predictions (most predicted probabilities were higher than observed probabilities). Post hoc analysis of 2020 data did not yield improved calibration on the 2019 validation set. Important risk factors for all models included increased age and higher ASA, BMI, hematocrit level, and sodium level. Hematocrit and ASA were in the top three most important features for all models. The factor with the strongest association for mortality and cardiac complication models was age, and for the respiratory model, chronic obstructive pulmonary disease. Risk related to sodium followed a U-shaped curve. Preoperative hyponatremia and hypernatremia predicted an increased risk of mortality and respiratory complications, with a nadir of 138 mmol/L; hyponatremia was more strongly associated with mortality than hypernatremia. A hematocrit level less than 36% predicted an increased risk of all three adverse outcomes. A BMI less than 24 kg/m 2 -and especially less than 20 kg/m 2 -predicted an increased risk of all three adverse outcomes, with little to no effect for higher BMI. CONCLUSION: This temporally validated model predicted 30-day mortality, cardiac complications, and respiratory complications after aseptic revision THA or TKA with c-statistics ranging from 0.78 to 0.88. This freely available risk calculator can be used preoperatively by surgeons to educate patients on their individual postoperative risk of these specific adverse outcomes. Unanswered questions that remain include whether altering the studied preoperative patient variables, such as sodium or hematocrit, would affect postoperative risk of adverse outcomes; however, a prospective cohort study is needed to answer this question. LEVEL OF EVIDENCE: Level III, therapeutic study.

A Machine-Learning Algorithm to Predict the Likelihood of Prolonged Opioid Use Following Arthroscopic Hip Surgery.

PURPOSE: To develop a machine-learning algorithm and clinician-friendly tool predicting the likelihood of prolonged opioid use (>90 days) following hip arthroscopy. METHODS: The Military Data Repository was queried for all adult patients undergoing arthroscopic hip surgery between 2012 and 2017. Demographic, health history, and prescription records were extracted for all included patients. Opioid use was divided into preoperative use (30-365 days before surgery), perioperative use (30 days before surgery through 14 days after surgery), postoperative use (14-90 days after surgery), and prolonged postoperative use (90-365 days after surgery). Six machine-learning algorithms (Naïve Bayes, Gradient Boosting Machine, Extreme Gradient Boosting, Random Forest, Elastic Net Regularization, and artificial neural network) were developed. Area under the receiver operating curve and Brier scores were calculated for each model. Decision curve analysis was applied to assess clinical utility. Local-Interpretable Model-Agnostic Explanations were used to demonstrate factor weights within the selected model. RESULTS: A total of 6,760 patients were included, of whom 2,762 (40.9%) filled at least 1 opioid prescription >90 days after surgery. The artificial neural network model showed superior discrimination and calibration with area under the receiver operating curve = 0.71 (95% confidence interval 0.68-0.74) and Brier score = 0.21 (95% confidence interval 0.20-0.22). Postsurgical opioid use, age, and preoperative opioid use had the most influence on model outcome. Lesser factors included the presence of a psychological comorbidity and strong history of a substance use disorder. CONCLUSIONS: The artificial neural network model shows sufficient validity and discrimination for use in clinical practice. The 5 identified factors (age, preoperative opioid use, postoperative opioid use, presence of a mental health comorbidity, and presence of a preoperative substance use disorder) accurately predict the likelihood of prolonged opioid use following hip arthroscopy. LEVEL OF EVIDENCE: III, retrospective comparative prognostic trial.

Machine Learning to Predict Fascial Dehiscence after Exploratory Laparotomy Surgery.

BACKGROUND: Fascial dehiscence following exploratory laparotomy is associated with significant morbidity and increased mortality. Previously published risk prediction models for fascial dehiscence are dated and limit a surgeon's ability to perform reliable risk assessment intraoperatively. We sought to determine if machine learning can predict fascial dehiscence after exploratory laparotomy. MATERIALS AND METHODS: A retrospective cohort study was conducted of 93,024 patients undergoing exploratory laparotomy from the 2011-2018 ACS NSQIP data files. Data were divided into training (2011-2016, n = 69,969) and temporal validation (2017-2018, n = 23,055) cohorts. A clinical decision support tool was developed using the model generated via machine learning techniques. RESULTS: 1,332 (1.9%) patients in the training cohort and 390 (1.7%) patients in the temporal validation cohort developed fascial dehiscence. The area under the receiver operating characteristic curve was 0.69 (95% CI 0.66 to 0.72) in the validation cohort. Model predictions demonstrated excellent probability calibration. Decision curve analysis calculates net clinical benefit within a threshold range of 0.8%-4.5%. Operative time, surgical site and deep space infections, and body mass index were among the most important features for model predictions. Finally, operative time, sodium level, and hematocrit demonstrated non-linear relationships with predicted risk. CONCLUSION: A clinical decision support tool for predicting fascial dehiscence after exploratory laparotomy was created and validated on a contemporary, national patient cohort using machine learning. The tool calculates net clinical benefit and can be used at the point of care. Some identified risk factor relationships were found to be complex and non-linear, highlighting the ability of some machine learning applications to capture nuanced, patient-specific risk profiles.

Competency-Based Assessments: Leveraging Artificial Intelligence to Predict Subcompetency Content.

PURPOSE: Faculty feedback on trainees is critical to guiding trainee progress in a competency-based medical education framework. The authors aimed to develop and evaluate a Natural Language Processing (NLP) algorithm that automatically categorizes narrative feedback into corresponding Accreditation Council for Graduate Medical Education Milestone 2.0 subcompetencies. METHOD: Ten academic anesthesiologists analyzed 5,935 narrative evaluations on anesthesiology trainees at 4 graduate medical education (GME) programs between July 1, 2019, and June 30, 2021. Each sentence (n = 25,714) was labeled with the Milestone 2.0 subcompetency that best captured its content or was labeled as demographic or not useful. Inter-rater agreement was assessed by Fleiss' Kappa. The authors trained an NLP model to predict feedback subcompetencies using data from 3 sites and evaluated its performance at a fourth site. Performance metrics included area under the receiver operating characteristic curve (AUC), positive predictive value, sensitivity, F1, and calibration curves. The model was implemented at 1 site in a self-assessment exercise. RESULTS: Fleiss' Kappa for subcompetency agreement was moderate (0.44). Model performance was good for professionalism, interpersonal and communication skills, and practice-based learning and improvement (AUC 0.79, 0.79, and 0.75, respectively). Subcompetencies within medical knowledge and patient care ranged from fair to excellent (AUC 0.66-0.84 and 0.63-0.88, respectively). Performance for systems-based practice was poor (AUC 0.59). Performances for demographic and not useful categories were excellent (AUC 0.87 for both). In approximately 1 minute, the model interpreted several hundred evaluations and produced individual trainee reports with organized feedback to guide a self-assessment exercise. The model was built into a web-based application. CONCLUSIONS: The authors developed an NLP model that recognized the feedback language of anesthesiologists across multiple GME programs. The model was operationalized in a self-assessment exercise. It is a powerful tool which rapidly organizes large amounts of narrative feedback.

Pulse Arrival Time Is Associated With Hemorrhagic Volume in a Porcine Model: A Pilot Study.

BACKGROUND: Hemorrhage is a major cause of preventable death worldwide, and early identification can be lifesaving. Pulse wave contour analysis has previously been used to infer hemodynamic variables in a variety of settings. We hypothesized that pulse arrival time (PAT), a form of pulse wave contour analysis which is assessed via electrocardiography (ECG) and photoplethysmography (PPG), is associated with hemorrhage volume. METHODS: Yorkshire-Cross swine were randomized to hemorrhage (30 mL/kg over 20 minutes) vs. control. Continuous ECG and PPG waveforms were recorded with a novel monitoring device, and algorithms were developed to calculate PAT and PAT variability throughout the respiratory cycle, termed "PAT index" or "PAT_I." Mixed effects models were used to determine associations between blood loss and PAT and between blood loss and PAT_I to account for clustering within subjects and investigate inter-subject variability in these relationships. RESULTS: PAT and PAT_I data were determined for ∼150 distinct intervals from five subjects. PAT and PAT_I were strongly associated with blood loss. Mixed effects modeling with PAT alone was substantially better than PAT_I alone (R2 0.93 vs. 0.57 and Akaike information criterion (AIC) 421.1 vs. 475.5, respectively). Modeling blood loss with PAT and PAT_I together resulted in slightly improved fit compared to PAT alone (R2 0.96, AIC 419.1). Mixed effects models demonstrated significant inter-subject variability in the relationships between blood loss and PAT. CONCLUSIONS: Findings from this pilot study suggest that PAT and PAT_I may be used to detect blood loss. Because of the simple design of a single-lead ECG and PPG, the technology could be packaged into a very small form factor device for use in austere or resource-constrained environments. Significant inter-subject variability in the relationship between blood loss and PAT highlights the importance of individualized hemodynamic monitoring.

Hemodynamic Effects of Cardiovascular Medications in a Normovolemic and Hemorrhaged Yorkshire-cross Swine Model.

The Yorkshire-cross swine model is a valuable translational model commonly used to study cardiovascular physiology and response to insult. Although the effects of vasoactive medications have been well described in healthy swine, the effects of these medications during hemorrhagic shock are less studied. In this study, we sought to expand the utility of the swine model by characterizing the hemodynamic changes that occurred after the administration of commonly available vasoactive medications during euvolemic and hypovolemic states. To this end, we anesthetized and established femoral arterial, central venous, and pulmonary arterial access in 15 juvenile Yorkshire-cross pigs. The pigs then received a series of rapidly metabolized but highly vasoactive medications in a standard dosing sequence. After completion of this sequence, each pig underwent a 30-mL/kg hemorrhage over 10 min, and the standard dosing sequence was repeated. We then used standard sta- tistical techniques to compare the effects of these vasoactive medications on a variety of hemodynamic parameters between the euvolemic and hemorrhagic states. All subjects completed the study protocol. The responses in the hemorrhagic state were often attenuated or even opposite of those in the euvolemic state. For example, phenylephrine decreased the mean arterial blood pressure during the euvolemic state but increased it in the hemorrhagic state. These results clarify previously poorly defined responses to commonly used vasoactive agents during the hemorrhagic state in swine. Our findings also demonstrate the need to consider the complex and dynamic physiologic state of hemorrhage when anticipating the effects of vasoactive drugs and planning study protocols.

Preemptive analgesia after lumbar spine surgery by pregabalin and celecoxib: a prospective study.

OBJECTIVE: To evaluate the preemptive analgesic effect of combination pregabalin with celecoxib for lumbar spine surgery. METHODS: A prospective, randomized study was conducted among 60 lumbar spine surgery patients and divided into two groups. Postoperative pain relief was achieved with intravenous patient-controlled analgesia with morphine. The preemptive analgesia group received oral pregabalin (150 mg) and celecoxib (200 mg) 2 hrs before surgery, and the control group received a placebo. Pain was assessed by visual analogue scale (VAS). Side effects and morphine consumption were monitored until 48 hrs after surgery. RESULTS: VAS score at rest and during movement was statistically significantly lower in the preemptive analgesia group at most time points (p<0.05). Morphine consumption was significantly lower in the preemptive analgesia group compared with control group in the 24 first hours (29.03±4.38 mg vs 24.43±4.94) and 48 hrs (52.23±9.57 mg vs 44.20±10.21 mg), p<0.05. Hemodynamics, respiratory rate, and SpO2 were similar for both groups. The sedation score was only statistically significant at H8 time point. The incidence of nausea/vomiting in the preemptive group did not statistically differ from the control group. CONCLUSION: Preoperative administration of pregabalin combined with celecoxib had a good preemptive analgesia effect and reduced intravenous morphine consumption after lumbar spine surgery. Side effects were mild and transient.

Patient-controlled paravertebral analgesia for video-assisted thoracoscopic surgery lobectomy.

BACKGROUND: Paravertebral block has been proven to be an efficient method to provide post-thoracotomy pain management. This study aimed to compare patient-controlled paravertebral analgesia (PCPA) and intravenous patient-controlled analgesia (IVPCA) in terms of analgesic efficiency, respiratory function, and adverse effects after video-assisted thoracoscopic surgery (VATS) lobectomy. PATIENTS AND METHODS: The prospective randomized trial study was carried out on 60 patients who underwent VATS lobectomy (randomly allocated 30 patients in each group). In the PCPA group, an initial dose of 0.3 mL/kg of 0.125% bupivacaine with fentanyl 2 µg/mL was administered, followed by a 3 mL/h continuous infusion with patient-controlled analgesia (2 mL bolus, 10-minute lockout interval, 25 mL/4 h limit). In the IVPCA group with morphine 1 mg/mL solution, an infusion device was programmed to deliver a 1.0 mL demand bolus with no basal infusion rate, with a 10-minute lockout interval and a maximum of 20 mL/4 h period. Postoperative pain was assessed by visual analog scale at rest and on coughing. Arterial blood gas and spirometry were monitored and recorded for the first 3 postoperative days. Side effects to include were also recorded. RESULTS: The PCPA group had statistically significant lower pain scores (P<0.0001) at rest at all times. Lower pain scores on coughing were statistically significant in PCPA group in the first 4 hours. Postoperative spirometry showed that both the groups had comparable recovery trajectories for their pulmonary function. Arterial blood gas analysis showed pH and PaCO2 were in a normal range in both the groups. The incidence of headache was higher in the IVPCA group (13.3% vs 0%; P=0.038). CONCLUSION: PCPA effectively managed pain after VATS lobectomy, with lower pain scores, similar respiratory function, and fewer side effects than standard IVPCA treatment.