Background:Traditional clinical trial enrollment relies on manual screening and coordinator-led recruitment, creating scalability barriers in high-volume perioperative environments. This study evaluated whether a fully automated, electronic health record (EHR)-integrated clinical decision support (CDS) system could identify eligible patients and engage clinicians in real time without manual screening or dedicated research staff. Methods:In this prospective implementation study, predefined respiratory-risk criteria were computed within the UCLA Perioperative Data Warehouse and transmitted to the EHR via Healthcare Level Seven interfaces. Patients meeting inclusion criteria automatically triggered Best Practice Advisories (BPAs) recommending an intervention. Outcomes included system accuracy in eligibility identification, provider adherence to BPA recommendations, and technical performance metrics. Results:The automated system processed 10 592 eligible patients and achieved 51.2% provider adherence (5424 patients) to CDS prompts without coordinator involvement. BPA allocation accuracy was 69.7% among patients recovering in the post-anesthesia care unit and 59.4% when including unanticipated ICU transfers. Adherence varied significantly by care team composition, with full teams (attending + CRNA + resident) achieving 57.4% adherence compared with 42.2% for solo attendings. Workflow factors were stronger predictors of adherence than patient clinical characteristics, indicating minimal selection bias. Conclusions:Fully automated, EHR-integrated CDS can enable large-scale, workflow-embedded enrollment into implementation-focused studies. While not a substitute for research designs requiring consent or randomization, this framework demonstrates a scalable approach for automated prescreening and CDS-driven prompting that reduces reliance on coordinator-dependent processes and supports real-world implementation science.

Background: Machine learning prediction models require prospective validation to ensure implementation fidelity and feasibility. Our primary objective was to prospectively validate a previously reported postoperative mortality prediction model in inpatients undergoing surgery. Our secondary objective was to evaluate feasibility of a pilot clinical decision support tool. Methods: We prospectively validated and implemented a random forest machine learning model trained to predict in-hospital mortality using data from a single academic medical centre. A reduced 32-feature model was implemented into the electronic health record (EHR) using a real-time data mart at the same institution. To assess model performance, the area under the receiver operating characteristic curve (AUROC), area under the curve precision-recall (AUCPR), and other performance measures were calculated. To assess feasibility, implementation workflow metrics were evaluated and a survey was administered to anaesthesiologists trained to use the pilot clinical decision support tool. Results: The AUROC for the prospectively implemented model was 0.874 (95% confidence interval [CI] 0.860-0.887), and the AUCPR was 0.111. By comparison, the AUROC for the 58-feature model was 0.925 (95% CI 0.900-0.947), and for ASA physical status the AUROC was 0.814 (95% CI 0.802-0.827) and the AUCPR was 0.103. The implementation demonstrated feasibility through real-time data updates, automated transfer of model outputs to the EHR, and provider survey entries. Conclusions: This prospective validation and EHR implementation of a previously published random forest machine learning model predicting postoperative in-hospital mortality demonstrated acceptable real-world performance of the implemented model and feasibility of integrating such a system into clinical practice.

Prolonged invasive mechanical ventilation and extubation failure are common and are associated with increased mortality, length of stay, and costs. Current studies involving airway events usually rely on manual chart review or are dependent on a single data source and typically work because the study populations are limited to homogeneous cohorts with consistent documentation workflow. To support quality improvement and research efforts addressing airway-related outcomes, ways to accurately and reliably identify airway events in electronic health records (EHR) are critical. We developed and validated an automated SQL-based algorithm to obtain high-quality airway event data that performs across diverse clinical settings within a single health system. Using data from all adult and pediatric patients at two large quaternary care hospitals from 2013 to 2025, the algorithm integrated multiple data sources to identify airway events, including ventilator flowsheets, tracheostomy procedures, anesthesia records, admission tables, and lines/drains/airway tables. Data quality was graded using a grading scale that incorporated both the number of data sources and temporal concordance between sources, with higher source count and closer temporal agreement indicating higher data quality (range 1 A [highest] to 3 C [lowest]). Qualitative validation was performed through manual review of 400 records. During the 12-year period, 285,157 unique airway events were identified. Of these, 11.8

Background:The risk of developing a persistent reduction in renal function after postoperative acute kidney injury (pAKI) is not well established. The goal of this investigation was to evaluate whether patients who develop pAKI have a greater decline in long-term renal function than patients who do not. Methods:In this multicentre retrospective propensity-matched study, anaesthesia data warehouses at three tertiary care hospitals were queried. Adult patients undergoing surgery with available preoperative and postoperative creatinine results and without baseline haemodialysis requirements were included. Patients were stratified by occurrence of pAKI as defined by the Acute Kidney Injury Network classification. The primary outcome was a decline in follow-up glomerular filtration rate (GFR) of 40% relative to baseline, based on follow-up outpatient visits from 0 to 36 months after hospital discharge. A propensity score-matched sample was used in Kaplan-Meier analysis and a piecewise Cox model to compare the time to reach a 40% decline in GFR for patients with and without pAKI. Results:In 95 213 patients, the rate of pAKI ranged from 9.9% to 13.7%. In the piecewise Cox model, pAKI was associated with a significantly increased hazard of a 40% decline in GFR. The common-effect hazard ratio was 13.35 (95% confidence interval [CI] 10.79-16.51, P<0.001) for 0-6 months, 7.07 (5.52-9.05, P<0.001) for 6-12 months, 6.02 (4.69-7.74, P<0.001) for 12-24 months, and 4.32 (2.65-7.05, P<0.001) for 24-36 months. Conclusions:pAKI is associated with a significantly increased hazard of a 40% decline in GFR up to 36 months after surgery across three institutions.

Objective: To develop and evaluate an open-source machine learning (ML) models for predicting hospital short stays (length of stay [LOS] under 48 and 72 hours) exclusively using data available at the time of ED admission, with a novel application of target encoding diagnostic codes. Materials and Methods: We trained two ML algorithms (Random Forest and XGBoost) on electronic health record (EHR) data from two hospitals to predict hospital short stays. We employed an innovative weighted target encoding method that converted categorical International Classification of Disease (ICD-10) codes into numeric representations of their probabilistic contribution to LOS. We measured area under the receiver operating characteristic curve (AUC) for correctly predicting LOS under 48 or 72 hours, which we compared to logistic regression. Results: The final sample included 8,693 adult patients admitted to an internal medicine service. Random Forest models achieved the highest performance for predicting LOS under 48 hours (AUROC=0.96, 95% CI 0.95-0.97; accuracy=91%) and under 72 hours (AUROC=0.94, 95% CI 0.93-0.95; accuracy=88%). These models outperformed logistic regression using the same features (48-hour AUROC=0.57, 95% CI 0.54-0.59 and accuracy=70%; 72-hour AUROC=0.59, 95% CI 0.57-0.61 and accuracy=56%). Discussion: Leveraging an innovative target encoding method, the Short Hospitalization Prediction (SHoP) model substantially outperforms previous ML approaches in accurately predicting LOS under both 48 and 72 hours using only ED pre-admission data (AUC 0.94-0.96). Conclusion: The technical innovation and predictive capability of the SHoP model enables powerful, real-time applications for optimizing patient flow and hospital resource utilization by identifying potentially divertible admissions while patients are still in the ED. ### Competing Interest Statement The authors have declared no competing interest. ### Funding Statement Dr. Leuchter is supported by funding from the NIH-NHLBI. Dr. Gabel is a co-founder of Extrico Health Inc., a healthcare analytics company. The funders/companies had no role in considering the study design or in the collection, analysis, interpretation of data, writing of the report, or decision to submit the article for publication. ### Author Declarations I confirm all relevant ethical guidelines have been followed, and any necessary IRB and/or ethics committee approvals have been obtained. Yes The details of the IRB/oversight body that provided approval or exemption for the research described are given below: The UCLA IRB determined that this secondary analysis of de-identified data did not constitute human subjects research so did not require review I confirm that all necessary patient/participant consent has been obtained and the appropriate institutional forms have been archived, and that any patient/participant/sample identifiers included were not known to anyone (e.g., hospital staff, patients or participants themselves) outside the research group so cannot be used to identify individuals. Yes I understand that all clinical trials and any other prospective interventional studies must be registered with an ICMJE-approved registry, such as ClinicalTrials.gov. I confirm that any such study reported in the manuscript has been registered and the trial registration ID is provided (note: if posting a prospective study registered retrospectively, please provide a statement in the trial ID field explaining why the study was not registered in advance). Yes I have followed all appropriate research reporting guidelines, such as any relevant EQUATOR Network research reporting checklist(s) and other pertinent material, if applicable. Yes Data was never abstracted for this study; training and testing of models was done on a secure cloud-based sever that prohibits the authors from extracting data for distribution. Thus, it is not technically feasible to share the training or test data. The code used to train the models is freely available on Github and open source, as noted in the manuscript text.

Background:Cerebral vasospasm (CV) is a feared complication occurring in 20-40% of patients following subarachnoid hemorrhage (SAH) and is known to contribute to delayed cerebral ischemia. It is standard practice to admit SAH patients to intensive care for an extended period of vigilant, resource-intensive, clinical monitoring. We used machine learning to predict CV requiring verapamil (CVRV) in the largest and only multi-center study to date. Methods:SAH patients admitted to UCLA from 2013-2022 and a validation cohort from VUMC from 2018-2023 were included. For each patient, 172 unique intensive care unit (ICU) variables were extracted through the primary endpoint, namely first verapamil administration or ICU downgrade. At each institution, a light gradient boosting machine (LightGBM) was trained using five- fold cross validation to predict the primary endpoint at various timepoints during hospital admission. Receiver-operator curves (ROC) and precision-recall (PR) curves were generated. Results:A total of 1,750 patients were included from UCLA, 125 receiving verapamil. LightGBM achieved an area under the ROC (AUC) of 0.88 an average of over one week in advance, and successfully ruled out 8% of non-verapamil patients with zero false negatives. Minimum leukocyte count, maximum platelet count, and maximum intracranial pressure were the variables with highest predictive accuracy. Our models predicted "no CVRV" vs "CVRV within three days" vs "CVRV after three days" with AUCs=0.88, 0.83, and 0.88, respectively. For external validation at VUMC, 1,654 patients were included, 75 receiving verapamil. Predictive models at VUMC performed very similarly to those at UCLA, averaging 0.01 AUC points lower. Conclusions:We present an accurate (AUC=0.88) and early (>1 week prior) predictor of CVRV using machine learning over two large cohorts of subarachnoid hemorrhage patients at separate institutions. This represents a significant step towards optimized clinical management and improved resource allocation in the intensive care setting of subarachnoid hemorrhage patients.

Abstract Background The mechanism for recording International Classification of Diseases (ICD) and diagnosis related groups (DRG) codes in a patient’s chart is through a certified medical coder who manually reviews the medical record at the completion of an admission. High-acuity ICD codes justify DRG modifiers, indicating the need for escalated hospital resources. In this manuscript, we demonstrate that value of rules-based computer algorithms that audit for omission of administrative codes and quantifying the downstream effects with regard to financial impacts and demographic findings did not indicate significant disparities. Methods All study data were acquired via the UCLA Department of Anesthesiology and Perioperative Medicine’s Perioperative Data Warehouse. The DataMart is a structured reporting schema that contains all the relevant clinical data entered into the EPIC (EPIC Systems, Verona, WI) electronic health record. Computer algorithms were created for eighteen disease states that met criteria for DRG modifiers. Each algorithm was run against all hospital admissions with completed billing from 2019. The algorithms scanned for the existence of disease, appropriate ICD coding, and DRG modifier appropriateness. Secondarily, the potential financial impact of ICD omissions was estimated by payor class and an analysis of ICD miscoding was done by ethnicity, sex, age, and financial class. Results Data from 34,104 hospital admissions were analyzed from January 1, 2019, to December 31, 2019. 11,520 (32.9%) hospital admissions were algorithm positive for a disease state with no corresponding ICD code. 1,990 (5.8%) admissions were potentially eligible for DRG modification/upgrade with an estimated lost revenue of $22,680,584.50. ICD code omission rates compared against reference groups (private payors, Caucasians, middle-aged patients) demonstrated significant p-values < 0.05; similarly significant p-value where demonstrated when comparing patients of opposite sexes. Conclusions We successfully used rules-based algorithms and raw structured EHR data to identify omitted ICD codes from inpatient medical record claims. These missing ICD codes often had downstream effects such as inaccurate DRG modifiers and missed reimbursement. Embedding augmented intelligence into this problematic workflow has the potential for improvements in administrative data, but more importantly, improvements in administrative data accuracy and financial outcomes.

Editor—Postoperative respiratory dysfunction, defined by requirements for supplemental oxygen, hypoxaemia, or tracheal reintubation, is a known potential complication of general anaesthesia. 1 International Surgical Outcomes Study groupGlobal patient outcomes after elective surgery: prospective cohort study in 27 low-, middle- and high-income countries. Br J Anaesth. 2016; 117: 601-609 Abstract Full Text Full Text PDF PubMed Scopus (359) Google Scholar ,2 Hafeez K.R. Tuteja A. Singh M. et al. Postoperative complications with neuromuscular blocking drugs and/or reversal agents in obstructive sleep apnoea patients: a systematic review. BMC Anesthesiol. 2018; 18: 91 Crossref PubMed Scopus (31) Google Scholar Respiratory dysfunction can encompass many respiratory complications occurring after emergence from anaesthesia, such as pneumonia, respiratory failure, atelectasis, and acute respiratory distress syndrome (ARDS). 3 Miskovic A. Lumb A.B. Postoperative pulmonary complications. Br J Anaesth. 2017; 118: 317-334 Abstract Full Text Full Text PDF PubMed Scopus (447) Google Scholar ,4 Jammer I. Wickboldt N. Sander M. et al. Standards for definitions and use of outcome measures for clinical effectiveness research in perioperative medicine: European Perioperative Clinical Outcome (EPCO) definitions: a statement from the ESA-ESICM joint taskforce on perioperative outcome measures. Eur J Anaesthesiol. 2015; 32: 88-105 Crossref PubMed Scopus (508) Google Scholar Risk factors include airway surgery, multiple tracheal intubation attempts, and patient factors such as pre-existing respiratory disease. 5 Nagappa M. Subramani Y. Chung F. Best perioperative practise in management of ambulatory patients with obstructive sleep apnoea. Curr Opin Anaesthesiol. 2018; 31: 700-706 Crossref Scopus (10) Google Scholar ,6 Gali B. Whalen F.X. Schroeder D.R. Gay P.C. Plevak D.J. Identification of patients at risk for postoperative respiratory complications using a preoperative obstructive sleep apnoea screening tool and postanaesthesia care assessment. Anesthesiology. 2009; 110: 869-877 Crossref PubMed Scopus (217) Google Scholar Although multifactorial, one probable contributor is continued weakness as a result of residual neuromuscular block. 7 Murphy G.S. Szokol J.W. Marymont J.H. Greenberg S.B. Avram M.J. Vender J.S. Residual neuromuscular blockade and critical respiratory events in the postanaesthesia care unit. Anesth Analg. 2008; 107: 130-137 Crossref PubMed Scopus (485) Google Scholar , 8 Patrocinio M.D. Shay D. Rudolph M.I. et al. Residual neuromuscular block prediction score versus train-of-four ratio and respiratory outcomes: a retrospective cohort study. Anesth Analg. 2021; 133: 610-619 Crossref Scopus (4) Google Scholar , 9 Murphy G.S. Szokol J.W. Avram M.J. et al. Residual neuromuscular block in the elderly: incidence and clinical implications. Anesthesiology. 2015; 123: 1322-1336 Crossref PubMed Scopus (84) Google Scholar Residual neuromuscular block occurs in 20–60% of patients in the PACU and can result from incomplete reversal prior to emergence from anaesthesia. In the immediate postoperative period, residual neuromuscular block is associated with increased postoperative hypoxaemia, atelectasis, and unplanned reintubations. 10 Grosse-Sundrup M. Henneman J.P. Sandberg W.S. et al. Intermediate acting non-depolarizing neuromuscular blocking agents and risk of postoperative respiratory complications: prospective propensity score matched cohort study. BMJ. 2012; 345: e6329 Crossref PubMed Scopus (212) Google Scholar

Background:Despite a renewed focus in recent years on pain management in the inpatient hospital setting, postoperative pain after elective craniotomy remains under investigated. This study aims to identify which perioperative factors associate most strongly with postoperative pain and opioid medication requirements after inpatient craniotomy. Materials and Methods:Using an existing dataset, we selected a restricted cohort of patients who underwent elective craniotomy surgery requiring an inpatient postoperative stay during a 7-year period at our institution (n=1832). We examined pain scores and opioid medication usage and analyzed the relative contribution of specific perioperative risk factors to postoperative pain and opioid medication intake (morphine milligram equivalents). Results:Postoperative pain was found to be highest on postoperative day 1 and decreased thereafter (up to day 5). Factors associated with greater postoperative opioid medication requirement were preoperative opioid medication use, duration of anesthesia, degree of pain in the preoperative setting, and patient age. Notably, the most significant factor associated with a higher postoperative pain score and Morphine milligram equivalents requirement was the time elapsed between the end of general anesthesia and a patient's first intravenous opioid medication. Conclusion:Postcraniotomy patients are at higher risk for requiring opioid pain medications if they have a history of preoperative opioid use, are of younger age, or undergo a longer surgery. Moreover, early requirement of intravenous opioid medications in the postoperative period should alert treating physicians that a patient's pain may require additional or alternative methods of pain control than routinely administered, to avoid over-reliance on opioid medications.

Background: Intraoperative hypotension is associated with postoperative complications. The use of vasopressors is often required to correct hypotension but the best vasopressor is unknown.Methods: A multicentre, cluster-randomised, crossover, feasibility and pilot trial was conducted across five hospitals in California. Phenylephrine (PE) vs norepinephrine (NE) infusion as the first-line vasopressor in patients under general anaesthesia alternated monthly at each hospital for 6 months. The primary endpoint was first-line vasopressor administration compliance of 80% or higher. Secondary endpoints were acute kidney injury (AKI), 30-day mortality, myocardial injury after noncardiac surgery (MINS), hospital length of stay, and rehospitalisation within 30 days.Results: A total of 3626 patients were enrolled over 6 months; 1809 patients were randomised in the NE group, 1817 in the PE group. Overall, 88.2% received the assigned first-line vasopressor. No drug infiltrations requiring treatment were reported in either group. Patients were median 63 yr old, 50% female, and 58% white. Randomisation in the NE group vs PE group did not reduce readmission within 30 days (adjusted odds ratio1/40.92; 95% confidence interval, 0.6-1.39), 30-day mortality (1.01; 0.48-2.09), AKI (1.1; 0.92-1.31), or MINS (1.63; 0.84-3.16).Conclusions: A large and diverse population undergoing major surgery under general anaesthesia was successfully enrolled and randomised to receive NE or PE infusion. This pilot and feasibility trial was not powered for adverse postoperative outcomes and a follow-up multicentre effectiveness trial is planned.Clinical trial registration: NCT04789330 (ClinicalTrials.gov).

Background. Intraoperative hypotension (IOH) is common and associated with mortality in major surgery. Although patients undergoing liver transplantation (LT) have low baseline blood pressure, the relation between blood pressure and mortality in LT is not well studied. We aimed to determine mean arterial pressure (MAP) that was associated with 30-d mortality in LT. Methods. We performed a retrospective cohort study. The data included patient demographics, pertinent preoperative and intraoperative variables, and MAP using various metrics and thresholds. The endpoint was 30-d mortality after LT. Results. One thousand one hundred seventy-eight patients from 2013 to 2020 were included. A majority of patients were exposed to IOH and many for a long period. Eighty-nine patients (7.6%) died within 30 d after LT. The unadjusted analysis showed that predicted mortality was associated with MAP <45 to 60 mm Hg but not MAP <65 mm Hg. The association between MAP and mortality was further tested using adjustment and various duration cutoffs. After adjustment, the shortest durations for MAPs <45, 50, and 55 mm Hg associated with 30-d mortality were 6, 10, and 25min (odds ratio, 1.911, 1.812, and 1.772; 95% confidence interval, 1.100-3.320, 1.039-3.158, and 1.008-3.114; P=0.002, 0.036, and 0.047), respectively. Exposure to MAP <60 mm Hg up to 120 min was not associated with increased mortality. Conclusion. In this large retrospective study, we found IOH was common during LT. Intraoperative MAP <55 mm Hg was associated with increased 30-d mortality after LT, and the duration associated with postoperative mortality was shorter with lower MAP than with higher MAP.

BACKGROUND:Many hospitals have replaced their legacy anesthesia information management system with an enterprise-wide electronic health record system. Integrating the anesthesia data within the context of the global hospital information infrastructure has created substantive challenges for many organizations. A process to build a perioperative data warehouse from Epic was recently published from the University of California Los Angeles (UCLA), but the generalizability of that process is unknown. We describe the implementation of their process at the University of Miami (UM). METHODS:The UCLA process was tested at UM, and performance was evaluated following the configuration of a reporting server and transfer of the required Clarity tables to that server. Modifications required for the code to execute correctly in the UM environment were identified and implemented, including the addition of locally specified elements in the database. RESULTS:The UCLA code to build the base tables in the perioperative data warehouse executed correctly after minor modifications to match the local server and database architecture at UM. The 26 stored procedures in the UCLA process all ran correctly using the default settings provided and populated the base tables. After modification of the item lists to reflect the UM implementation of Epic (eg, medications, laboratory tests, physiologic monitors, and anesthesia machine parameters), the UCLA code ran correctly and populated the base tables. The data from those tables were used successfully to populate the existing perioperative data warehouse at UM, which housed data from the legacy anesthesia information management system of the institution. The time to pull data from Epic and populate the perioperative data warehouse was 197 ± 47 minutes (standard deviation [SD]) on weekdays and 260 ± 56 minutes (SD) on weekend days, measured over 100 consecutive days. The longer times on weekends reflect the simultaneous execution of database maintenance tasks on the reporting server. The UCLA extract process has been in production at UM for the past 18 months and has been invaluable for quality assurance, business process, and research activities. CONCLUSIONS:The data schema developed at UCLA proved to be a practical and scalable method to extract information from the Epic electronic health system database into the perioperative data warehouse in use at UM. Implementing the process developed at UCLA to build a comprehensive perioperative data warehouse from Epic is an extensible process that other hospitals seeking more efficient access to their electronic health record data should consider.

The interest in applying machine learning in healthcare has grown rapidly in recent years. Most predictive algorithms requiring pathway implementations are evaluated using metrics focused on predictive performance, such as the c statistic. However, these metrics are of limited clinical value, for two reasons: (1) they do not account for the algorithm's role within a provider workflow; and (2) they do not quantify the algorithm's value in terms of patient outcomes and cost savings. We propose a model for simulating the selection of patients over time by a clinician using a machine learning algorithm, and quantifying the expected patient outcomes and cost savings. Using data on unplanned emergency department surgical readmissions, we show that factors such as the provider's schedule and postoperative prediction timing can have major effects on the pathway cohort size and potential cost reductions from preventing hospital readmissions.

During the perioperative period patients often suffer complications, including acute kidney injury (AKI), reintubation, and mortality. In order to effectively prevent these complications, high-risk patients must be readily identified. However, most current risk scores are designed to predict a single postoperative complication and often lack specificity on the patient level. In other fields, machine learning (ML) has been shown to successfully create models to predict multiple end points using a single input feature set. We hypothesized that ML can be used to create models to predict postoperative mortality, AKI, reintubation, and a combined outcome using a single set of features available at the end of surgery. A set of 46 features available at the end of surgery, including drug dosing, blood loss, vital signs, and others were extracted. Additionally, six additional features accounting for total intraoperative hypotension were extracted and trialed for different models. A total of 59,981 surgical procedures met inclusion criteria and the deep neural networks (DNN) were trained on 80% of the data, with 20% reserved for testing. The network performances were then compared to ASA Physical Status. In addition to creating separate models for each outcome, a multitask learning model was trialed that used information on all outcomes to predict the likelihood of each outcome individually. The overall rate of the examined complications in this data set was 0.79% for mortality, 22.3% (of 21,676 patients with creatinine values) for AKI, and 1.1% for reintubation. Overall, there was significant overlap between the various model types for each outcome, with no one modeling technique consistently performing the best. However, the best DNN models did beat the ASA score for all outcomes other than mortality. The highest area under the receiver operating characteristic curve (AUC) models were 0.792 (0.775-0.808) for AKI, 0.879 (0.851-0.905) for reintubation, 0.907 (0.872-0.938) for mortality, and 0.874 (0.864-0.866) for any outcome. The ASA score alone achieved AUCs of 0.652 (0.636-0.669) for AKI, 0.787 (0.757-0.818) for reintubation, 0.839 (0.804-0.875) for mortality, and 0.76 (0.748-0.773) for any outcome. Overall, the DNN architecture was able to create models that outperformed the ASA physical status to predict all outcomes based on a single feature set, consisting of objective data available at the end of surgery. No one model architecture consistently performed the best.

Background: Although prediction of hospital readmissions has been studied in medical patients, it has received relatively little attention in surgical patient populations. Published predictors require information only available at the moment of discharge. The authors hypothesized that machine learning approaches can be leveraged to accurately predict readmissions in postoperative patients from the emergency department. Further, the authors hypothesize that these approaches can accurately predict the risk of readmission much sooner than hospital discharge. Methods: Using a cohort of surgical patients at a tertiary care academic medical center, surgical, demographic, lab, medication, care team, and current procedural terminology data were extracted from the electronic health record. The primary outcome was whether there existed a future hospital readmission originating from the emergency department within 30 days of surgery. Secondarily, the time interval from surgery to the prediction was analyzed at 0, 12, 24, 36, 48, and 60 h. Different machine learning models for predicting the primary outcome were evaluated with respect to the area under the receiver-operator characteristic curve metric using different permutations of the available features. Results: Surgical hospital admissions (N = 34,532) from April 2013 to December 2016 were included in the analysis. Surgical and demographic features led to moderate discrimination for prediction after discharge (area under the curve: 0.74 to 0.76), whereas medication, consulting team, and current procedural terminology features did not improve the discrimination. Lab features improved discrimination, with gradient-boosted trees attaining the best performance (area under the curve: 0.866, SD 0.006). This performance was sustained during temporal validation with 2017 to 2018 data (area under the curve: 0.85 to 0.88). Lastly, the discrimination of the predictions calculated 36 h after surgery (area under the curve: 0.88 to 0.89) nearly matched those from time of discharge. Conclusions: A machine learning approach to predicting postoperative readmission can produce hospital-specific models for accurately predicting 30-day readmissions via the emergency department. Moreover, these predictions can be confidently calculated at 36 h after surgery without consideration of discharge-level data.