Surgical scheduling via optimization and machine learning with long-tailed data : Health care management science, in press

Using data from cardiovascular surgery patients with long and highly variable post-surgical lengths of stay (LOS), we develop a modeling framework to reduce recovery unit congestion. We estimate the LOS and its probability distribution using machine learning models, schedule procedures on a rolling basis using a variety of optimization models, and estimate performance with simulation. The machine learning models achieved only modest LOS prediction accuracy, despite access to a very rich set of patient characteristics. Compared to the current paper-based system used in the hospital, most optimization models failed to reduce congestion without increasing wait times for surgery. A conservative stochastic optimization with sufficient sampling to capture the long tail of the LOS distribution outperformed the current manual process and other stochastic and robust optimization approaches. These results highlight the perils of using oversimplified distributional models of LOS for scheduling procedures and the importance of using optimization methods well-suited to dealing with long-tailed behavior.

Optimization of Tree-Based Machine Learning Models to Predict the Length of Hospital Stay Using Genetic Algorithm

The length of hospital stay (LOS) is a significant indicator of the quality of patient care, hospital efficiency, and operational resilience. Considering the importance of LOS in hospital resource management, this research aims to improve the accuracy of LOS prediction using hyperparameter optimization (HPO). Expert physicians and related studies were reviewed to determine the variables affecting LOS. The electronic medical records of 200 patients in the department of internal medicine of a hospital in Iran were collected randomly. As the performance of machine learning (ML) models can vary based on the characteristics of the features, several models were applied and evaluated in this study. In particular, k-nearest neighbors (KNN), multivariate regression, decision tree (DT), random forest (RF), artificial neural network (ANN), and XGBoost have been evaluated and improved. The genetic algorithm (GA) was applied to optimize the tree-based models. In addition, the dummy coding technique, sometimes called the One-Hot encoding, was used to encode categorical features to increase prediction accuracy. Compared with other algorithms, the XGBoost model optimized by GA (XGB_GA) achieved higher accuracy and better prediction performance. The mean and median of absolute errors in the test dataset for this model were 1.54 and 1.14 days, respectively. In other words, the XGB_GA model reduced the mean absolute error by 37%, which is beneficial in the reliable design of a clinical decision support system.

Team Communication and Expectations Following Pediatric Cardiac Surgery: A Multi-Disciplinary Survey

Patients and families desire an accurate understanding of the expected recovery following congenital cardiac surgery. Variation in knowledge and expectations within the care team may be under-recognized and impact communication and care delivery. Our objective was to assess knowledge of common postoperative milestones and perceived efficacy of communication with patients and families and within the care team. An 18-question survey measuring knowledge of expected milestones for recovery after four index operations and team communication in the postoperative period was distributed electronically to multidisciplinary care team members at 16 academic pediatric heart centers. Answers were compared to local median data for each respondent's heart center to assess accuracy and stratified by heart center role and years of experience. We obtained 874 responses with broad representation of disciplines. More than half of all respondent predictions (55.3%) did not match their local median data. Percent matching did not vary by care team role but improved with increasing experience (35.8% < 2 years vs. 46.4% > 10 years, p = 0.2133). Of all respondents, 62.7% expressed confidence discussing the anticipated postoperative course, 78.6% denoted confidence discussing postoperative complications, and 55.3% conveyed that not all members of their care team share a common expectation for typical postoperative recovery. Most respondents (94.6%) stated that increased knowledge of local data would positively impact communication. Confidence in communication exceeded accuracy in predicting the timing of postoperative milestones. Important variation in knowledge and expectations for postoperative recovery in pediatric cardiac surgery exists and may impact communication and clinical effectiveness.

C-reactive protein as a non-linear predictor of prolonged length of intensive care unit stay after gastrointestinal cancer surgery

BACKGROUND

The relationship between C-reactive protein (CRP) levels and prolonged intensive care unit (ICU) length of stay (LoS) has not been well defined.

AIM

To explore the association between CRP levels at ICU admission and prolonged ICU LoS in gastrointestinal cancer (GC) patients after major surgery.

METHODS

A retrospective study was performed to quantify serum CRP levels and to establish their association with prolonged ICU LoS (≥ 72 h) in GC patients admitted to the ICU. Univariate and multivariate regression analyses were conducted, and restricted cubic spline curves with four knots (5%, 35%, 65%, 95%) were used to explore non-linearity assumptions.

RESULTS

A total of 408 patients were enrolled. Among them, 83 (20.3%) patients had an ICU LoS longer than 72 h. CRP levels were independently associated with the risk of prolonged ICU LoS [odds ratio (OR) 1.47, 95% confidence interval (CI) 1.00–2.17]. Restricted cubic spline analysis revealed a non-linear relationship between CRP levels and OR for the prolonged ICU LoS (P = 0.035 for non-linearity). After the cut-off of 2.6 (log transformed mg/L), the OR for prolonged ICU LoS significantly increased with CRP levels. The adjusted regression coefficient was 0.70 (95%CI 0.31–1.57, P = 0.384) for CRP levels less than 2.6, whereas it was 2.43 (95%CI 1.39–4.24, P = 0.002) for CRP levels higher than 2.6.

CONCLUSION

Among the GC patients, CRP levels at ICU admission were non-linearly associated with prolonged ICU LoS in survivors. An admission CRP level > 2.6 (log transformed mg/L) was associated with increased risk of prolonged ICU LoS.

Length of Stay in Patients Undergoing Tracheoplasty: A NSQIP Study

BACKGROUND

Prolonged length of stay (LOS) has been associated with increased morbidity and resource utilization in various surgical procedures. We aim to determine factors associated with increased hospital stay in patient undergoing tracheoplasty.

METHODS

The 2012-2018 National Surgical Quality Improvement Program (NSQIP) database was queried for patients undergoing tracheoplasty. Patient LOS was the primary clinical outcome. A LOS >75th percentile was considered as prolonged and was utilized for bivariate analysis of demographic, comorbidity, and operative characteristics. LOS was utilized as a continuous variable for multivariate linear regression analysis.

RESULTS

A total of 252 patients were queried. The majority of patients were female (67.5%), white (82.4%), and over the age of 65 (77.0%). Patients had a median LOS of 7 days with the 75th percentile cutoff being defined at 10 days. On bivariate analysis of associated comorbidities, patients with prolonged LOS were more commonly obese (72.4% vs. 53.1%, p = 0.009), diabetic (37.9% vs. 16.5%, p < 0.001), dyspneic (58.6% vs. 40.7%, p = 0.016), and had chronic steroid use (25.9% vs. 12.9%, p = 0.018). Multivariable logistic regression analysis demonstrated significant associations between prolonged LOS and both chronic obstructive pulmonary disorder (COPD) (OR: 3.43, p = 0.020) and chronic steroid use (OR: 3.81, p = 0.018).

CONCLUSIONS

This study elucidates factors associated with prolonged LOS in patients undergoing tracheoplasty. Patients with COPD and chronic steroid use were significantly associated with prolonged LOS.

LEVEL OF EVIDENCE

4 Laryngoscope, 2022.

Comparison of Four Machine Learning Techniques for Prediction of Intensive Care Unit Length of Stay in Heart Transplantation Patients

Background

Post-operative heart transplantation patients often require admission to an intensive care unit (ICU). Early prediction of the ICU length of stay (ICU-LOS) of these patients is of great significance and can guide treatment while reducing the mortality rate among patients. However, conventional linear models have tended to perform worse than non-linear models.

Materials and Methods

We collected the clinical data of 365 patients from Wuhan Union Hospital who underwent heart transplantation surgery between April 2017 and August 2020. The patients were randomly divided into training data (N = 256) and test data (N = 109) groups. 84 clinical features were collected for each patient. Features were validated using the Least Absolute Shrinkage and Selection Operator (LASSO) regression’s fivefold cross-validation method. We obtained Shapley Additive explanations (SHAP) values by executing package “shap” to interpret model predictions. Four machine learning models and logistic regression algorithms were developed. The area under the receiver operating characteristic curve (AUC-ROC) was used to compare the prediction performance of different models. Finally, for the convenience of clinicians, an online web-server was established and can be freely accessed via the website https://wuhanunion.shinyapps.io/PredictICUStay/.

Results

In this study, 365 consecutive patients undergoing heart transplantation surgery for moderate (NYHA grade 3) or severe (NYHA grade 4) heart failure were collected in Wuhan Union Hospital from 2017 to 2020. The median age of the recipient patients was 47.2 years, while the median age of the donors was 35.58 years. 330 (90.4%) of the donor patients were men, and the average surgery duration was 260.06 min. Among this cohort, 47 (12.9%) had renal complications, 25 (6.8%) had hepatic complications, 11 (3%) had undergone chest re-exploration and 19 (5.2%) had undergone extracorporeal membrane oxygenation (ECMO). The following six important clinical features were selected using LASSO regression, and according to the result of SHAP, the rank of importance was (1) the use of extracorporeal membrane oxygenation (ECMO); (2) donor age; (3) the use of an intra-aortic balloon pump (IABP); (4) length of surgery; (5) high creatinine (Cr); and (6) the use of continuous renal replacement therapy (CRRT). The eXtreme Gradient Boosting (XGBoost) algorithm presented significantly better predictive performance (AUC-ROC = 0.88) than other models [Accuracy: 0.87; sensitivity: 0.98; specificity: 0.51; positive predictive value (PPV): 0.86; negative predictive value (NPV): 0.93].

Conclusion

Using the XGBoost classifier with heart transplantation patients can provide an accurate prediction of ICU-LOS, which will not only improve the accuracy of clinical decision-making but also contribute to the allocation and management of medical resources; it is also a real-world example of precision medicine in hospitals.

Ignorance Isn't Bliss: We Must Close the Machine Learning Knowledge Gap in Pediatric Critical Care

Pediatric intensivists are bombarded with more patient data than ever before. Integration and interpretation of data from patient monitors and the electronic health record (EHR) can be cognitively expensive in a manner that results in delayed or suboptimal medical decision making and patient harm. Machine learning (ML) can be used to facilitate insights from healthcare data and has been successfully applied to pediatric critical care data with that intent. However, many pediatric critical care medicine (PCCM) trainees and clinicians lack an understanding of foundational ML principles. This presents a major problem for the field. We outline the reasons why in this perspective and provide a roadmap for competency-based ML education for PCCM trainees and other stakeholders.

[Predicting prolonged length of intensive care unit stay via machine learning]

OBJECTIVE

To construct length of intensive care unit (ICU) stay (LOS-ICU) prediction models for ICU patients, based on three machine learning models support vector machine (SVM), classification and regression tree (CART), and random forest (RF), and to compare the prediction perfor-mance of the three machine learning models with the customized simplified acute physiology score Ⅱ(SAPS-Ⅱ) model.

METHODS

We used medical information mart for intensive care (MIMIC)-Ⅲ database for model development and validation. The primary outcome was prolonged LOS-ICU(pLOS-ICU), defined as longer than the third quartile of patients' LOS-ICU in the studied dataset. The recursive feature elimination method was used to do feature selection for three machine learning models. We utilized 5-fold cross validation to evaluate model prediction performance. The Brier value, area under the receiver operation characteristic curve (AUROC), and estimated calibration index (ECI) were used as perfor-mance measures. Performances of the four models were compared, and performance differences between the models were assessed using two-sided t test. The model with the best prediction performance was employed to generate variable importance ranking, and the identified top five important predictors were pre-sented.

RESULTS

The final cohort in our study consisted of 40 200 eligible ICU patients, of whom 23.7% were with pLOS-ICU. The proportion of the male patients was 57.6%, and the age of all the ICU patients was (61.9±16.5) years.Results showed that the three machine learning models outperformed the customized SAPS-Ⅱ model in terms of all the performance measures with statistical significance (P < 0.01). Among the three machine learning models, the RF model achieved the best overall performance (Brier value, 0.145), discrimination (AUROC, 0.770) and calibration (ECI, 7.259). The calibration curve showed that the RF model slightly overestimated the risk of pLOS-ICU in high-risk ICU patients, but underestimated the risk of pLOS-ICU in low-risk ICU patients. Top five important predictors for pLOS-ICU identified by the RF model included age, heart rate, systolic blood pressure, body tempe-rature, and ratio of arterial oxygen tension to the fraction of inspired oxygen(PaO₂/FiO₂).

CONCLUSION

The RF algorithm-based pLOS-ICU prediction model had a best prediction performance in this study. It lays a foundation for future application of the RF-based pLOS-ICU prediction model in ICU clinical practice.

Predicting Prolonged Length of ICU Stay through Machine Learning

This study aimed to construct machine learning (ML) models for predicting prolonged length of stay (pLOS) in intensive care units (ICU) among general ICU patients. A multicenter database called eICU (Collaborative Research Database) was used for model derivation and internal validation, and the Medical Information Mart for Intensive Care (MIMIC) III database was used for external validation. We used four different ML methods (random forest, support vector machine, deep learning, and gradient boosting decision tree (GBDT)) to develop prediction models. The prediction performance of the four models were compared with the customized simplified acute physiology score (SAPS) II. The area under the receiver operation characteristic curve (AUROC), area under the precision-recall curve (AUPRC), estimated calibration index (ECI), and Brier score were used to measure performance. In internal validation, the GBDT model achieved the best overall performance (Brier score, 0.164), discrimination (AUROC, 0.742; AUPRC, 0.537), and calibration (ECI, 8.224). In external validation, the GBDT model also achieved the best overall performance (Brier score, 0.166), discrimination (AUROC, 0.747; AUPRC, 0.536), and calibration (ECI, 8.294). External validation showed that the calibration curve of the GBDT model was an optimal fit, and four ML models outperformed the customized SAPS II model. The GBDT-based pLOS-ICU prediction model had the best prediction performance among the five models on both internal and external datasets. Furthermore, it has the potential to assist ICU physicians to identify patients with pLOS-ICU risk and provide appropriate clinical interventions to improve patient outcomes.

Admission criteria for a cardiovascular short stay unit: a retrospective analysis on a pilot unit

Rapid intensive observation (RIO) units have been created to guarantee high standards of care in a sustainable health-care system. Within short stay units (SSUs), which are a subgroup of RIOs, only rapidly manageable patients should be admitted. Physicians are unable to predict the length of stay (LOS) as objective criteria to make such a prediction are missing. A retrospective observational study was carried out to identify the objective criteria for admission within a cardiovascular care-oriented SSU. Over a period of 317 days, 340 patients (age 69.4 ± 14.7 years) were admitted to a pilot SSU within our internal medicine department. The most frequent diagnoses were chest pain (45.9%), syncope (12.9%), and supraventricular arrhythmias (11.2%). The median LOS was 4 days (quartile 1:3; quartile 3:7). Predictors of LOS ≤ 96 h were age < 80, hemoglobin > 115 g/L, estimated glomerular filtration rate > 45 mL/min/1.73 m², Charlson Comorbidity Index < 3, Barthel Index > 40, diagnosis of chest pain, syncope, supraventricular arrhythmias, or acute heart failure. The HEART (history, ECG, age, risk factors, troponin) score was found to be excellent in risk stratification of patients admitted for chest pain. Blood tests and anamnestic variables can be used to predict the LOS and thus SSU admission. The HEART score may help in the classification of patients with chest pain admitted to an SSU.

Hospital Length of Stay Prediction Methods: A Systematic Review

OBJECTIVE

This systematic review sought to establish a picture of length of stay (LOS) prediction methods based on available hospital data and study protocols designed to measure their performance.

MATERIALS AND METHODS

An English literature search was done relative to hospital LOS prediction from 1972 to September 2019 according to the PRISMA guidelines. Articles were retrieved from PubMed, ScienceDirect, and arXiv databases. Information were extracted from the included papers according to a standardized assessment of population setting and study sample, data sources and input variables, LOS prediction methods, validation study design, and performance evaluation metrics.

RESULTS

Among 74 selected articles, 98.6% (73/74) used patients' data to predict LOS; 27.0% (20/74) used temporal data; and 21.6% (16/74) used the data about hospitals. Overall, regressions were the most popular prediction methods (64.9%, 48/74), followed by machine learning (20.3%, 15/74) and deep learning (17.6%, 13/74). Regarding validation design, 35.1% (26/74) did not use a test set, whereas 47.3% (35/74) used a separate test set, and 17.6% (13/74) used cross-validation. The most used performance metrics were R2 (47.3%, 35/74), mean squared (or absolute) error (24.4%, 18/74), and the accuracy (14.9%, 11/74). Over the last decade, machine learning and deep learning methods became more popular (P=0.016), and test sets and cross-validation got more and more used (P=0.014).

CONCLUSIONS

Methods to predict LOS are more and more elaborate and the assessment of their validity is increasingly rigorous. Reducing heterogeneity in how these methods are used and reported is key to transparency on their performance.

Using nursing notes to improve clinical outcome prediction in intensive care patients: A retrospective cohort study

OBJECTIVE

Electronic health record documentation by intensive care unit (ICU) clinicians may predict patient outcomes. However, it is unclear whether physician and nursing notes differ in their ability to predict short-term ICU prognosis. We aimed to investigate and compare the ability of physician and nursing notes, written in the first 48 hours of admission, to predict ICU length of stay and mortality using 3 analytical methods.

MATERIALS AND METHODS

This was a retrospective cohort study with split sampling for model training and testing. We included patients ≥18 years of age admitted to the ICU at Beth Israel Deaconess Medical Center in Boston, Massachusetts, from 2008 to 2012. Physician or nursing notes generated within the first 48 hours of admission were used with standard machine learning methods to predict outcomes.

RESULTS

For the primary outcome of composite score of ICU length of stay ≥7 days or in-hospital mortality, the gradient boosting model had better performance than the logistic regression and random forest models. Nursing and physician notes achieved area under the curves (AUCs) of 0.826 and 0.796, respectively, with even better predictive power when combined (AUC, 0.839).

DISCUSSION

Models using only nursing notes more accurately predicted short-term prognosis than did models using only physician notes, but in combination, the models achieved the greatest accuracy in prediction.

CONCLUSIONS

Our findings demonstrate that statistical models derived from text analysis in the first 48 hours of ICU admission can predict patient outcomes. Physicians' and nurses' notes are both uniquely important in mortality prediction and combining these notes can produce a better predictive model.

PLASMA INTERLEUKIN-6 LEVELS MAY BE ASSOCIATED WITH THE LENGTH OF STAY TIME OF ADULT HYPERGLYCEMIC PATIENTS IN AN INTENSIVE CARE UNIT

Context

Estimation of intensive care unit (ICU) length of stay time (LOS) may be challenging, and pro-inflammatory cytokines can be used as a marker for this purpose.

Objective

The current study aimed to investigate the association between pro-inflammatory cytokine levels and LOS in hyperglycemic patients admitted to adult ICU.

Design

This is a prospective observational study.

Subjects and Methods

All adult ICU patients with a blood glucose level higher than 250 mg/dL, during the study period were included. Hospitalization day demographics were recorded, and plasma IL-6, IL1-ß, and TNF-α concentrations were measured.

Results

A total of 74 patients were enrolled in the study. Diabetic ketoacidosis (DKA) was positive in 31 patients, and the remaining 43 were in the non-DKA (NDKA) group. There was no difference between the two groups in terms of age, gender, LOS, hemoglobin, hematocrit, lactate levels, and platelets count. IL-6, IL-1ß, and TNF-α levels did not show any difference between DKA and NDKA groups (p=0.784, 0.413, and 0.288, respectively). There was a positive correlation between IL-6 levels and LOS (n=74, Pearson correlation=0.330; p=0.004).

Conclusions

Among pro-inflammatory cytokines, IL-6 showed a better performance for the prediction of LOS than IL-1ß, TNF-α, and CRP.

Risk factors and socio-economic burden in pancreatic ductal adenocarcinoma operation: a machine learning based analysis

BACKGROUND

Surgical resection is the major way to cure pancreatic ductal adenocarcinoma (PDAC). However, this operation is complex, and the peri-operative risk is high, making patients more likely to be admitted to the intensive care unit (ICU). Therefore, establishing a risk model that predicts admission to ICU is meaningful in preventing patients from post-operation deterioration and potentially reducing socio-economic burden.

METHODS

We retrospectively collected 120 clinical features from 1242 PDAC patients, including demographic data, pre-operative and intra-operative blood tests, in-hospital duration, and ICU status. Machine learning pipelines, including Supporting Vector Machine (SVM), Logistic Regression, and Lasso Regression, were employed to choose an optimal model in predicting ICU admission. Ordinary least-squares regression (OLS) and Lasso Regression were adopted in the correlation analysis of post-operative bleeding, total in-hospital duration, and discharge costs.

RESULTS

SVM model achieved higher performance than the other two models, resulted in an AU-ROC of 0.80. The features, such as age, duration of operation, monocyte count, and intra-operative partial arterial pressure of oxygen (PaO2), are risk factors in the ICU admission. The protective factors include RBC count, analgesic pump dexmedetomidine (DEX), and intra-operative maintenance of DEX. Basophil percentage, duration of the operation, and total infusion volume were risk variables for staying in ICU. The bilirubin, CA125, and pre-operative albumin were associated with the post-operative bleeding volume. The operation duration was the most important factor for discharge costs, while pre-lymphocyte percentage and the absolute count are responsible for less cost.

CONCLUSIONS

We observed that several new indicators such as DEX, monocyte count, basophil percentage, and intra-operative PaO2 showed a good predictive effect on the possibility of admission to ICU and duration of stay in ICU. This work provided an essential reference for indication in advance to PDAC operation.

Artificial Intelligence in the Intensive Care Unit

This article is one of ten reviews selected from the Annual Update in Intensive Care and Emergency Medicine 2020. Other selected articles can be found online at https://www.biomedcentral.com/collections/annualupdate2020. Further information about the Annual Update in Intensive Care and Emergency Medicine is available from http://www.springer.com/series/8901.

Can We Improve Prediction of Adverse Surgical Outcomes? Development of a Surgical Complexity Score Using a Novel Machine Learning Technique

BACKGROUND

An optimal method to quantify surgical complexity using patient comorbidities derived from administrative billing data is lacking. We sought to develop a novel, easy-to-use surgical Complexity Score to accurately predict adverse outcomes among patients undergoing elective surgery.

STUDY DESIGN

A novel surgical Complexity Score was developed using 100% Medicare Inpatient and Outpatient Standard Analytic Files (SAFs) from years 2012 to 2016 (n = 1,049,160). Comorbid conditions were entered into a machine learning algorithm to assign weights to maximize the correlation with multiple postoperative outcomes including morbidity, readmission, mortality, and postoperative super-use. Predictive ability was compared against 3 of the most commonly used risk adjustment indices: the Charlson Comorbidity Index (CCI), Elixhauser Comorbidity Index (ECI), and the Centers for Medicare and Medicaid Service's Hierarchical Condition Category (CMS-HCC).

RESULTS

Patients underwent colectomy (12.6%), abdominal aortic aneurysm repair (4.4%), coronary artery bypass grafting (13.0%), total hip replacement (22.0%), total knee replacement (43.0%), or lung resection (5.0%). The Complexity Score had a good to very good predictive ability for all adverse outcomes. The Complexity Score had the highest accuracy in predicting perioperative morbidity (area under the curve [AUC]: 0.868, 95% CI 0.866 to 0.869); this performed better than the CCI (AUC: 0.717, 95% CI 0.715 to 0.719), ECI (AUC: 0.799, 95% CI 0.797 to 0.800), and similar to the CMS-HCC (AUC: 0.862, 95% CI 0.861 to 0.863). Similarly, the Complexity Score outperformed each of the 3 other comorbidity indices in predicting 90-day readmission (AUC: 0.707, 95% CI 0.705 to 0.709), 30-day readmission (AUC: 0.717, 95% CI 0.715 to 0.720), and postoperative super-use (AUC: 0.817, 95% CI 0.814 to 0.820).

CONCLUSIONS

Compared with the most commonly used comorbidity and surgical risk scores, the novel surgical Complexity Score outperformed the CCI, ECI, and CMS-HCC in predicting postoperative morbidity, 30-day readmission, 90-day readmission, and postoperative super-use.

The Critical Care Society of Southern Africa Consensus Guideline on ICU Triage and Rationing (ConICTri)

Background

In South Africa (SA), administrators and intensive care practitioners are faced with the challenge of resource scarcity as well as an increasing demand for intensive care unit (ICU) services. ICU services are expensive, and practitioners in low- to middle-income countries experience the consequences of limited resources daily. Critically limited resources necessitate that rationing and triage (prioritisation) decisions are routinely necessary in SA, particularly in the publicly funded health sector.

Purpose

The purpose of this guideline is to utilise the relevant recommendations of the associated consensus meeting document and other internationally accepted principles to develop a guideline to inform frontline triage policy and ensure the best utilisation of adult intensive care in SA, while maintaining the fair distribution of available resources.

Recommendations

An overall conceptual framework for the triage process was developed. The components of the framework were developed on the basis that patients should be admitted preferentially when the likely incremental medical benefit derived from ICU admission justifies admission. An estimate of likely resource use should also form part of the triage decision, with those patients requiring relatively less resources to achieve substantial benefit receiving priority for admission. Thus, the triage system should maximise the benefits obtained from ICU resources available for the community. Where possible, practical examples of what the consensus group agreed would be considered appropriate practice under specified South African circumstances were provided, to assist clinicians with practical decision-making. It must be stressed that this guideline is not intended to be prescriptive for individual hospital or regional practice, and hospitals and regions are encouraged to develop specified local guidelines with locally relevant examples. The guideline should be reviewed and revised if appropriate within 5 years.

Conclusion

In recognition of the absolute need to limit patient access to ICU because of the lack of sufficient intensive care resources in public hospitals, this guideline has been developed to guide policy-making and assist frontline triage decision-making in SA. This document is not a complete plan for quality practice, but rather a template to support frontline clinicians, guide administrators and inform the public regarding appropriate triage decision-making.

Inclusion of Unstructured Clinical Text Improves Early Prediction of Death or Prolonged ICU Stay

OBJECTIVES

Early prediction of undesired outcomes among newly hospitalized patients could improve patient triage and prompt conversations about patients' goals of care. We evaluated the performance of logistic regression, gradient boosting machine, random forest, and elastic net regression models, with and without unstructured clinical text data, to predict a binary composite outcome of in-hospital death or ICU length of stay greater than or equal to 7 days using data from the first 48 hours of hospitalization.

DESIGN

Retrospective cohort study with split sampling for model training and testing.

SETTING

A single urban academic hospital.

PATIENTS

All hospitalized patients who required ICU care at the Beth Israel Deaconess Medical Center in Boston, MA, from 2001 to 2012.

INTERVENTIONS

None.

MEASUREMENTS AND MAIN RESULTS

Among eligible 25,947 hospital admissions, we observed 5,504 (21.2%) in which patients died or had ICU length of stay greater than or equal to 7 days. The gradient boosting machine model had the highest discrimination without (area under the receiver operating characteristic curve, 0.83; 95% CI, 0.81-0.84) and with (area under the receiver operating characteristic curve, 0.89; 95% CI, 0.88-0.90) text-derived variables. Both gradient boosting machines and random forests outperformed logistic regression without text data (p < 0.001), whereas all models outperformed logistic regression with text data (p < 0.02). The inclusion of text data increased the discrimination of all four model types (p < 0.001). Among those models using text data, the increasing presence of terms "intubated" and "poor prognosis" were positively associated with mortality and ICU length of stay, whereas the term "extubated" was inversely associated with them.

CONCLUSIONS

Variables extracted from unstructured clinical text from the first 48 hours of hospital admission using natural language processing techniques significantly improved the abilities of logistic regression and other machine learning models to predict which patients died or had long ICU stays. Learning health systems may adapt such models using open-source approaches to capture local variation in care patterns.

Physician predictions of length of stay of patients admitted with heart failure

Physicians' ability to predict length of stay is understudied, particularly for patients with heart failure (HF) admissions. The objective of this prospective, observational cohort study was to measure the accuracy of inpatient physicians' predictions of length of stay at the time of admission of patients admitted to an academic tertiary care hospital with HF and to determine whether level of experience improves accuracy. The patients included 165 adults consecutively admitted with heart failure, about whom 415 predictions were made within 24 hours of admission. Mean and median lengths of stay were 10.9 and 8 days, respectively. The mean difference between predicted and actual length of stay was statistically significant for all groups: interns, -5.9 days (95% confidence interval [CI]: -8.2 to -3.6, P < 0.0001); residents, -4.3 days (95% CI: -6.0 to -2.7, P = 0.0001); attending cardiologists, -3.5 days (95% CI: -5.1 to -2.0, P < 0.0001). There were no differences in accuracy by level of experience (P = 0.61). Physicians, regardless of experience, underestimate length of stay of patients admitted with HF. Journal of Hospital Medicine 2016;11:642-645. © 2016 Society of Hospital Medicine.