Predicting Prolonged Length of ICU Stay through Machine Learning

Urea to creatinine ratio as a predictor of persistent critical illness

INTRODUCTION

Persistent critical illness (PCI) is a syndrome in which the acute presenting problem has been stabilized, but the patient's clinical state does not allow ICU discharge. The burden associated with PCI is substantial. The most obvious marker of PCI is prolonged ICU length of stay (LOS), usually greater than 10 days. Urea to Creatinine ratio (UCr) has been suggested as an early marker of PCI development.

METHODS

A single-center retrospective study. Data of patients admitted to a general mixed medical-surgical ICU during Jan 1st 2018 till Dec 31st 2022 was extracted, including demographic data, baseline characteristics, daily urea and creatinine results, renal replacement therapy (RRT) provided, and outcome measures - length of stay, and mortality (ICU, and 90 days). Patients were defined as PCI patients if their LOS was >10 days. We used Fisher exact test or Chi-square to compare PCI and non-PCI patients. The association between UCr with PCI development was assessed by repeated measures linear model. Multivariate Cox regression was used for 1 year mortality assessment.

RESULTS

2098 patients were included in the analysis. Patients who suffered from PCI were older, with higher admission prognostic scores. Their 90-day mortality was significantly higher than non-PCI patients (34.58% vs 12.18%, p < 0.0001). A significant difference in UCr was found only on the first admission day among all patients. This was not found when examining separately surgical, trauma, or transplantation patients. We did not find a difference in UCr in different KDIGO (Kidney Disease Improving Global Outcomes) stages. Elevated UCr and PCI were found to be significantly associated with 1 year mortality.

CONCLUSION

In this single center retrospective cohort study, UCr was not found to be associated with PCI development.

Supervised machine learning algorithms to predict the duration and risk of long-term hospitalization in HIV-infected individuals: a retrospective study

Objective

The study aimed to use supervised machine learning models to predict the length and risk of prolonged hospitalization in PLWHs to help physicians timely clinical intervention and avoid waste of health resources.

Methods

Regression models were established based on RF, KNN, SVM, and XGB to predict the length of hospital stay using RMSE, MAE, MAPE, and R2, while classification models were established based on RF, KNN, SVM, NN, and XGB to predict risk of prolonged hospital stay using accuracy, PPV, NPV, specificity, sensitivity, and kappa, and visualization evaluation based on AUROC, AUPRC, calibration curves and decision curves of all models were used for internally validation.

Results

In regression models, XGB model performed best in the internal validation (RMSE = 16.81, MAE = 10.39, MAPE = 0.98, R2 = 0.47) to predict the length of hospital stay, while in classification models, NN model presented good fitting and stable features and performed best in testing sets, with excellent accuracy (0.7623), PPV (0.7853), NPV (0.7092), sensitivity (0.8754), specificity (0.5882), and kappa (0.4672), and further visualization evaluation indicated that the largest AUROC (0.9779), AUPRC (0.773) and well-performed calibration curve and decision curve in the internal validation.

Conclusion

This study showed that XGB model was effective in predicting the length of hospital stay, while NN model was effective in predicting the risk of prolonged hospitalization in PLWH. Based on predictive models, an intelligent medical prediction system may be developed to effectively predict the length of stay and risk of HIV patients according to their medical records, which helped reduce the waste of healthcare resources.

Predicting Duration of Invasive Mechanical Ventilation in the Pediatric ICU

BACKGROUND

Timely ventilator liberation can prevent morbidities associated with invasive mechanical ventilation in the pediatric ICU (PICU). There currently exists no standard benchmark for duration of invasive mechanical ventilation in the PICU. This study sought to develop and validate a multi-center prediction model of invasive mechanical ventilation duration to determine a standardized duration of invasive mechanical ventilation ratio.

METHODS

This was a retrospective cohort study using registry data from 157 institutions in the Virtual Pediatric Systems database. The study population included encounters in the PICU between 2012-2021 involving endotracheal intubation and invasive mechanical ventilation in the first day of PICU admission who received invasive mechanical ventilation for > 24 h. Subjects were stratified into a training cohort (2012-2017) and 2 validation cohorts (2018-2019/2020-2021). Four models to predict the duration of invasive mechanical ventilation were trained using data from the first 24 h, validated, and compared.

RESULTS

The study included 112,353 unique encounters. All models had observed-to-expected (O/E) ratios close to one but low mean squared error and R2 values. The random forest model was the best performing model and achieved an O/E ratio of 1.043 (95% CI 1.030-1.056) and 1.004 (95% CI 0.990-1.019) in the validation cohorts and 1.009 (95% CI 1.004-1.016) in the full cohort. There was a high degree of institutional variation, with single-unit O/E ratios ranging between 0.49-1.91. When stratified by time period, there were observable changes in O/E ratios at the individual PICU level over time.

CONCLUSIONS

We derived and validated a model to predict the duration of invasive mechanical ventilation that performed well in aggregated predictions at the PICU and the cohort level. This model could be beneficial in quality improvement and institutional benchmarking initiatives for use at the PICU level and for tracking of performance over time.

Machine Learning for Benchmarking Critical Care Outcomes

OBJECTIVES

Enhancing critical care efficacy involves evaluating and improving system functioning. Benchmarking, a retrospective comparison of results against standards, aids risk-adjusted assessment and helps healthcare providers identify areas for improvement based on observed and predicted outcomes. The last two decades have seen the development of several models using machine learning (ML) for clinical outcome prediction. ML is a field of artificial intelligence focused on creating algorithms that enable computers to learn from and make predictions or decisions based on data. This narrative review centers on key discoveries and outcomes to aid clinicians and researchers in selecting the optimal methodology for critical care benchmarking using ML.

METHODS

We used PubMed to search the literature from 2003 to 2023 regarding predictive models utilizing ML for mortality (592 articles), length of stay (143 articles), or mechanical ventilation (195 articles). We supplemented the PubMed search with Google Scholar, making sure relevant articles were included. Given the narrative style, papers in the cohort were manually curated for a comprehensive reader perspective.

RESULTS

Our report presents comparative results for benchmarked outcomes and emphasizes advancements in feature types, preprocessing, model selection, and validation. It showcases instances where ML effectively tackled critical care outcome-prediction challenges, including nonlinear relationships, class imbalances, missing data, and documentation variability, leading to enhanced results.

CONCLUSIONS

Although ML has provided novel tools to improve the benchmarking of critical care outcomes, areas that require further research include class imbalance, fairness, improved calibration, generalizability, and long-term validation of published models.

Predictive models for short-term mortality and length of hospital stay among adults with community-onset bacteraemia before and during the COVID-19 pandemic: application of early data dynamics

BACKGROUND

The development of scoring systems to predict the short-term mortality and the length of hospital stay (LOS) in patients with bacteraemia is essential to improve the quality of care and reduce the occupancy variance in the hospital bed.

METHODS

Adults hospitalised with community-onset bacteraemia in the coronavirus disease 2019 (COVID-19) and pre-COVID-19 eras were captured as the validation and derivation cohorts in the multicentre study, respectively. Model I incorporated all variables available on day 0, Model II incorporated all variables available on day 3, and Models III, IV, and V incorporated the variables that changed from day 0 to day 3. This study adopted the statistical and machine learning (ML) methods to jointly determine the prediction performance of these models in two study cohorts.

RESULTS

A total of 3,639 (81.4%) and 834 (18.6%) patients were included in the derivation and validation cohorts, respectively. Model IV achieved the best performance in predicting 30-day mortality in both cohorts. The most frequently identified variables incorporated into Model IV were deteriorated consciousness from day 0 to day 3 and deteriorated respiration from day 0 to day 3. Model V achieved the best performance in predicting LOS in both cohorts. The most frequently identified variables in Model V were deteriorated consciousness from day 0 to day 3, a body temperature ≤ 36.0 °C or ≥ 39.0 °C on day 3, and a diagnosis of complicated bacteraemia.

CONCLUSIONS

For hospitalised adults with community-onset bacteraemia, clinical variables that dynamically changed from day 0 to day 3 were crucial in predicting the short-term mortality and LOS.

Research Hotspots and Trends of Deep Learning in Critical Care Medicine: A Bibliometric and Visualized Study

Background

Interest in the application of deep learning (DL) in critical care medicine (CCM) is growing rapidly. However, comprehensive bibliometric research that analyze and measure the global literature is still lacking.

Objective

The present study aimed to systematically evaluate the research hotspots and trends of DL in CCM worldwide based on the output of publications, cooperative relationships of research, citations, and the co-occurrence of keywords.

Methods

A total of 1708 articles in all were obtained from Web of Science. Bibliometric analysis was performed by Bibliometrix package in R software (4.2.2), Microsoft Excel 2019, VOSviewer (1.6.18), and CiteSpace (5.8.R3).

Results

The annual publications increased steeply in the past five years, accounting for 95.67% (1634/1708) of all the included literature. China and USA contributed to approximately 71.66% (1244/1708) of all publications. Seven of the top ten most productive organizations rank in the top 100 universities globally. Hot spots in research on the application of DL in CCM have focused on classifying disease phenotypes, predicting early signs of clinical deterioration, and forecasting disease progression, prognosis, and death. Convolutional neural networks, long and short-term memory networks, recurrent neural networks, transformer models, and attention mechanisms were all commonly used DL technologies.

Conclusion

Hot spots in research on the application of DL in CCM have focused on classifying disease phenotypes, predicting early signs of clinical deterioration, and forecasting disease progression, prognosis, and death. Extensive collaborative research to improve the maturity and robustness of the model remains necessary to make DL-based model applications sufficiently compelling for conventional CCM practice.

Machine learning-based prediction of hospital prolonged length of stay admission at emergency department: a Gradient Boosting algorithm analysis

Objective

This study aims to develop and compare different models to predict the Length of Stay (LoS) and the Prolonged Length of Stay (PLoS) of inpatients admitted through the emergency department (ED) in general patient settings. This aim is not only to promote any specific model but rather to suggest a decision-supporting tool (i.e., a prediction framework).

Methods

We analyzed a dataset of patients admitted through the ED to the "Sant"Orsola Malpighi University Hospital of Bologna, Italy, between January 1 and October 26, 2022. PLoS was defined as any hospitalization with LoS longer than 6 days. We deployed six classification algorithms for predicting PLoS: Random Forest (RF), Support Vector Machines (SVM), Gradient Boosting (GB), AdaBoost, K-Nearest Neighbors (KNN), and logistic regression (LoR). We evaluated the performance of these models with the Brier score, the area under the ROC curve (AUC), accuracy, sensitivity (recall), specificity, precision, and F1-score. We further developed eight regression models for LoS prediction: Linear Regression (LR), including the penalized linear models Least Absolute Shrinkage and Selection Operator (LASSO), Ridge and Elastic-net regression, Support vector regression, RF regression, KNN, and eXtreme Gradient Boosting (XGBoost) regression. The model performances were measured by their mean square error, mean absolute error, and mean relative error. The dataset was randomly split into a training set (70%) and a validation set (30%).

Results

A total of 12,858 eligible patients were included in our study, of whom 60.88% had a PloS. The GB classifier best predicted PloS (accuracy 75%, AUC 75.4%, Brier score 0.181), followed by LoR classifier (accuracy 75%, AUC 75.2%, Brier score 0.182). These models also showed to be adequately calibrated. Ridge and XGBoost regressions best predicted LoS, with the smallest total prediction error. The overall prediction error is between 6 and 7 days, meaning there is a 6-7 day mean difference between actual and predicted LoS.

Conclusion

Our results demonstrate the potential of machine learning-based methods to predict LoS and provide valuable insights into the risks behind prolonged hospitalizations. In addition to physicians' clinical expertise, the results of these models can be utilized as input to make informed decisions, such as predicting hospitalizations and enhancing the overall performance of a public healthcare system.

A risk nomogram for predicting prolonged intensive care unit stays in patients with chronic obstructive pulmonary disease

Background

Providing intensive care is increasingly expensive, and the aim of this study was to construct a risk column line graph (nomograms)for prolonged length of stay (LOS) in the intensive care unit (ICU) for patients with chronic obstructive pulmonary disease (COPD).

Methods

This study included 4,940 patients, and the data set was randomly divided into training (n = 3,458) and validation (n = 1,482) sets at a 7:3 ratio. First, least absolute shrinkage and selection operator (LASSO) regression analysis was used to optimize variable selection by running a tenfold k-cyclic coordinate descent. Second, a prediction model was constructed using multifactorial logistic regression analysis. Third, the model was validated using receiver operating characteristic (ROC) curves, Hosmer-Lemeshow tests, calibration plots, and decision-curve analysis (DCA), and was further internally validated.

Results

This study selected 11 predictors: sepsis, renal replacement therapy, cerebrovascular disease, respiratory failure, ventilator associated pneumonia, norepinephrine, bronchodilators, invasive mechanical ventilation, electrolytes disorders, Glasgow Coma Scale score and body temperature. The models constructed using these 11 predictors indicated good predictive power, with the areas under the ROC curves being 0.826 (95%CI, 0.809-0.842) and 0.827 (95%CI, 0.802-0.853) in the training and validation sets, respectively. The Hosmer-Lemeshow test indicated a strong agreement between the predicted and observed probabilities in the training (χ² = 8.21, p = 0.413) and validation (χ² = 0.64, p = 0.999) sets. In addition, decision-curve analysis suggested that the model had good clinical validity.

Conclusion

This study has constructed and validated original and dynamic nomograms for prolonged ICU stay in patients with COPD using 11 easily collected parameters. These nomograms can provide useful guidance to medical and nursing practitioners in ICUs and help reduce the disease and economic burdens on patients.

Explainable time-series deep learning models for the prediction of mortality, prolonged length of stay and 30-day readmission in intensive care patients

Background

In-hospital mortality, prolonged length of stay (LOS), and 30-day readmission are common outcomes in the intensive care unit (ICU). Traditional scoring systems and machine learning models for predicting these outcomes usually ignore the characteristics of ICU data, which are time-series forms. We aimed to use time-series deep learning models with the selective combination of three widely used scoring systems to predict these outcomes.

Materials and methods

A retrospective cohort study was conducted on 40,083 patients in ICU from the Medical Information Mart for Intensive Care-IV (MIMIC-IV) database. Three deep learning models, namely, recurrent neural network (RNN), gated recurrent unit (GRU), and long short-term memory (LSTM) with attention mechanisms, were trained for the prediction of in-hospital mortality, prolonged LOS, and 30-day readmission with variables collected during the initial 24 h after ICU admission or the last 24 h before discharge. The inclusion of variables was based on three widely used scoring systems, namely, APACHE II, SOFA, and SAPS II, and the predictors consisted of time-series vital signs, laboratory tests, medication, and procedures. The patients were randomly divided into a training set (80%) and a test set (20%), which were used for model development and model evaluation, respectively. The area under the receiver operating characteristic curve (AUC), sensitivity, specificity, and Brier scores were used to evaluate model performance. Variable significance was identified through attention mechanisms.

Results

A total of 33 variables for 40,083 patients were enrolled for mortality and prolonged LOS prediction and 36,180 for readmission prediction. The rates of occurrence of the three outcomes were 9.74%, 27.54%, and 11.79%, respectively. In each of the three outcomes, the performance of RNN, GRU, and LSTM did not differ greatly. Mortality prediction models, prolonged LOS prediction models, and readmission prediction models achieved AUCs of 0.870 ± 0.001, 0.765 ± 0.003, and 0.635 ± 0.018, respectively. The top significant variables co-selected by the three deep learning models were Glasgow Coma Scale (GCS), age, blood urea nitrogen, and norepinephrine for mortality; GCS, invasive ventilation, and blood urea nitrogen for prolonged LOS; and blood urea nitrogen, GCS, and ethnicity for readmission.

Conclusion

The prognostic prediction models established in our study achieved good performance in predicting common outcomes of patients in ICU, especially in mortality prediction. In addition, GCS and blood urea nitrogen were identified as the most important factors strongly associated with adverse ICU events.