Developing well-calibrated illness severity scores for decision support in the critically ill

Illness severity assessment of older adults in critical illness using machine learning (ELDER-ICU): an international multicentre study with subgroup bias evaluation

BACKGROUND

Comorbidity, frailty, and decreased cognitive function lead to a higher risk of death in elderly patients (more than 65 years of age) during acute medical events. Early and accurate illness severity assessment can support appropriate decision making for clinicians caring for these patients. We aimed to develop ELDER-ICU, a machine learning model to assess the illness severity of older adults admitted to the intensive care unit (ICU) with cohort-specific calibration and evaluation for potential model bias.

METHODS

In this retrospective, international multicentre study, the ELDER-ICU model was developed using data from 14 US hospitals, and validated in 171 hospitals from the USA and Netherlands. Data were extracted from the Medical Information Mart for Intensive Care database, electronic ICU Collaborative Research Database, and Amsterdam University Medical Centers Database. We used six categories of data as predictors, including demographics and comorbidities, physical frailty, laboratory tests, vital signs, treatments, and urine output. Patient data from the first day of ICU stay were used to predict in-hospital mortality. We used the eXtreme Gradient Boosting algorithm (XGBoost) to develop models and the SHapley Additive exPlanations method to explain model prediction. The trained model was calibrated before internal, external, and temporal validation. The final XGBoost model was compared against three other machine learning algorithms and five clinical scores. We performed subgroup analysis based on age, sex, and race. We assessed the discrimination and calibration of models using the area under receiver operating characteristic (AUROC) and standardised mortality ratio (SMR) with 95% CIs.

FINDINGS

Using the development dataset (n=50 366) and predictive model building process, the XGBoost algorithm performed the best in all types of validations compared with other machine learning algorithms and clinical scores (internal validation with 5037 patients from 14 US hospitals, AUROC=0·866 [95% CI 0·851-0·880]; external validation in the US population with 20 541 patients from 169 hospitals, AUROC=0·838 [0·829-0·847]; external validation in European population with 2411 patients from one hospital, AUROC=0·833 [0·812-0·853]; temporal validation with 4311 patients from one hospital, AUROC=0·884 [0·869-0·897]). In the external validation set (US population), the median AUROCs of bias evaluations covering eight subgroups were above 0·81, and the overall SMR was 0·99 (0·96-1·03). The top ten risk predictors were the minimum Glasgow Coma Scale score, total urine output, average respiratory rate, mechanical ventilation use, best state of activity, Charlson Comorbidity Index score, geriatric nutritional risk index, code status, age, and maximum blood urea nitrogen. A simplified model containing only the top 20 features (ELDER-ICU-20) had similar predictive performance to the full model.

INTERPRETATION

The ELDER-ICU model reliably predicts the risk of in-hospital mortality using routinely collected clinical features. The predictions could inform clinicians about patients who are at elevated risk of deterioration. Prospective validation of this model in clinical practice and a process for continuous performance monitoring and model recalibration are needed.

FUNDING

National Institutes of Health, National Natural Science Foundation of China, National Special Health Science Program, Health Science and Technology Plan of Zhejiang Province, Fundamental Research Funds for the Central Universities, Drug Clinical Evaluate Research of Chinese Pharmaceutical Association, and National Key R&D Program of China.

Attention-based multimodal fusion with contrast for robust clinical prediction in the face of missing modalities

OBJECTIVE

With the increasing amount and growing variety of healthcare data, multimodal machine learning supporting integrated modeling of structured and unstructured data is an increasingly important tool for clinical machine learning tasks. However, it is non-trivial to manage the differences in dimensionality, volume, and temporal characteristics of data modalities in the context of a shared target task. Furthermore, patients can have substantial variations in the availability of data, while existing multimodal modeling methods typically assume data completeness and lack a mechanism to handle missing modalities.

METHODS

We propose a Transformer-based fusion model with modality-specific tokens that summarize the corresponding modalities to achieve effective cross-modal interaction accommodating missing modalities in the clinical context. The model is further refined by inter-modal, inter-sample contrastive learning to improve the representations for better predictive performance. We denote the model as Attention-based cRoss-MOdal fUsion with contRast (ARMOUR). We evaluate ARMOUR using two input modalities (structured measurements and unstructured text), six clinical prediction tasks, and two evaluation regimes, either including or excluding samples with missing modalities.

RESULTS

Our model shows improved performances over unimodal or multimodal baselines in both evaluation regimes, including or excluding patients with missing modalities in the input. The contrastive learning improves the representation power and is shown to be essential for better results. The simple setup of modality-specific tokens enables ARMOUR to handle patients with missing modalities and allows comparison with existing unimodal benchmark results.

CONCLUSION

We propose a multimodal model for robust clinical prediction to achieve improved performance while accommodating patients with missing modalities. This work could inspire future research to study the effective incorporation of multiple, more complex modalities of clinical data into a single model.

Evaluation of medication regimen complexity as a predictor for mortality

While medication regimen complexity, as measured by a novel medication regimen complexity-intensive care unit (MRC-ICU) score, correlates with baseline severity of illness and mortality, whether the MRC-ICU improves hospital mortality prediction is not known. After characterizing the association between MRC-ICU, severity of illness and hospital mortality we sought to evaluate the incremental benefit of adding MRC-ICU to illness severity-based hospital mortality prediction models. This was a single-center, observational cohort study of adult intensive care units (ICUs). A random sample of 991 adults admitted ≥ 24 h to the ICU from 10/2015 to 10/2020 were included. The logistic regression models for the primary outcome of mortality were assessed via area under the receiver operating characteristic (AUROC). Medication regimen complexity was evaluated daily using the MRC-ICU. This previously validated index is a weighted summation of medications prescribed in the first 24 h of ICU stay [e.g., a patient prescribed insulin (1 point) and vancomycin (3 points) has a MRC-ICU = 4 points]. Baseline demographic features (e.g., age, sex, ICU type) were collected and severity of illness (based on worst values within the first 24 h of ICU admission) was characterized using both the Acute Physiology and Chronic Health Evaluation (APACHE II) and the Sequential Organ Failure Assessment (SOFA) score. Univariate analysis of 991 patients revealed every one-point increase in the average 24-h MRC-ICU score was associated with a 5% increase in hospital mortality [Odds Ratio (OR) 1.05, 95% confidence interval 1.02-1.08, p = 0.002]. The model including MRC-ICU, APACHE II and SOFA had a AUROC for mortality of 0.81 whereas the model including only APACHE-II and SOFA had a AUROC for mortality of 0.76. Medication regimen complexity is associated with increased hospital mortality. A prediction model including medication regimen complexity only modestly improves hospital mortality prediction.

Development and Validation of a Dynamic Nomogram for Predicting in-Hospital Mortality in Patients with Acute Pancreatitis: A Retrospective Cohort Study in the Intensive Care Unit

Purpose

The aim of this study is to develop and validate a predictive model for the prediction of in-hospital mortality in patients with acute pancreatitis (AP) based on the intensive care database.

Patients and Methods

We analyzed the data of patients with AP in the Medical Information Mart for Intensive Care-IV (MIMIC-IV) database and Electronic Intensive Care Unit Collaborative Research Database (eICU-CRD). Then, patients from MIMIC-IV were divided into a development group and a validation group according to the ratio of 8:2, and eICU-CRD was assigned as an external validation group. Univariate logistic regression and least absolute shrinkage and selection operator regression were used for screening the best predictors, and multivariate logistic regression was used to establish a dynamic nomogram. We evaluated the discrimination, calibration, and clinical efficacy of the nomogram, and compared the performance of the nomogram with Acute Physiology and Chronic Health Evaluation II (APACHE-II) score and Bedside Index of Severity in AP (BISAP) score.

Results

A total of 1030 and 514 patients with AP in MIMIC-IV database and eICU-CRD were included in the study. After stepwise analysis, 8 out of a total of 37 variables were selected to construct the nomogram. The dynamic nomogram can be obtained by visiting https://model.sci-inn.com/KangZou/. The area under receiver operating characteristic curve (AUC) of the nomogram was 0.859, 0.871, and 0.847 in the development, internal, and external validation set respectively. The nomogram had a similar performance with APACHE-II (AUC = 0.841, p = 0.537) but performed better than BISAP (AUC = 0.690, p = 0.001) score in the validation group. Moreover, the calibration curve presented a satisfactory predictive accuracy, and the decision curve analysis suggested great clinical application value of the nomogram.

Conclusion

Based on the results of internal and external validation, the nomogram showed favorable discrimination, calibration, and clinical practicability in predicting the in-hospital mortality of patients with AP.

Development, validation, and feature extraction of a deep learning model predicting in-hospital mortality using Japan's largest national ICU database: a validation framework for transparent clinical Artificial Intelligence (cAI) development

OBJECTIVE

While clinical Artificial Intelligence (cAI) mortality prediction models and relevant studies have increased, limitations including the lack of external validation studies and inadequate model calibration leading to decreased overall accuracy have been observed. To combat this, we developed and evaluated a novel deep neural network (DNN) and a validation framework to promote transparent cAI development.

METHODS

Data from Japan's largest ICU database was used to develop the DNN model, predicting in-hospital mortality including ICU and post-ICU mortality by days since ICU discharge. The most important variables to the model were extracted with SHapley Additive exPlanations (SHAP) to examine the DNN's efficacy as well as develop models that were also externally validated.

MAIN RESULTS

The area under the receiver operating characteristic curve (AUC) for predicting ICU mortality was 0.94 [0.93-0.95], and 0.91 [0.90-0.92] for in-hospital mortality, ranging between 0.91-0.95 throughout one year since ICU discharge. An external validation using only the top 20 variables resulted with higher AUCs than traditional severity scores.

CONCLUSIONS

Our DNN model consistently generated AUCs between 0.91-0.95 regardless of days since ICU discharge. The 20 most important variables to our DNN, also generated higher AUCs than traditional severity scores regardless of days since ICU discharge. To our knowledge, this is the first study that predicts ICU and in-hospital mortality using cAI by post-ICU discharge days up to over a year. This finding could contribute to increased transparency on cAI applications.

Association of APACHE-II Scores With 30-Day Mortality After Tracheostomy: A Retrospective Study

OBJECTIVE

The objective of this study was to assess whether the Acute Physiology, Age, Chronic Health Evaluation II (APACHE-II) score is a reliable predictor of 30-day mortality in the setting of adult patients with ventilator-dependent respiratory failure (VDRF) who undergo tracheostomy.

METHODS

This is a retrospective, single-institution study. Potential subjects were identified using the current procedural terminology codes for the tracheostomy procedure and International Classification of Diseases, 10th Revision, codes for VDRF. APACHE-II scores were retrospectively calculated. Tracheostomies were performed in our population over an 18-month period (November 2018 through April 2020). Our study population did not include patients with novel coronavirus. The primary outcome was mortality at 30 days after tracheostomy.

RESULTS

A total of 238 patients with VDRF who had a tracheostomy were included in this study. Twenty-eight (11.8%) patients died within 30 days of tracheostomy. The mean (standard deviation) APACHE-II score was 22.5 (10.2) for patients who died within 30 days of tracheostomy and 19.8 (7.4) for patients living within 30 days of tracheostomy (p = 0.30). Patients with APACHE-II scores greater than or equal to 30 showed higher odds of death within 30 days of tracheostomy (odds ratio, 3.0; 95% CI, 1.14-7.89, p = 0.03).

CONCLUSION

An APACHE-II score of 30 and above is associated with mortality within 30 days of tracheostomy in patients with VDRF. APACHE-II scores may be a promising tool for assessing risk of mortality in patients with VDRF after tracheostomy.

LEVEL OF EVIDENCE

3 Laryngoscope, 133:273-278, 2023.

Multisystemic Inflammatory Syndrome in Children: a step towards a better understanding of this entity

As pediatricians, we all have to deal with new childhood inflammatory disorder due to COVID 19: the Multisystem Inflammatory Syndrome in Children (MIS-C). The recent article by Savorgnan et al. on the physiologic profiles associated with MIS-C proposed a classification through the "MIS-C severity score" (MSS). The authors also identified a combination of seven variables collected during the first 3 h of admission in the PICU that contributes to stratify MIS-C severity with an area under the receiver operating characteristic curve (AUC) >0.90. This work represents an important first step in the development of a MIS-C severity score and is a call for collaborative groups to validate the prediction model through multicenter studies and thereby refine the management of MIS-C. IMPACT: The recent article by Savorgnan et al. on physiologic profile associated with MIS-C represents an important first step in the development of an MIS-C severity score and is a call for collaborative groups to validate the prediction model through multicenter studies and thereby refine the management of MIS-C. Our manuscript helps in the methodology interpretation of the manuscript by Savorgnan et al. And our manuscript promotes collaborative work on MIS-C.

Outcome Prediction in Critically-Ill Patients with Venous Thromboembolism and/or Cancer Using Machine Learning Algorithms: External Validation and Comparison with Scoring Systems

Intensive care unit (ICU) patients with venous thromboembolism (VTE) and/or cancer suffer from high mortality rates. Mortality prediction in the ICU has been a major medical challenge for which several scoring systems exist but lack in specificity. This study focuses on two target groups, namely patients with thrombosis or cancer. The main goal is to develop and validate interpretable machine learning (ML) models to predict early and late mortality, while exploiting all available data stored in the medical record. To this end, retrospective data from two freely accessible databases, MIMIC-III and eICU, were used. Well-established ML algorithms were implemented utilizing automated and purposely built ML frameworks for addressing class imbalance. Prediction of early mortality showed excellent performance in both disease categories, in terms of the area under the receiver operating characteristic curve (AUC–ROC): VTE-MIMIC-III 0.93, eICU 0.87, cancer-MIMIC-III 0.94. On the other hand, late mortality prediction showed lower performance, i.e., AUC–ROC: VTE 0.82, cancer 0.74–0.88. The predictive model of early mortality developed from 1651 VTE patients (MIMIC-III) ended up with a signature of 35 features and was externally validated in 2659 patients from the eICU dataset. Our model outperformed traditional scoring systems in predicting early as well as late mortality. Novel biomarkers, such as red cell distribution width, were identified.

Developing machine learning models for prediction of mortality in the medical intensive care unit

BACKGROUND AND OBJECTIVE

Alert of patient deterioration is essential for prompt medical intervention in the Medical Intensive Care Unit (MICU). Logistic Regression (LR) has been used for the development of most conventional severity-of-illness scoring systems to anticipate the risk of mortality in the MICU. Machine Learning (ML) models such as probabilistic graphical models and Extreme Gradient Boosting (XGB) have demonstrated improved prediction accuracy in patient outcomes compared to LR. The aim was to compare three ML models to the SAPS, SAPS II, SAPS III, SOFA, serial SOFA, LODS, and OASIS for prediction of MICU mortality.

METHODS

A Bayesian Network (BN), Naïve Bayes network (NB), and a XGB model were developed. 9893 adult MICU-stays from the MIMIC-III database were studied. The primary outcome was MICU mortality prediction and the secondary outcome was 1-year mortality prediction. Data analyzed consisted on routine physiological measurements collected during 5 hours in the MICU, demographic and diagnoses/procedure features. The performance was evaluated by accuracy statistics, discrimination and calibration measures. Limitations of the study were discussed.

RESULTS

The AUROC for MICU mortality prediction was 0.919 for XGB, 0.905 for BN, and 0.864 for NB, while the conventional systems displayed much lower values with the serial SOFA having the best value (0.814). The Diagnostic Odds Ratio was ≤7.099 for all the conventional systems, reaching values of 30.115 for XGB and 22.648 for BN. The XGB achieved a sensitivity of 0.831 and specificity of 0.86 assuring an acceptable precision (0.528), whose values were much lower for the conventional systems. The Brier score was better for the ML models, except for the NB (0.119), with 0.072 for XGB and 0.081 for BN.

CONCLUSIONS

The XGB and BN substantially outperformed the conventional systems for discrimination, calibration and the accuracy statistics assessed. The NB showed inferior performance to the XGB and BN but improved the discrimination and all accuracy statistics of the conventional systems except for an inferior calibration and 1-year mortality discrimination. The XGB showed the best performance among all models. These ML models have the potential to improve the monitoring of MICU patients, which must be evaluated in future studies.

Identifying early-measured variables associated with APACHE IVa providing incorrect in-hospital mortality predictions for critical care patients

APACHE IVa provides typically useful and accurate predictions on in-hospital mortality and length of stay for patients in critical care. However, there are factors which may preclude APACHE IVa from reaching its ceiling of predictive accuracy. Our primary aim was to determine which variables available within the first 24 h of a patient’s ICU stay may be indicative of the APACHE IVa scoring system making occasional but potentially illuminating errors in predicting in-hospital mortality. We utilized the publicly available multi-institutional ICU database, eICU, available since 2018, to identify a large observational cohort for our investigation. APACHE IVa scores are provided by eICU for each patient’s ICU stay. We used Lasso logistic regression in an aim to build parsimonious final models, using cross-validation to select the penalization parameter, separately for each of our two responses, i.e., errors, of interest, which are APACHE falsely predicting in-hospital death (Type I error), and APACHE falsely predicting in-hospital survival (Type II error). We then assessed the performance of the models with a random holdout validation sample. While the extremeness of the APACHE prediction led to dependable predictions for preventing either type of error, distinct variables were identified as being strongly associated with the two different types of errors occurring. These included a primary set of predictors consisting of mean SpO2 and worst lactate for predicting Type I errors, and worst albumin and mean heart rate for Type II. In addition, a secondary set of predictors including changes recorded in care limitations for the patient’s treatment plan, worst pH, whether cardiac arrest occurred at admission, and whether vasopressor was provided for predicting Type I error; age, whether the patient was ventilated in day 1, mean respiratory rate, worst lactate, worst blood urea nitrogen test, and mean aperiodic vitals for Type II. The two models also differed in their performance metrics in their holdout validation samples, in large part due to the lower prevalence of Type II errors compared to Type I. The eICU database was a good resource for evaluating our objective, and important recommendations are provided, particularly identifying key variables that could lead to APACHE prediction errors when APACHE scores are sufficiently low to predict in-hospital survival.

Prediction of in-hospital mortality of Clostriodiodes difficile infection using critical care database: a big data-driven, machine learning approach

RESEARCH OBJECTIVES

Clostriodiodes difficile infection (CDI) is a major cause of healthcare-associated diarrhoea with high mortality. There is a lack of validated predictors for severe outcomes in CDI. The aim of this study is to derive and validate a clinical prediction tool for CDI in-hospital mortality using a large critical care database.

METHODOLOGY

The demographics, clinical parameters, laboratory results and mortality of CDI were extracted from the Medical Information Mart for Intensive Care-III (MIMIC-III) database. We subsequently trained three machine learning models: logistic regression (LR), random forest (RF) and gradient boosting machine (GBM) to predict in-hospital mortality. The individual performances of the models were compared against current severity scores (Clostridiodes difficile Associated Risk of Death Score (CARDS) and ATLAS (Age, Treatment with systemic antibiotics, leukocyte count, Albumin and Serum creatinine as a measure of renal function) by calculating area under receiver operating curve (AUROC). We identified factors associated with higher mortality risk in each model.

SUMMARY OF RESULTS

From 61 532 intensive care unit stays in the MIMIC-III database, there were 1315 CDI cases. The mortality rate for CDI in the study cohort was 18.33%. AUROC was 0.69 (95% CI, 0.60 to 0.76) for LR, 0.71 (95% CI, 0.62 to 0.77) for RF and 0.72 (95% CI, 0.64 to 0.78) for GBM, while previously AUROC was 0.57 (95% CI, 0.51 to 0.65) for CARDS and 0.63 (95% CI, 0.54 to 0.70) for ATLAS. Albumin, lactate and bicarbonate were significant mortality factors for all the models. Free calcium, potassium, white blood cell, urea, platelet and mean blood pressure were present in at least two of the three models.

CONCLUSION

Our machine learning derived CDI in-hospital mortality prediction model identified pertinent factors that can assist critical care clinicians in identifying patients at high risk of dying from CDI.

Utilizing Artificial Intelligence in Critical Care: Adding A Handy Tool to Our Armamentarium

We have witnessed rapid advancement in technology over the last few decades. With the advent of artificial intelligence (AI), newer avenues have opened for researchers. AI has added an entirely new dimension to this technological boom. Researchers in medical science have been excited about the tantalizing prospect of utilizing AI for the benefit of patient care. Lately, we have come across studies trying to test and validate various models based on AI to improve patient care strategies in critical care medicine as well. Thus, in this review, we will attempt to succinctly review current literature discussing AI in critical care medicine and analyze its future utility based on prevailing evidence.

Characterizing the Patients, Hospitals, and Data Quality of the eICU Collaborative Research Database

OBJECTIVES

The eICU Collaborative Research Database is a publicly available repository of granular data from more than 200,000 ICU admissions. The quantity and variety of its entries hold promise for observational critical care research. We sought to understand better the data available within this resource to guide its future use.

DESIGN

We conducted a descriptive analysis of the eICU Collaborative Research Database, including patient, practitioner, and hospital characteristics. We investigated the completeness of demographic and hospital data, as well as those values required to calculate an Acute Physiology and Chronic Health Evaluation score. We also assessed the rates of ventilation, intubation, and dialysis, and looked for potential errors in the vital sign data.

SETTING

American ICUs that participated in the Philips Healthcare eICU program between 2014 and 2015.

PATIENTS

A total of 139,367 individuals who were admitted to one of the 335 participating ICUs between 2014 and 2015.

INTERVENTIONS

None.

MEASUREMENTS AND MAIN RESULTS

Most encounters were from small- and medium-sized hospitals, and managed by nonintensivists. The median ICU length of stay was 1.57 days (interquartile range, 0.82-2.97 d). The median Acute Physiology and Chronic Health Evaluation IV-predicted ICU mortality was 2.2%, with an observed mortality of 5.4%. Rates of ventilation (20-33%), intubation (15-24%), and dialysis (3-5%) varied according to the query method used. Most vital sign readings fell into realistic ranges, with manually curated data less likely to contain implausible results than automatically entered data.

CONCLUSIONS

Data in the eICU Collaborative Research Database are for the most part complete and plausible. Some ambiguity exists in determining which encounters are associated with various interventions, most notably mechanical ventilation. Caution is warranted in extrapolating findings from the eICU Collaborative Research Database to larger ICUs with higher acuity.

Explainable Machine Learning Model for Predicting GI Bleed Mortality in the Intensive Care Unit

INTRODUCTION

Acute gastrointestinal (GI) bleed is a common reason for hospitalization with 2%-10% risk of mortality. In this study, we developed a machine learning (ML) model to calculate the risk of mortality in intensive care unit patients admitted for GI bleed and compared it with APACHE IVa risk score. We used explainable ML methods to provide insight into the model's prediction and outcome.

METHODS

We analyzed the patient data in the Electronic Intensive Care Unit Collaborative Research Database and extracted data for 5,691 patients (mean age = 67.4 years; 61% men) admitted with GI bleed. The data were used in training a ML model to identify patients who died in the intensive care unit. We compared the predictive performance of the ML model with the APACHE IVa risk score. Performance was measured by area under receiver operating characteristic curve (AUC) analysis. This study also used explainable ML methods to provide insights into the model's outcome or prediction using the SHAP (SHapley Additive exPlanations) method.

RESULTS

The ML model performed better than the APACHE IVa risk score in correctly classifying the low-risk patients. The ML model had a specificity of 27% (95% confidence interval [CI]: 25-36) at a sensitivity of 100% compared with the APACHE IVa score, which had a specificity of 4% (95% CI: 3-31) at a sensitivity of 100%. The model identified patients who died with an AUC of 0.85 (95% CI: 0.80-0.90) in the internal validation set, whereas the APACHE IVa clinical scoring systems identified patients who died with AUC values of 0.80 (95% CI: 0.73-0.86) with P value <0.001.

DISCUSSION

We developed a ML model that predicts the mortality in patients with GI bleed with a greater accuracy than the current scoring system. By making the ML model explainable, clinicians would be able to better understand the reasoning behind the outcome.

Linking of global intensive care (LOGIC): An international benchmarking in critical care initiative

Benchmarking is a common and effective method for measuring and analyzing ICU performance. With the existence of national registries, objective information can now be obtained to allow benchmarking of ICU care within and between countries. The present manuscript briefly describes the current status of benchmarking in healthcare and critical care and presents the LOGIC project, an initiative to promote international benchmarking for intensive care units. Currently 13 registries have joined LOGIC. We showed large differences in the utilization of ICU as well as resources and in outcomes. Despite the need for careful interpretation of differences due to variation in definitions and limited risk adjustment, LOGIC is a growing worldwide initiative that allows access to insightful epidemiologic data from ICUs in multiple databases and registries.

Risk of mortality and cardiopulmonary arrest in critical patients presenting to the emergency department using machine learning and natural language processing

Emergency department triage is the first point in time when a patient's acuity level is determined. The time to assign a priority at triage is short and it is vital to accurately stratify patients at this stage, since under-triage can lead to increased morbidity, mortality and costs. Our aim was to present a model that can assist healthcare professionals in triage decision making, namely in the stratification of patients through the risk prediction of a composite critical outcome-mortality and cardiopulmonary arrest. Our study cohort consisted of 235826 adult patients triaged at a Portuguese Emergency Department from 2012 to 2016. Patients were assigned to emergent, very urgent or urgent priorities of the Manchester Triage System (MTS). Demographics, clinical variables routinely collected at triage and the patients' chief complaint were used. Logistic regression, random forests and extreme gradient boosting were developed using all available variables. The term frequency-inverse document frequency (TF-IDF) natural language processing weighting factor was applied to vectorize the chief complaint. Stratified random sampling was used to split the data into train (70%) and test (30%) data sets. Ten-fold cross validation was performed in train to optimize model hyper-parameters. The performance obtained with the best model was compared against the reference model-a regularized logistic regression trained using only triage priorities. Extreme gradient boosting exhibited good calibration properties and yielded areas under the receiver operating characteristic and precision-recall curves of 0.96 (95% CI 0.95-0.97) and 0.31 (95% CI 0.26-0.36), respectively. The predictors ranked with higher importance by this model were the Glasgow coma score, the patients' age, pulse oximetry and arrival mode. Compared to the reference, the extreme gradient boosting model using clinical variables and the chief complaint presented higher recall for patients assigned MTS-3 and can identify those who are at risk of the composite outcome.

Screening for diabetic retinopathy: new perspectives and challenges

Although the prevalence of all stages of diabetic retinopathy has been declining since 1980 in populations with improved diabetes control, the crude prevalence of visual impairment and blindness caused by diabetic retinopathy worldwide increased between 1990 and 2015, largely because of the increasing prevalence of type 2 diabetes, particularly in low-income and middle-income countries. Screening for diabetic retinopathy is essential to detect referable cases that need timely full ophthalmic examination and treatment to avoid permanent visual loss. In the past few years, personalised screening intervals that take into account several risk factors have been proposed, with good cost-effectiveness ratios. However, resources for nationwide screening programmes are scarce in many countries. New technologies, such as scanning confocal ophthalmology with ultrawide field imaging and handheld mobile devices, teleophthalmology for remote grading, and artificial intelligence for automated detection and classification of diabetic retinopathy, are changing screening strategies and improving cost-effectiveness. Additionally, emerging evidence suggests that retinal imaging could be useful for identifying individuals at risk of cardiovascular disease or cognitive impairment, which could expand the role of diabetic retinopathy screening beyond the prevention of sight-threatening disease.

Machine Learning for Pulmonary and Critical Care Medicine: A Narrative Review

Machine learning (ML) is a discipline of computer science in which statistical methods are applied to data in order to classify, predict, or optimize, based on previously observed data. Pulmonary and critical care medicine have seen a surge in the application of this methodology, potentially delivering improvements in our ability to diagnose, treat, and better understand a multitude of disease states. Here we review the literature and provide a detailed overview of the recent advances in ML as applied to these areas of medicine. In addition, we discuss both the significant benefits of this work as well as the challenges in the implementation and acceptance of this non-traditional methodology for clinical purposes.

Developing well-calibrated illness severity scores for decision support in the critically ill

Illness severity scores are regularly employed for quality improvement and benchmarking in the intensive care unit, but poor generalization performance, particularly with respect to probability calibration, has limited their use for decision support. These models tend to perform worse in patients at a high risk for mortality. We hypothesized that a sequential modeling approach wherein an initial regression model assigns risk and all patients deemed high risk then have their risk quantified by a second, high-risk-specific, regression model would result in a model with superior calibration across the risk spectrum. We compared this approach to a logistic regression model and a sophisticated machine learning approach, the gradient boosting machine. The sequential approach did not have an effect on the receiver operating characteristic curve or the precision-recall curve but resulted in improved reliability curves. The gradient boosting machine achieved a small improvement in discrimination performance and was similarly calibrated to the sequential models.