Prospective and External Evaluation of a Machine Learning Model to Predict In-Hospital Mortality of Adults at Time of Admission

Early Prognostication of Critical Patients With Spinal Cord Injury: A Machine Learning Study With 1485 Cases

STUDY DESIGN

A retrospective case-series.

OBJECTIVE

The study aims to use machine learning to predict the discharge destination of spinal cord injury (SCI) patients in the intensive care unit.

SUMMARY OF BACKGROUND DATA

Prognostication following SCI is vital, especially for critical patients who need intensive care.

PATIENTS AND METHODS

Clinical data of patients diagnosed with SCI were extracted from a publicly available intensive care unit database. The first recorded data of the included patients were used to develop a total of 98 machine learning classifiers, seeking to predict discharge destination (eg, death, further medical care, home, etc.). The microaverage area under the curve (AUC) was the main indicator to assess discrimination. The best average-AUC classifier and the best death-sensitivity classifier were integrated into an ensemble classifier. The discrimination of the ensemble classifier was compared with top death-sensitivity classifiers and top average-AUC classifiers. In addition, prediction consistency and clinical utility were also assessed.

RESULTS

A total of 1485 SCI patients were included. The ensemble classifier had a microaverage AUC of 0.851, which was only slightly inferior to the best average-AUC classifier ( P =0.10). The best average-AUC classifier death sensitivity was much lower than that of the ensemble classifier. The ensemble classifier had a death sensitivity of 0.452, which was inferior to the top 8 death-sensitivity classifiers, whose microaverage AUC were inferior to the ensemble classifier ( P <0.05). In addition, the ensemble classifier demonstrated a comparable Brier score and superior net benefit in the DCA when compared with the performance of the origin classifiers.

CONCLUSIONS

The ensemble classifier shows an overall superior performance in predicting discharge destination, considering discrimination ability, prediction consistency, and clinical utility. This classifier system may aid in the clinical management of critical SCI patients in the early phase following injury.

LEVEL OF EVIDENCE

Level 3.

Vancomycin: An analysis and evaluation of eight population pharmacokinetic models for clinical application in general adult population

INTRODUCTION

Based on the recent guidelines for vancomycin therapeutic drug monitoring (TDM), the area under the curve to minimum inhibitory concentration ratio was to be employed combined with the usage of population pharmacokinetic (popPK) model for dosing adaptation. Yet, deploying these models in a clinical setting requires an external evaluation of their performance.

OBJECTIVES

This study aimed to evaluate existing vancomycin popPK models from the literature for the use in TDM within the general patient population in a clinical setting.

METHODS

The models under external evaluation were chosen based on a review of literature covering vancomycin popPK models developed in general adult populations. Patients' data were collected from Charles-Le Moyne Hospital (CLMH). The external evaluation was performed with NONMEM® (v7.5). Additional analyses such as evaluating the impact of number of samples on external evaluation, Bayesian forecasting, and a priori dosing regimen simulations were performed on the best performing model.

RESULTS

Eight popPK models were evaluated with an independent dataset that included 40 patients and 252 samples. The model developed by Goti and colleagues demonstrated superior performance in diagnostic plots and population predictive performance, with bias and inaccuracy values of 0.251% and 22.7%, respectively, and for individual predictive performance, bias and inaccuracy were -4.90% and 12.1%, respectively. When limiting the independent dataset to one or two samples per patient, the Goti model exhibited inadequate predictive performance for inaccuracy, with values exceeding 30%. Moreover, the Goti model is suitable for Bayesian forecasting with at least two samples as prior for the prediction of the next trough concentration. Furthermore, the vancomycin dosing regimen that would maximize therapeutic targets of area under the curve to minimum inhibitory concentration ratio (AUC24/MIC) and trough concentrations (Ctrough) for the Goti model was 20 mg/kg/dose twice daily.

CONCLUSION

Considering the superior predictive performance and potential for Bayesian forecasting in the Goti model, future research aims to test its applicability in clinical settings at CLMH, both in a priori and a posteriori scenario.

Predicting Colonic Neoplasia Surgical Complications: A Machine Learning Approach

BACKGROUND

A range of statistical approaches have been used to help predict outcomes associated with colectomy. The multifactorial nature of complications suggests that machine learning algorithms may be more accurate in determining postoperative outcomes by detecting nonlinear associations, which are not readily measured by traditional statistics.

OBJECTIVE

The aim of this study was to investigate the utility of machine learning algorithms to predict complications in patients undergoing colectomy for colonic neoplasia.

DESIGN

Retrospective analysis using decision tree, random forest, and artificial neural network classifiers to predict postoperative outcomes.

SETTINGS

National Inpatient Sample database (2003-2017).

PATIENTS

Adult patients who underwent elective colectomy with anastomosis for neoplasia.

MAIN OUTCOME MEASURES

Performance was quantified using sensitivity, specificity, accuracy, and area under the receiver operating characteristic curve to predict the incidence of anastomotic leak, prolonged length of stay, and inpatient mortality.

RESULTS

A total of 14,935 patients (4731 laparoscopic, 10,204 open) were included. They had an average age of 67 ± 12.2 years, and 53% of patients were women. The 3 machine learning models successfully identified patients who developed the measured complications. Although differences between model performances were largely insignificant, the neural network scored highest for most outcomes: predicting anastomotic leak, area under the receiver operating characteristic curve 0.88/0.93 (open/laparoscopic, 95% CI, 0.73-0.92/0.80-0.96); prolonged length of stay, area under the receiver operating characteristic curve 0.84/0.88 (open/laparoscopic, 95% CI, 0.82-0.85/0.85-0.91); and inpatient mortality, area under the receiver operating characteristic curve 0.90/0.92 (open/laparoscopic, 95% CI, 0.85-0.96/0.86-0.98).

LIMITATIONS

The patients from the National Inpatient Sample database may not be an accurate sample of the population of all patients undergoing colectomy for colonic neoplasia and does not account for specific institutional and patient factors.

CONCLUSIONS

Machine learning predicted postoperative complications in patients with colonic neoplasia undergoing colectomy with good performance. Although validation using external data and optimization of data quality will be required, these machine learning tools show great promise in assisting surgeons with risk-stratification of perioperative care to improve postoperative outcomes. See Video Abstract .

PREDICCIN DE LAS COMPLICACIONES QUIRRGICAS DE LA NEOPLASIA DE COLON UN ENFOQUE DE MODELO DE APRENDIZAJE AUTOMTICO

ANTECEDENTES:Se han utilizado una variedad de enfoques estadísticos para ayudar a predecir los resultados asociados con la colectomía. La naturaleza multifactorial de las complicaciones sugiere que los algoritmos de aprendizaje automático pueden ser más precisos en determinar los resultados posoperatorios al detectar asociaciones no lineales, que generalmente no se miden en las estadísticas tradicionales.OBJETIVO:El objetivo de este estudio fue investigar la utilidad de los algoritmos de aprendizaje automático para predecir complicaciones en pacientes sometidos a colectomía por neoplasia de colon.DISEÑO:Análisis retrospectivo utilizando clasificadores de árboles de decisión, bosques aleatorios y redes neuronales artificiales para predecir los resultados posoperatorios.AJUSTE:Base de datos de la Muestra Nacional de Pacientes Hospitalizados (2003-2017).PACIENTES:Pacientes adultos sometidos a colectomía electiva con anastomosis por neoplasia.INTERVENCIONES:N/A.PRINCIPALES MEDIDAS DE RESULTADO:El rendimiento se cuantificó utilizando la sensibilidad, especificidad, precisión y la característica operativa del receptor del área bajo la curva para predecir la incidencia de fuga anastomótica, duración prolongada de la estancia hospitalaria y mortalidad de los pacientes hospitalizados.RESULTADOS:Se incluyeron un total de 14.935 pacientes (4.731 laparoscópicos, 10.204 abiertos). Presentaron una edad promedio de 67 ± 12,2 años y el 53% eran mujeres. Los tres modelos de aprendizaje automático identificaron con éxito a los pacientes que desarrollaron las complicaciones medidas. Aunque las diferencias entre el rendimiento del modelo fueron en gran medida insignificantes, la red neuronal obtuvo la puntuación más alta para la mayoría de los resultados: predicción de fuga anastomótica, característica operativa del receptor del área bajo la curva 0,88/0,93 (abierta/laparoscópica, IC del 95%: 0,73-0,92/0,80-0,96); duración prolongada de la estancia hospitalaria, característica operativa del receptor del área bajo la curva 0,84/0,88 (abierta/laparoscópica, IC del 95%: 0,82-0,85/0,85-0,91); y mortalidad de pacientes hospitalizados, característica operativa del receptor del área bajo la curva 0,90/0,92 (abierto/laparoscópico, IC del 95%: 0,85-0,96/0,86-0,98).LIMITACIONES:Los pacientes de la base de datos de la Muestra Nacional de Pacientes Hospitalizados pueden no ser una muestra precisa de la población de todos los pacientes sometidos a colectomía por neoplasia de colon y no tienen en cuenta factores institucionales y específicos del paciente.CONCLUSIONES:El aprendizaje automático predijo con buen rendimiento las complicaciones postoperatorias en pacientes con neoplasia de colon sometidos a colectomía. Aunque será necesaria la validación mediante datos externos y la optimización de la calidad de los datos, estas herramientas de aprendizaje automático son muy prometedoras para ayudar a los cirujanos con la estratificación de riesgos de la atención perioperatoria para mejorar los resultados posoperatorios. (Traducción-Dr. Fidel Ruiz Healy ).

AI in medical diagnosis: AI prediction & human judgment

AI has long been regarded as a panacea for decision-making and many other aspects of knowledge work; as something that will help humans get rid of their shortcomings. We believe that AI can be a useful asset to support decision-makers, but not that it should replace decision-makers. Decision-making uses algorithmic analysis, but it is not solely algorithmic analysis; it also involves other factors, many of which are very human, such as creativity, intuition, emotions, feelings, and value judgments. We have conducted semi-structured open-ended research interviews with 17 dermatologists to understand what they expect from an AI application to deliver to medical diagnosis. We have found four aggregate dimensions along which the thinking of dermatologists can be described: the ways in which our participants chose to interact with AI, responsibility, 'explainability', and the new way of thinking (mindset) needed for working with AI. We believe that our findings will help physicians who might consider using AI in their diagnosis to understand how to use AI beneficially. It will also be useful for AI vendors in improving their understanding of how medics want to use AI in diagnosis. Further research will be needed to examine if our findings have relevance in the wider medical field and beyond.

Developing machine learning models to predict primary graft dysfunction after lung transplantation

Primary graft dysfunction (PGD) is the leading cause of morbidity and mortality in the first 30 days after lung transplantation. Risk factors for the development of PGD include donor and recipient characteristics, but how multiple variables interact to impact the development of PGD and how clinicians should consider these in making decisions about donor acceptance remain unclear. This was a single-center retrospective cohort study to develop and evaluate machine learning pipelines to predict the development of PGD grade 3 within the first 72 hours of transplantation using donor and recipient variables that are known at the time of donor offer acceptance. Among 576 bilateral lung recipients, 173 (30%) developed PGD grade 3. The cohort underwent a 75% to 25% train-test split, and lasso regression was used to identify 11 variables for model development. A K-nearest neighbor's model showing the best calibration and performance with relatively small confidence intervals was selected as the final predictive model with an area under the receiver operating characteristics curve of 0.65. Machine learning models can predict the risk for development of PGD grade 3 based on data available at the time of donor offer acceptance. This may improve donor-recipient matching and donor utilization in the future.

Value Creation Through Artificial Intelligence and Cardiovascular Imaging: A Scientific Statement From the American Heart Association

Multiple applications for machine learning and artificial intelligence (AI) in cardiovascular imaging are being proposed and developed. However, the processes involved in implementing AI in cardiovascular imaging are highly diverse, varying by imaging modality, patient subtype, features to be extracted and analyzed, and clinical application. This article establishes a framework that defines value from an organizational perspective, followed by value chain analysis to identify the activities in which AI might produce the greatest incremental value creation. The various perspectives that should be considered are highlighted, including clinicians, imagers, hospitals, patients, and payers. Integrating the perspectives of all health care stakeholders is critical for creating value and ensuring the successful deployment of AI tools in a real-world setting. Different AI tools are summarized, along with the unique aspects of AI applications to various cardiac imaging modalities, including cardiac computed tomography, magnetic resonance imaging, and positron emission tomography. AI is applicable and has the potential to add value to cardiovascular imaging at every step along the patient journey, from selecting the more appropriate test to optimizing image acquisition and analysis, interpreting the results for classification and diagnosis, and predicting the risk for major adverse cardiac events.

Scope of practice regulation in medicine: balancing patient safety, access to care and professional autonomy

Scope of practice regulation in medicine is crucial for ensuring patient safety, access to care and professional autonomy. This paper explores the impact of scope of practice regulation on healthcare delivery, professional responsibilities and patient outcomes. It discusses the variability in standards for safe practice, the challenges in defining boundaries between medical specialties and the recent controversies in cosmetic surgery practice. The paper also examines the potential benefits and drawbacks of rigorous scope of practice regulations, including their impact on clinical innovation, flexibility and access to care. Furthermore, it delves into the implications of defensive medicine and the consequences of restrictive regulations on patient care. The author proposes implementing a proactive, national, artificial intelligence-powered, real-time outcome monitoring system to address these challenges. This system aims to cover every patient undergoing a surgical procedure and could be gradually extended to non-surgical conditions, benefiting all key stakeholders in the health system. The paper emphasises the need for a balanced approach to scope of practice regulation to avoid stifling clinical innovation and professional autonomy, while ensuring patient safety and professional accountability.

External evaluation of the Dynamic Criticality Index: A machine learning model to predict future need for ICU care in hospitalized pediatric patients

OBJECTIVE

To assess the single site performance of the Dynamic Criticality Index (CI-D) models developed from a multi-institutional database to predict future care. Secondarily, to assess future care-location predictions in a single institution when CI-D models are re-developed using single-site data with identical variables and modeling methods. Four CI-D models were assessed for predicting care locations >6-12 hours, >12-18 hours, >18-24 hours, and >24-30 hours in the future.

DESIGN

Prognostic study comparing multi-institutional CI-D models' performance in a single-site electronic health record dataset to an institution-specific CI-D model developed using identical variables and modelling methods. The institution did not participate in the multi-institutional dataset.

PARTICIPANTS

All pediatric inpatients admitted from January 1st 2018 -February 29th 2020 through the emergency department.

MAIN OUTCOME(S) AND MEASURE(S)

The main outcome was inpatient care in routine or ICU care locations.

RESULTS

A total of 29,037 pediatric hospital admissions were included, with 5,563 (19.2%) admitted directly to the ICU, 869 (3.0%) transferred from routine to ICU care, and 5,023 (17.3%) transferred from ICU to routine care. Patients had a median [IQR] age 68 months (15-157), 47.5% were female and 43.4% were black. The area under the receiver operating characteristic curve (AUROC) for the multi-institutional CI-D models applied to a single-site test dataset was 0.493-0.545 and area under the precision-recall curve (AUPRC) was 0.262-0.299. The single-site CI-D models applied to an independent single-site test dataset had an AUROC 0.906-0.944 and AUPRC range from 0.754-0.824. Accuracy at 0.95 sensitivity for those transferred from routine to ICU care was 72.6%-81.0%. Accuracy at 0.95 specificity was 58.2%-76.4% for patients who transferred from ICU to routine care.

CONCLUSION AND RELEVANCE

Models developed from multi-institutional datasets and intended for application to individual institutions should be assessed locally and may benefit from re-development with site-specific data prior to deployment.

Systematic review and longitudinal analysis of implementing Artificial Intelligence to predict clinical deterioration in adult hospitals: what is known and what remains uncertain

OBJECTIVE

To identify factors influencing implementation of machine learning algorithms (MLAs) that predict clinical deterioration in hospitalized adult patients and relate these to a validated implementation framework.

MATERIALS AND METHODS

A systematic review of studies of implemented or trialed real-time clinical deterioration prediction MLAs was undertaken, which identified: how MLA implementation was measured; impact of MLAs on clinical processes and patient outcomes; and barriers, enablers and uncertainties within the implementation process. Review findings were then mapped to the SALIENT end-to-end implementation framework to identify the implementation stages at which these factors applied.

RESULTS

Thirty-seven articles relating to 14 groups of MLAs were identified, each trialing or implementing a bespoke algorithm. One hundred and seven distinct implementation evaluation metrics were identified. Four groups reported decreased hospital mortality, 1 significantly. We identified 24 barriers, 40 enablers, and 14 uncertainties and mapped these to the 5 stages of the SALIENT implementation framework.

DISCUSSION

Algorithm performance across implementation stages decreased between in silico and trial stages. Silent plus pilot trial inclusion was associated with decreased mortality, as was the use of logistic regression algorithms that used less than 39 variables. Mitigation of alert fatigue via alert suppression and threshold configuration was commonly employed across groups.

CONCLUSIONS

: There is evidence that real-world implementation of clinical deterioration prediction MLAs may improve clinical outcomes. Various factors identified as influencing success or failure of implementation can be mapped to different stages of implementation, thereby providing useful and practical guidance for implementers.

A framework for understanding label leakage in machine learning for health care

INTRODUCTION

The pitfalls of label leakage, contamination of model input features with outcome information, are well established. Unfortunately, avoiding label leakage in clinical prediction models requires more nuance than the common advice of applying "no time machine rule."

FRAMEWORK

We provide a framework for contemplating whether and when model features pose leakage concerns by considering the cadence, perspective, and applicability of predictions. To ground these concepts, we use real-world clinical models to highlight examples of appropriate and inappropriate label leakage in practice.

RECOMMENDATIONS

Finally, we provide recommendations to support clinical and technical stakeholders as they evaluate the leakage tradeoffs associated with model design, development, and implementation decisions. By providing common language and dimensions to consider when designing models, we hope the clinical prediction community will be better prepared to develop statistically valid and clinically useful machine learning models.

Predicting 90-day prognosis for patients with stroke: a machine learning approach

Background

Stroke is a significant global health burden and ranks as the second leading cause of death worldwide.

Objective

This study aims to develop and evaluate a machine learning-based predictive tool for forecasting the 90-day prognosis of stroke patients after discharge as measured by the modified Rankin Score.

Methods

The study utilized data from a large national multiethnic stroke registry comprising 15,859 adult patients diagnosed with ischemic or hemorrhagic stroke. Of these, 7,452 patients satisfied the study's inclusion criteria. Feature selection was performed using the correlation and permutation importance methods. Six classifiers, including Random Forest (RF), Classification and Regression Tree, Linear Discriminant Analysis, Support Vector Machine, and k-Nearest Neighbors, were employed for prediction.

Results

The RF model demonstrated superior performance, achieving the highest accuracy (0.823) and excellent discrimination power (AUC 0.893). Notably, stroke type, hospital acquired infections, admission location, and hospital length of stay emerged as the top-ranked predictors.

Conclusion

The RF model shows promise in predicting stroke prognosis, enabling personalized care plans and enhanced preventive measures for stroke patients. Prospective validation is essential to assess its real-world clinical performance and ensure successful implementation across diverse healthcare settings.

Early prediction of in-hospital mortality utilizing multivariate predictive modelling of electronic medical records and socio-determinants of health of the first day of hospitalization

BACKGROUND

In France an average of 4% of hospitalized patients die during their hospital stay. To aid medical decision making and the attribution of resources, within a few days of admission the identification of patients at high risk of dying in hospital is essential.

METHODS

We used de-identified routine patient data available in the first 2 days of hospitalization in a French University Hospital (between 2016 and 2018) to build models predicting in-hospital mortality (at ≥ 2 and ≤ 30 days after admission). We tested nine different machine learning algorithms with repeated 10-fold cross-validation. Models were trained with 283 variables including age, sex, socio-determinants of health, laboratory test results, procedures (Classification of Medical Acts), medications (Anatomical Therapeutic Chemical code), hospital department/unit and home address (urban, rural etc.). The models were evaluated using various performance metrics. The dataset contained 123,729 admissions, of which the outcome for 3542 was all-cause in-hospital mortality and 120,187 admissions (no death reported within 30 days) were controls.

RESULTS

The support vector machine, logistic regression and Xgboost algorithms demonstrated high discrimination with a balanced accuracy of 0.81 (95%CI 0.80-0.82), 0.82 (95%CI 0.80-0.83) and 0.83 (95%CI 0.80-0.83) and AUC of 0.90 (95%CI 0.88-0.91), 0.90 (95%CI 0.89-0.91) and 0.90 (95%CI 0.89-0.91) respectively. The most predictive variables for in-hospital mortality in all three models were older age (greater risk), and admission with a confirmed appointment (reduced risk).

CONCLUSION

We propose three highly discriminating machine-learning models that could improve clinical and organizational decision making for adult patients at hospital admission.

Identifying inpatient mortality in MarketScan claims data using machine learning

PURPOSE

Inpatient mortality is an important variable in epidemiology studies using claims data. In 2016, MarketScan data began obscuring specific hospital discharge status types for patient privacy, including inpatient deaths, by setting the values to missing. We used a machine learning approach to correctly identify hospitalizations that resulted in inpatient death using data prior to 2016.

METHODS

All hospitalizations from 2011 to 2015 with discharge status of missing, died, or one of the other subsequently obscured values were identified and divided into a training set and two test sets. Predictor variables included age, sex, elapsed time from hospital discharge until last observed claim and until healthcare plan disenrollment, and absence of any discharge diagnoses. Four machine learning methods were used to train statistical models and assess sensitivity and positive predictive value (PPV) for inpatient mortality.

RESULTS

Overall 1 307 917 hospitalizations were included. All four machine learning approaches performed well in all datasets. Random forest performed best with 88% PPV and 93% sensitivity for the training set and both test sets. The two factors with the highest relative importance for identifying inpatient mortality were having no observed claims for the patient on days 2-91 following hospital discharge and patient disenrollment from the healthcare plan within 60 days following hospital discharge.

CONCLUSION

We successfully developed machine learning algorithms to identify inpatient mortality. This approach can be applied to obscured data to accurately identify inpatient mortality among hospitalizations with missing discharge status.

Making the Improbable Possible: Generalizing Models Designed for a Syndrome-Based, Heterogeneous Patient Landscape

Syndromic conditions, such as sepsis, are commonly encountered in the intensive care unit. Although these conditions are easy for clinicians to grasp, these conditions may limit the performance of machine-learning algorithms. Individual hospital practice patterns may limit external generalizability. Data missingness is another barrier to optimal algorithm performance and various strategies exist to mitigate this. Recent advances in data science, such as transfer learning, conformal prediction, and continual learning, may improve generalizability of machine-learning algorithms in critically ill patients. Randomized trials with these approaches are indicated to demonstrate improvements in patient-centered outcomes at this point.

Predictive Modeling Using Artificial Intelligence and Machine Learning Algorithms on Electronic Health Record Data: Advantages and Challenges

The rapid adoption of electronic health record (EHR) systems in US hospitals from 2008 to 2014 produced novel data elements for analysis. Concurrent innovations in computing architecture and machine learning (ML) algorithms have made rapid consumption of health data feasible and a powerful engine for clinical innovation. In critical care research, the net convergence of these trends has resulted in an exponential increase in outcome prediction research. In the following article, we explore the history of outcome prediction in the intensive care unit (ICU), the growing use of EHR data, and the rise of artificial intelligence and ML (AI) in critical care.

The Use of Templates for Documenting Advance Care Planning Conversations: A Descriptive Analysis

CONTEXT

While professional societies and expert panels have recommended quality indicators related to advance care planning (ACP) documentation, including using structured documentation templates, it is unclear how clinicians document these conversations.

OBJECTIVE

To explore how clinicians document ACP, specifically, which components of these conversations are documented.

METHODS

A codebook was developed based on existing frameworks for ACP conversations and documentation. ACP documentation from a hospital medicine quality improvement project conducted from November 2019 to April 2021 were included and assessed. Documentation was examined for the presence or absence of each component within the coding schema. Clinician documented ACP using three different note types: template (only template prompts were used), template plus (authors added additional text to the template), and free text only. ACP note components were analyzed by note type and author department.

RESULTS

A total of 182 ACP notes were identified and reviewed. The most common note type was template plus (58%), followed by free text (28%) and template (14%). The most frequent components across all note types were: important relationships to patient (92%), and discussion of life-sustaining treatment preferences (87%). There was considerable heterogeneity in the components across note types. The presence of components focused on treatment decisions and legal paperwork differed significantly between note types (P < 0.05). Components on preference for medical information, emotional state, or spiritual support were rarely included across all note types.

CONCLUSION

This study provides a preliminary exploration of ACP documentation and found that templates may influence what information is documented after an ACP conversation.

Artificial Intelligence-Driven Respiratory Distress Syndrome Prediction for Very Low Birth Weight Infants: Korean Multicenter Prospective Cohort Study

BACKGROUND

Respiratory distress syndrome (RDS) is a disease that commonly affects premature infants whose lungs are not fully developed. RDS results from a lack of surfactant in the lungs. The more premature the infant is, the greater is the likelihood of having RDS. However, even though not all premature infants have RDS, preemptive treatment with artificial pulmonary surfactant is administered in most cases.

OBJECTIVE

We aimed to develop an artificial intelligence model to predict RDS in premature infants to avoid unnecessary treatment.

METHODS

In this study, 13,087 very low birth weight infants who were newborns weighing less than 1500 grams were assessed in 76 hospitals of the Korean Neonatal Network. To predict RDS in very low birth weight infants, we used basic infant information, maternity history, pregnancy/birth process, family history, resuscitation procedure, and test results at birth such as blood gas analysis and Apgar score. The prediction performances of 7 different machine learning models were compared, and a 5-layer deep neural network was proposed in order to enhance the prediction performance from the selected features. An ensemble approach combining multiple models from the 5-fold cross-validation was subsequently developed.

RESULTS

Our proposed ensemble 5-layer deep neural network consisting of the top 20 features provided high sensitivity (83.03%), specificity (87.50%), accuracy (84.07%), balanced accuracy (85.26%), and area under the curve (0.9187). Based on the model that we developed, a public web application that enables easy access for the prediction of RDS in premature infants was deployed.

CONCLUSIONS

Our artificial intelligence model may be useful for preparations for neonatal resuscitation, particularly in cases involving the delivery of very low birth weight infants, as it can aid in predicting the likelihood of RDS and inform decisions regarding the administration of surfactant.

Prospective evaluation of data-driven models to predict daily risk of Clostridioides difficile infection at 2 large academic health centers

Many data-driven patient risk stratification models have not been evaluated prospectively. We performed and compared the prospective and retrospective evaluations of 2 Clostridioides difficile infection (CDI) risk-prediction models at 2 large academic health centers, and we discuss the models' robustness to data-set shifts.

Performance of Machine Learning Algorithms for Predicting Disease Activity in Inflammatory Bowel Disease

This study aimed to explore the effectiveness of predicting disease activity in patients with inflammatory bowel disease (IBD), using machine learning (ML) models. A retrospective research was undertaken on IBD patients who were admitted into the First Affiliated Hospital of Wenzhou Medical University between September 2011 and September 2019. At first, data were randomly split into a 3:1 ratio of training to test set. The least absolute shrinkage and selection operator (LASSO) algorithm was applied to reduce the dimension of variables. These variables were used to generate seven ML algorithms, namely random forests (RFs), adaptive boosting (AdaBoost), K-nearest neighbors (KNNs), support vector machines (SVMs), naïve Bayes (NB), ridge regression, and eXtreme gradient boosting (XGBoost) to train to predict disease activity in IBD patients. SHapley Additive exPlanation (SHAP) analysis was performed to rank variable importance. A total of 876 participants with IBD, consisting of 275 ulcerative colitis (UC) and 601 Crohn's disease (CD), were retrospectively enrolled in the study. Thirty-three variables were obtained from the clinical characteristics and laboratory tests of the participants. Finally, after LASSO analysis, 11 and 5 variables were screened out to construct ML models for CD and UC, respectively. All seven ML models performed well in predicting disease activity in the CD and UC test sets. Among these ML models, SVM was more effective in predicting disease activity in the CD group, whose AUC reached 0.975, sensitivity 0.947, specificity 0.920, and accuracy 0.933. AdaBoost performed best for the UC group, with an AUC of 0.911, sensitivity 0.844, specificity 0.875, and accuracy 0.855. ML algorithms were available and capable of predicting disease activity in IBD patients. Based on clinical and laboratory variables, ML algorithms demonstrate great promise in guiding physicians' decision-making.

Artificial Intelligence Technologies in Cardiology

As the world produces exabytes of data, there is a growing need to find new methods that are more suitable for dealing with complex datasets. Artificial intelligence (AI) has significant potential to impact the healthcare industry, which is already on the road to change with the digital transformation of vast quantities of information. The implementation of AI has already achieved success in the domains of molecular chemistry and drug discoveries. The reduction in costs and in the time needed for experiments to predict the pharmacological activities of new molecules is a milestone in science. These successful applications of AI algorithms provide hope for a revolution in healthcare systems. A significant part of artificial intelligence is machine learning (ML), of which there are three main types-supervised learning, unsupervised learning, and reinforcement learning. In this review, the full scope of the AI workflow is presented, with explanations of the most-often-used ML algorithms and descriptions of performance metrics for both regression and classification. A brief introduction to explainable artificial intelligence (XAI) is provided, with examples of technologies that have developed for XAI. We review important AI implementations in cardiology for supervised, unsupervised, and reinforcement learning and natural language processing, emphasizing the used algorithm. Finally, we discuss the need to establish legal, ethical, and methodical requirements for the deployment of AI models in medicine.

Accelerating health system innovation: principles and practices from the Duke Institute for Health Innovation

The Duke Institute for Health Innovation (DIHI) was launched in 2013. Frontline staff members submit proposals for innovation projects that align with strategic priorities set by organizational leadership. Funded projects receive operational and technical support from institute staff members and a transdisciplinary network of collaborators to develop and implement solutions as part of routine clinical care, ranging from machine learning algorithms to mobile applications. DIHI's operations are shaped by four guiding principles: build to show value, build to integrate, build to scale, and build responsibly. Between 2013 and 2021, more than 600 project proposals have been submitted to DIHI. More than 85 innovation projects, both through the application process and other strategic partnerships, have been supported and implemented. DIHI's funding has incubated 12 companies, engaged more than 300 faculty members, staff members, and students, and contributed to more than 50 peer-reviewed publications. DIHI's practices can serve as a model for other health systems to systematically source, develop, implement, and scale innovations.

Advanced Care Planning for Hospitalized Patients Following Clinician Notification of Patient Mortality by a Machine Learning Algorithm

Importance

Goal-concordant care is an ongoing challenge in hospital settings. Identification of high mortality risk within 30 days may call attention to the need to have serious illness conversations, including the documentation of patient goals of care.

Objective

To examine goals of care discussions (GOCDs) in a community hospital setting with patients identified as having a high risk of mortality by a machine learning mortality prediction algorithm.

Design, Setting, and Participants

This cohort study took place at community hospitals within 1 health care system. Participants included adult patients with a high risk of 30-day mortality who were admitted to 1 of 4 hospitals between January 2 and July 15, 2021. Patient encounters of inpatients in the intervention hospital where physicians were notified of the computed high risk mortality score were compared with patient encounters of inpatients in 3 community hospitals without the intervention (ie, matched control).

Intervention

Physicians of patients with a high risk of mortality within 30 days received notification and were encouraged to arrange for GOCDs.

Main Outcomes and Measures

The primary outcome was the percentage change of documented GOCDs prior to discharge. Propensity-score matching was completed on a preintervention and postintervention period using age, sex, race, COVID-19 status, and machine learning-predicted mortality risk scores. A difference-in-difference analysis validated the results.

Results

Overall, 537 patients were included in this study with 201 in the preintervention period (94 in the intervention group; 104 in the control group) and 336 patients in the postintervention period. The intervention and control groups included 168 patients per group and were well-balanced in age (mean [SD], 79.3 [9.60] vs 79.6 [9.21] years; standardized mean difference [SMD], 0.03), sex (female, 85 [51%] vs 85 [51%]; SMD, 0), race (White patients, 145 [86%] vs 144 [86%]; SMD 0.006), and Charlson comorbidities (median [range], 8.00 [2.00-15.0] vs 9.00 [2.00 to 19.0]; SMD, 0.34). Patients in the intervention group from preintervention to postintervention period were associated with being 5 times more likely to have documented GOCDs (OR, 5.11 [95% CI, 1.93 to 13.42]; P = .001) by discharge compared with matched controls, and GOCD occurred significantly earlier in the hospitalization in the intervention patients as compared with matched controls (median, 4 [95% CI, 3 to 6] days vs 16 [95% CI, 15 to not applicable] days; P < .001). Similar findings were observed for Black patient and White patient subgroups.

Conclusions and Relevance

In this cohort study, patients whose physicians had knowledge of high-risk predictions from machine learning mortality algorithms were associated with being 5 times more likely to have documented GOCDs than matched controls. Additional external validation is needed to determine if similar interventions would be helpful at other institutions.

Quantifying and Interpreting the Prediction Accuracy of Models for the Time of a Cardiovascular Event-Moving Beyond C Statistic: A Review

Importance

For personalized or stratified medicine, it is critical to establish a reliable and efficient prediction model for a clinical outcome of interest. The goal is to develop a parsimonious model with fewer predictors for broad future application without compromising predictability. A general approach is to construct various empirical models via individual patients' specific baseline characteristics/biomarkers and then evaluate their relative merits. When the outcome of interest is the timing of a cardiovascular event, a commonly used metric to assess the adequacy of the fitted models is based on C statistics. These measures quantify a model's ability to separate those who develop events earlier from those who develop them later or not at all (discrimination), but they do not measure how closely model estimates match observed outcomes (prediction accuracy). Metrics that provide clinically interpretable measures to quantify prediction accuracy are needed.

Observations

C statistics measure the concordance between the risk scores derived from the model and the observed event time observations. However, C statistics do not quantify the model prediction accuracy. The integrated Brier Score, which calculates the mean squared distance between the empirical cumulative event-free curve and its individual patient's counterparts, estimates the prediction accuracy, but it is not clinically intuitive. A simple alternative measure is the average distance between the observed and predicted event times over the entire study population. This metric directly quantifies the model prediction accuracy and has often been used to evaluate the goodness of fit of the assumed models in settings other than survival data. This time-scale measure is easier to interpret than the C statistics or the Brier score.

Conclusions and Relevance

This article enhances our understanding of the model selection/evaluation process with respect to prediction accuracy. A simple, intuitive measure for quantifying such accuracy beyond C statistics can improve the reliability and efficiency of the selected model for personalized and stratified medicine.

Development and External Validation of a Machine Learning Model for the Early Prediction of Doses of Harmful Intracranial Pressure in Patients with Severe Traumatic Brain Injury

Treatment and prevention of elevated intracranial pressure (ICP) is crucial in patients with severe traumatic brain injury (TBI). Elevated ICP is associated with secondary brain injury, and both intensity and duration of an episode of intracranial hypertension, often referred to as "ICP dose," are associated with worse outcomes. Prediction of such harmful episodes of ICP dose could allow for a more proactive and preventive management of TBI, with potential implications on patients' outcomes. The goal of this study was to develop and validate a machine-learning (ML) model to predict potentially harmful ICP doses in patients with severe TBI. The prediction target was defined based on previous studies and included a broad range of doses of elevated ICP that have been associated with poor long-term neurological outcomes. The ML models were used, with minute-by-minute ICP and mean arterial blood pressure signals as inputs. Harmful ICP episodes were predicted with a 30 min forewarning. Models were developed in a multi-center dataset of 290 adult patients with severe TBI and externally validated on 264 patients from the Collaborative European NeuroTrauma Effectiveness Research in Traumatic Brain Injury (CENTER-TBI) dataset. The external validation of the prediction model on the CENTER-TBI dataset demonstrated good discrimination and calibration (area under the curve: 0.94, accuracy: 0.89, precision: 0.87, sensitivity: 0.78, specificity: 0.94, calibration-in-the-large: 0.03, calibration slope: 0.93). The proposed prediction model provides accurate and timely predictions of harmful doses of ICP on the development and external validation dataset. A future interventional study is needed to assess whether early intervention on the basis of ICP dose predictions will result in improved outcomes.

A hybrid machine learning framework to improve prediction of all-cause rehospitalization among elderly patients in Hong Kong

BACKGROUND

Accurately estimating elderly patients' rehospitalisation risk benefits clinical decisions and service planning. However, research in rehospitalisation and repeated hospitalisation yielded only models with modest performance, and the model performance deteriorates rapidly as the prediction timeframe expands beyond 28 days and for older participants.

METHODS

A temporal zero-inflated Poisson (tZIP) regression model was developed and validated retrospectively and prospectively. The data of the electronic health records (EHRs) contain cohorts (aged 60+) in a major public hospital in Hong Kong. Two temporal offset functions accounted for the associations between exposure time and parameters corresponding to the zero-inflated logistic component and the Poisson distribution's expected count. tZIP was externally validated with a retrospective cohort's rehospitalisation events up to 12 months after the discharge date. Subsequently, tZIP was validated prospectively after piloting its implementation at the study hospital. Patients discharged within the pilot period were tagged, and the proposed model's prediction of their rehospitalisation was verified monthly. Using a hybrid machine learning (ML) approach, the tZIP-based risk estimator's marginal effect on 28-day rehospitalisation was further validated, competing with other factors representing different post-acute and clinical statuses.

RESULTS

The tZIP prediction of rehospitalisation from 28 days to 365 days was achieved at above 80% discrimination accuracy retrospectively and prospectively in two out-of-sample cohorts. With a large margin, it outperformed the Cox proportional and linear models built with the same predictors. The hybrid ML revealed that the risk estimator's contribution to 28-day rehospitalisation outweighed other features relevant to service utilisation and clinical status.

CONCLUSIONS

A novel rehospitalisation risk model was introduced, and its risk estimators, whose importance outweighed all other factors of diverse post-acute care and clinical conditions, were derived. The proposed approach relies on four easily accessible variables easily extracted from EHR. Thus, clinicians could visualise patients' rehospitalisation risk from 28 days to 365 days after discharge and screen high-risk older patients for follow-up care at the proper time.

Prediction of intrapartum fever using continuously monitored vital signs and heart rate variability

Objectives

Neonatal early onset sepsis (EOS), bacterial infection during the first seven days of life, is difficult to diagnose because presenting signs are non-specific, but early diagnosis before birth can direct life-saving treatment for mother and baby. Specifically, maternal fever during labor from placental infection is the strongest predictor of EOS. Alterations in maternal heart rate variability (HRV) may precede development of intrapartum fever, enabling incipient EOS detection. The objective of this work was to build a predictive model for intrapartum fever.

Methods

Continuously measured temperature, heart rate, and beat-to-beat RR intervals were obtained from wireless sensors on women (n = 141) in labor; traditional manual vital signs were taken every 3-6 hours. Validated measures of HRV were calculated in moving 5-minute windows of RR intervals: standard deviation of normal-to-normal intervals (SDNN) and root mean square of successive differences (RMSSD) between normal heartbeats.

Results

Fever (>38.0 °C) was detected by manual or continuous measurements in 48 women. Compared to afebrile mothers, average SDNN and RMSSD in febrile mothers decreased significantly (p < 0.001) at 2 and 3 hours before fever onset, respectively. This observed HRV divergence and raw recorded vitals were applied to a logistic regression model at various time horizons, up to 4-5 hours before fever onset. Model performance increased with decreasing time horizons, and a model built using continuous vital signs as input variables consistently outperformed a model built from episodic vital signs.

Conclusions

HRV-based predictive models could identify mothers at risk for fever and infants at risk for EOS, guiding maternal antibiotic prophylaxis and neonatal monitoring.

Development and validation of self-monitoring auto-updating prognostic models of survival for hospitalized COVID-19 patients

Clinical prognostic models can assist patient care decisions. However, their performance can drift over time and location, necessitating model monitoring and updating. Despite rapid and significant changes during the pandemic, prognostic models for COVID-19 patients do not currently account for these drifts. We develop a framework for continuously monitoring and updating prognostic models and apply it to predict 28-day survival in COVID-19 patients. We use demographic, laboratory, and clinical data from electronic health records of 34912 hospitalized COVID-19 patients from March 2020 until May 2022 and compare three modeling methods. Model calibration performance drift is immediately detected with minor fluctuations in discrimination. The overall calibration on the prospective validation cohort is significantly improved when comparing the dynamically updated models against their static counterparts. Our findings suggest that, using this framework, models remain accurate and well-calibrated across various waves, variants, race and sex and yield positive net-benefits.

Machine learning methods to predict 30-day hospital readmission outcome among US adults with pneumonia: analysis of the national readmission database

Background

Hospital readmissions for pneumonia are a growing concern in the US, with significant consequences for costs and quality of care. This study developed the rule-based model and other machine learning (ML) models to predict 30-day readmission risk in patients with pneumonia and compared model performance.

Methods

This population-based study involved patients aged ≥ 18 years hospitalized with pneumonia from January 1, 2016, through November 30, 2016, using the Healthcare Cost and Utilization Project-National Readmission Database (HCUP-NRD). Rule-based algorithms and other ML algorithms, specifically decision trees, random forest, extreme gradient descent boosting (XGBoost), and Least Absolute Shrinkage and Selection Operator (LASSO), were used to model all-cause readmissions 30 days post-discharge from index pneumonia hospitalization. A total of 61 clinically relevant variables were included for ML model development. Models were trained on randomly partitioned 50% of the data and evaluated using the remaining dataset. Model hyperparameters were tuned using the ten-fold cross-validation on the resampled training dataset. The area under the receiver operating curves (AUROC) and area under precision-recall curves (AUPRC) were calculated for the testing set to evaluate the model performance.

Results

Of the 372,293 patients with an index hospital hospitalization for pneumonia, 48,280 (12.97%) were readmitted within 30 days. Judged by AUROC in the testing data, rule-based model (0.6591) significantly outperformed decision tree (0.5783, p value < 0.001), random forest (0.6509, p value < 0.01) and LASSO (0.6087, p value < 0.001), but was less superior than XGBoost (0.6606, p value = 0.015). The AUPRC of the rule-based model in the testing data (0.2146) was higher than the decision tree (0.1560), random forest (0.2052), and LASSO (0.2042), but was similar to XGBoost (0.2147). The top risk-predictive rules captured by the rule-based algorithm were comorbidities, illness severity, disposition locations, payer type, age, and length of stay. These predictive risk factors were also identified by other ML models with high variable importance.

Conclusion

The performance of machine learning models for predicting readmission in pneumonia patients varied. The XGboost was better than the rule-based model based on the AUROC. However, important risk factors for predicting readmission remained consistent across ML models.

An external validation study of the Score for Emergency Risk Prediction (SERP), an interpretable machine learning-based triage score for the emergency department

Emergency departments (EDs) are experiencing complex demands. An ED triage tool, the Score for Emergency Risk Prediction (SERP), was previously developed using an interpretable machine learning framework. It achieved a good performance in the Singapore population. We aimed to externally validate the SERP in a Korean cohort for all ED patients and compare its performance with Korean triage acuity scale (KTAS). This retrospective cohort study included all adult ED patients of Samsung Medical Center from 2016 to 2020. The outcomes were 30-day and in-hospital mortality after the patients' ED visit. We used the area under the receiver operating characteristic curve (AUROC) to assess the performance of the SERP and other conventional scores, including KTAS. The study population included 285,523 ED visits, of which 53,541 were after the COVID-19 outbreak (2020). The whole cohort, in-hospital, and 30 days mortality rates were 1.60%, and 3.80%. The SERP achieved an AUROC of 0.821 and 0.803, outperforming KTAS of 0.679 and 0.729 for in-hospital and 30-day mortality, respectively. SERP was superior to other scores for in-hospital and 30-day mortality prediction in an external validation cohort. SERP is a generic, intuitive, and effective triage tool to stratify general patients who present to the emergency department.

Predicting Mortality Using Machine Learning Algorithms in Patients Who Require Renal Replacement Therapy in the Critical Care Unit

Background: General severity of illness scores are not well calibrated to predict mortality among patients receiving renal replacement therapy (RRT) for acute kidney injury (AKI). We developed machine learning models to make mortality prediction and compared their performance to that of the Sequential Organ Failure Assessment (SOFA) and HEpatic failure, LactatE, NorepInephrine, medical Condition, and Creatinine (HELENICC) scores. Methods: We extracted routinely collected clinical data for AKI patients requiring RRT in the MIMIC and eICU databases. The development models were trained in 80% of the pooled dataset and tested in the rest of the pooled dataset. We compared the area under the receiver operating characteristic curves (AUCs) of four machine learning models (multilayer perceptron [MLP], logistic regression, XGBoost, and random forest [RF]) to that of the SOFA, nonrenal SOFA, and HELENICC scores and assessed calibration, sensitivity, specificity, positive (PPV) and negative (NPV) predicted values, and accuracy. Results: The mortality AUC of machine learning models was highest for XGBoost (0.823; 95% confidence interval [CI], 0.791−0.854) in the testing dataset, and it had the highest accuracy (0.758). The XGBoost model showed no evidence of lack of fit with the Hosmer−Lemeshow test (p > 0.05). Conclusion: XGBoost provided the highest performance of mortality prediction for patients with AKI requiring RRT compared with previous scoring systems.

Machine Learning for the Orthopaedic Surgeon: Uses and Limitations

➤

Machine learning is a subset of artificial intelligence in which computer algorithms are trained to make classifications and predictions based on patterns in data. The utilization of these techniques is rapidly expanding in the field of orthopaedic research.

➤

There are several domains in which machine learning has application to orthopaedics, including radiographic diagnosis, gait analysis, implant identification, and patient outcome prediction.

➤

Several limitations prevent the widespread use of machine learning in the daily clinical environment. However, future work can overcome these issues and enable machine learning tools to be a useful adjunct for orthopaedic surgeons in their clinical decision-making.

Performance of Machine Learning Algorithms for Predicting Adverse Outcomes in Community-Acquired Pneumonia

Background: The ability to assess adverse outcomes in patients with community-acquired pneumonia (CAP) could improve clinical decision-making to enhance clinical practice, but the studies remain insufficient, and similarly, few machine learning (ML) models have been developed.

Objective: We aimed to explore the effectiveness of predicting adverse outcomes in CAP through ML models.

Methods: A total of 2,302 adults with CAP who were prospectively recruited between January 2012 and March 2015 across three cities in South America were extracted from DryadData. After a 70:30 training set: test set split of the data, nine ML algorithms were executed and their diagnostic accuracy was measured mainly by the area under the curve (AUC). The nine ML algorithms included decision trees, random forests, extreme gradient boosting (XGBoost), support vector machines, Naïve Bayes, K-nearest neighbors, ridge regression, logistic regression without regularization, and neural networks. The adverse outcomes included hospital admission, mortality, ICU admission, and one-year post-enrollment status.

Results: The XGBoost algorithm had the best performance in predicting hospital admission. Its AUC reached 0.921, and accuracy, precision, recall, and F1-score were better than those of other models. In the prediction of ICU admission, a model trained with the XGBoost algorithm showed the best performance with AUC 0.801. XGBoost algorithm also did a good job at predicting one-year post-enrollment status. The results of AUC, accuracy, precision, recall, and F1-score indicated the algorithm had high accuracy and precision. In addition, the best performance was seen by the neural network algorithm when predicting death (AUC 0.831).

Conclusions: ML algorithms, particularly the XGBoost algorithm, were feasible and effective in predicting adverse outcomes of CAP patients. The ML models based on available common clinical features had great potential to guide individual treatment and subsequent clinical decisions.

Machine learning-based suggestion for critical interventions in the management of potentially severe conditioned patients in emergency department triage

Providing timely intervention to critically ill patients is a challenging task in emergency departments (ED). Our study aimed to predict early critical interventions (CrIs), which can be used as clinical recommendations. This retrospective observational study was conducted in the ED of a tertiary hospital located in a Korean metropolitan city. Patient who visited ED from January 1, 2016, to December 31, 2018, were included. Need of six CrIs were selected as prediction outcomes, namely, arterial line (A-line) insertion, oxygen therapy, high-flow nasal cannula (HFNC), intubation, Massive Transfusion Protocol (MTP), and inotropes and vasopressor. Extreme gradient boosting (XGBoost) prediction model was built by using only data available at the initial stage of ED. Overall, 137,883 patients were included in the study. The areas under the receiver operating characteristic curve for the prediction of A-line insertion was 0·913, oxygen therapy was 0.909, HFNC was 0.962, intubation was 0.945, MTP was 0.920, and inotropes or vasopressor administration was 0.899 in the XGBoost method. In addition, an increase in the need for CrIs was associated with worse ED outcomes. The CrIs model was integrated into the study site's electronic medical record and could be used to suggest early interventions for emergency physicians.

Palliative Care Exposure Relative to Predicted Risk of Six-Month Mortality in Hospitalized Adults

CONTEXT

The optimal strategy for implementing mortality-predicting algorithms to facilitate clinical care, prognostic discussions, and palliative care interventions remains unknown.

OBJECTIVES

To develop and validate a real-time predictive model for 180 day mortality using routinely available clinical and laboratory admission data and determine if palliative care exposure varies with predicted mortality risk.

METHODS

Adult admissions between October 1, 2013 and October.1, 2017 were included for the model derivation. A separate cohort was collected between January 1, 2018 and July 31, 2020 for validation. Patients were followed for 180 days from discharge, and logistic regression with selected variables was used to estimate patients' risk for mortality.

RESULTS

In the model derivation cohort, 7963 events of 180 day mortality (4.5% event rate) were observed. Median age was 53.0 (IQR 24.0-66.0) with 92,734 females (52.5%). Variables with strongest association with 180 day mortality included: Braden Score (OR 0.83; 95% CI 0.82-0.84); admission Do Not Resuscitate orders (OR 2.61; 95% CI 2.43-2.79); admission service and admission status. The model yielded excellent discriminatory ability in both the derivation (c-statistic 0.873; 95% CI 0.870-0.877; Brier score 0.04) and validation cohorts (c-statistic 0.844; 95% CI 0.840-0.847; Brier score 0.072). Inpatient palliative care consultations increased from 3% of minimal-risk encounters to 41% of high-risk encounters (P < 0.01).

CONCLUSION

We developed and temporally validated a predictive mortality model for adults from a large retrospective cohort, which helps quantify the potential need for palliative care referrals based on risk strata. Machine learning algorithms for mortality require clinical interpretation, and additional studies are needed to design patient-centered and risk-specific interventions.

Development and Structure of an Accurate Machine Learning Algorithm to Predict Inpatient Mortality and Hospice Outcomes in the Coronavirus Disease 2019 Era

BACKGROUND

The coronavirus disease 2019 (COVID-19) pandemic has challenged the accuracy and racial biases present in traditional mortality scores. An accurate prognostic model that can be applied to hospitalized patients irrespective of race or COVID-19 status may benefit patient care.

RESEARCH DESIGN

This cohort study utilized historical and ongoing electronic health record features to develop and validate a deep-learning model applied on the second day of admission predicting a composite outcome of in-hospital mortality, discharge to hospice, or death within 30 days of admission. Model features included patient demographics, diagnoses, procedures, inpatient medications, laboratory values, vital signs, and substance use history. Conventional performance metrics were assessed, and subgroup analysis was performed based on race, COVID-19 status, and intensive care unit admission.

SUBJECTS

A total of 35,521 patients hospitalized between April 2020 and October 2020 at a single health care system including a tertiary academic referral center and 9 community hospitals.

RESULTS

Of 35,521 patients, including 9831 non-White patients and 2020 COVID-19 patients, 2838 (8.0%) met the composite outcome. Patients who experienced the composite outcome were older (73 vs. 61 y old) with similar sex and race distributions between groups. The model achieved an area under the receiver operating characteristic curve of 0.89 (95% confidence interval: 0.88, 0.91) and an average positive predictive value of 0.46 (0.40, 0.52). Model performance did not differ significantly in White (0.89) and non-White (0.90) subgroups or when grouping by COVID-19 status and intensive care unit admission.

CONCLUSION

A deep-learning model using large-volume, structured electronic health record data can effectively predict short-term mortality or hospice outcomes on the second day of admission in the general inpatient population without significant racial bias.

Dynamic Mortality Risk Predictions for Children in ICUs: Development and Validation of Machine Learning Models

OBJECTIVES

Assess a machine learning method of serially updated mortality risk.

DESIGN

Retrospective analysis of a national database (Health Facts; Cerner Corporation, Kansas City, MO).

SETTING

Hospitals caring for children in ICUs.

PATIENTS

A total of 27,354 admissions cared for in ICUs from 2009 to 2018.

INTERVENTIONS

None.

MAIN OUTCOME

Hospital mortality risk estimates determined at 6-hour time periods during care in the ICU. Models were truncated at 180 hours due to decreased sample size secondary to discharges and deaths.

MEASUREMENTS AND MAIN RESULTS

The Criticality Index, based on physiology, therapy, and care intensity, was computed for each admission for each time period and calibrated to hospital mortality risk (Criticality Index-Mortality [CI-M]) at each of 29 time periods (initial assessment: 6 hr; last assessment: 180 hr). Performance metrics and clinical validity were determined from the held-out test sample (n = 3,453, 13%). Discrimination assessed with the area under the receiver operating characteristic curve was 0.852 (95% CI, 0.843-0.861) overall and greater than or equal to 0.80 for all individual time periods. Calibration assessed by the Hosmer-Lemeshow goodness-of-fit test showed good fit overall (p = 0.196) and was statistically not significant for 28 of the 29 time periods. Calibration plots for all models revealed the intercept ranged from--0.002 to 0.009, the slope ranged from 0.867 to 1.415, and the R2 ranged from 0.862 to 0.989. Clinical validity assessed using population trajectories and changes in the risk status of admissions (clinical volatility) revealed clinical trajectories consistent with clinical expectations and greater clinical volatility in deaths than survivors (p < 0.001).

CONCLUSIONS

Machine learning models incorporating physiology, therapy, and care intensity can track changes in hospital mortality risk during intensive care. The CI-M's framework and modeling method are potentially applicable to monitoring clinical improvement and deterioration in real time.

External validation of the 4C Mortality Score for hospitalised patients with COVID-19 in the RECOVER network

OBJECTIVES

Estimating mortality risk in hospitalised SARS-CoV-2+ patients may help with choosing level of care and discussions with patients. The Coronavirus Clinical Characterisation Consortium Mortality Score (4C Score) is a promising COVID-19 mortality risk model. We examined the association of risk factors with 30-day mortality in hospitalised, full-code SARS-CoV-2+ patients and investigated the discrimination and calibration of the 4C Score. This was a retrospective cohort study of SARS-CoV-2+ hospitalised patients within the RECOVER (REgistry of suspected COVID-19 in EmeRgency care) network.

SETTING

99 emergency departments (EDs) across the USA.

PARTICIPANTS

Patients ≥18 years old, positive for SARS-CoV-2 in the ED, and hospitalised.

PRIMARY OUTCOME

Death within 30 days of the index visit. We performed logistic regression analysis, reporting multivariable risk ratios (MVRRs) and calculated the area under the ROC curve (AUROC) and mean prediction error for the original 4C Score and after dropping the C reactive protein (CRP) component.

RESULTS

Of 6802 hospitalised patients with COVID-19, 1149 (16.9%) died within 30 days. The 30-day mortality was increased with age 80+ years (MVRR=5.79, 95% CI 4.23 to 7.34); male sex (MVRR=1.17, 1.05 to 1.28); and nursing home/assisted living facility residence (MVRR=1.29, 1.1 to 1.48). The 4C Score had comparable discrimination in the RECOVER dataset compared with the original 4C validation dataset (AUROC: RECOVER 0.786 (95% CI 0.773 to 0.799), 4C validation 0.763 (95% CI 0.757 to 0.769). Score-specific mortalities in our sample were lower than in the 4C validation sample (mean prediction error 6.0%). Dropping the CRP component from the 4C Score did not substantially affect discrimination and 4C risk estimates were now close (mean prediction error 0.7%).

CONCLUSIONS

We independently validated 4C Score as predicting risk of 30-day mortality in hospitalised SARS-CoV-2+ patients. We recommend dropping the CRP component of the score and using our recalibrated mortality risk estimates.

Generalizability challenges of mortality risk prediction models: A retrospective analysis on a multi-center database

Modern predictive models require large amounts of data for training and evaluation, absence of which may result in models that are specific to certain locations, populations in them and clinical practices. Yet, best practices for clinical risk prediction models have not yet considered such challenges to generalizability. Here we ask whether population- and group-level performance of mortality prediction models vary significantly when applied to hospitals or geographies different from the ones in which they are developed. Further, what characteristics of the datasets explain the performance variation? In this multi-center cross-sectional study, we analyzed electronic health records from 179 hospitals across the US with 70,126 hospitalizations from 2014 to 2015. Generalization gap, defined as difference between model performance metrics across hospitals, is computed for area under the receiver operating characteristic curve (AUC) and calibration slope. To assess model performance by the race variable, we report differences in false negative rates across groups. Data were also analyzed using a causal discovery algorithm "Fast Causal Inference" that infers paths of causal influence while identifying potential influences associated with unmeasured variables. When transferring models across hospitals, AUC at the test hospital ranged from 0.777 to 0.832 (1st-3rd quartile or IQR; median 0.801); calibration slope from 0.725 to 0.983 (IQR; median 0.853); and disparity in false negative rates from 0.046 to 0.168 (IQR; median 0.092). Distribution of all variable types (demography, vitals, and labs) differed significantly across hospitals and regions. The race variable also mediated differences in the relationship between clinical variables and mortality, by hospital/region. In conclusion, group-level performance should be assessed during generalizability checks to identify potential harms to the groups. Moreover, for developing methods to improve model performance in new environments, a better understanding and documentation of provenance of data and health processes are needed to identify and mitigate sources of variation.

Current Status and Future Directions of Neuromonitoring With Emerging Technologies in Neonatal Care

Neonatology has experienced a significant reduction in mortality rates of the preterm population and critically ill infants over the last few decades. Now, the emphasis is directed toward improving long-term neurodevelopmental outcomes and quality of life. Brain-focused care has emerged as a necessity. The creation of neonatal neurocritical care units, or Neuro-NICUs, provides strategies to reduce brain injury using standardized clinical protocols, methodologies, and provider education and training. Bedside neuromonitoring has dramatically improved our ability to provide assessment of newborns at high risk. Non-invasive tools, such as continuous electroencephalography (cEEG), amplitude-integrated electroencephalography (aEEG), and near-infrared spectroscopy (NIRS), allow screening for seizures and continuous evaluation of brain function and cerebral oxygenation at the bedside. Extended and combined uses of these techniques, also described as multimodal monitoring, may allow practitioners to better understand the physiology of critically ill neonates. Furthermore, the rapid growth of technology in the Neuro-NICU, along with the increasing use of telemedicine and artificial intelligence with improved data mining techniques and machine learning (ML), has the potential to vastly improve decision-making processes and positively impact outcomes. This article will cover the current applications of neuromonitoring in the Neuro-NICU, recent advances, potential pitfalls, and future perspectives in this field.

The derivation of an International Classification of Diseases, Tenth Revision-based trauma-related mortality model using machine learning

BACKGROUND

Existing mortality prediction models have attempted to quantify injury burden following trauma-related admissions with the most notable being the Injury Severity Score (ISS). Although easy to calculate, it requires additional administrative coding. International Classification of Diseases (ICD)-based models such as the Trauma Mortality Prediction Model (TMPM-ICD10) circumvent these limitations, but they use linear modeling, which may not adequately capture the intricate relationships of injuries on mortality. Using ICD-10 coding and machine learning (ML) algorithms, the present study used the National Trauma Data Bank to develop mortality prediction models whose performance was compared with logistic regression, ISS, and TMPM-ICD10.

METHODS

The 2015 to 2017 National Trauma Data Bank was used to identify adults following trauma-related admissions. Of 8,021 ICD-10 codes, injuries were categorized into 1,495 unique variables. The primary outcome was in-hospital mortality. eXtreme Gradient Boosting (XGBoost), a ML technique that uses iterations of decision trees, was used to develop mortality models. Model discrimination was compared with logistic regression, ISS, and TMPM-ICD10 using receiver operating characteristic curve and probabilistic accuracy with calibration curves.

RESULTS

Of 1,611,063 patients, 54,870 (3.41%) experienced in-hospital mortality. Compared with those who survived, those who died more frequently suffered from penetrating trauma and had a greater number of injuries. The XGBoost model exhibited superior receiver operating characteristic curve (0.863 [95% confidence interval (CI), 0.862-0.864]) compared with logistic regression (0.845 [95% CI, 0.844-0.846]), ISS (0.828 [95% CI, 0.827-0.829]), and TMPM-ICD10 (0.861 [95% CI, 0.860-0.862]) (all p < 0.001). Importantly, the ML model also had significantly improved calibration compared with other methodologies (XGBoost, coefficient of determination (R2) = 0.993; logistic regression, R2 = 0.981; ISS, R2 = 0.649; TMPM-ICD10, R2 = 0.830).

CONCLUSION

Machine learning models using XGBoost demonstrated superior performance and calibration compared with logistic regression, ISS, and TMPM-ICD10. Such approaches in quantifying injury severity may improve its utility in mortality prognostication, quality improvement, and trauma research.

LEVEL OF EVIDENCE

Prognostic and Epidemiologic; level III.

Data Science Trends Relevant to Nursing Practice: A Rapid Review of the 2020 Literature

BACKGROUND

The term "data science" encompasses several methods, many of which are considered cutting edge and are being used to influence care processes across the world. Nursing is an applied science and a key discipline in health care systems in both clinical and administrative areas, making the profession increasingly influenced by the latest advances in data science. The greater informatics community should be aware of current trends regarding the intersection of nursing and data science, as developments in nursing practice have cross-professional implications.

OBJECTIVES

This study aimed to summarize the latest (calendar year 2020) research and applications of nursing-relevant patient outcomes and clinical processes in the data science literature.

METHODS

We conducted a rapid review of the literature to identify relevant research published during the year 2020. We explored the following 16 topics: (1) artificial intelligence/machine learning credibility and acceptance, (2) burnout, (3) complex care (outpatient), (4) emergency department visits, (5) falls, (6) health care-acquired infections, (7) health care utilization and costs, (8) hospitalization, (9) in-hospital mortality, (10) length of stay, (11) pain, (12) patient safety, (13) pressure injuries, (14) readmissions, (15) staffing, and (16) unit culture.

RESULTS

Of 16,589 articles, 244 were included in the review. All topics were represented by literature published in 2020, ranging from 1 article to 59 articles. Numerous contemporary data science methods were represented in the literature including the use of machine learning, neural networks, and natural language processing.

CONCLUSION

This review provides an overview of the data science trends that were relevant to nursing practice in 2020. Examinations of such literature are important to monitor the status of data science's influence in nursing practice.

Development and validation of a machine learning method to predict intraoperative red blood cell transfusions in cardiothoracic surgery

Accurately predicting red blood cell (RBC) transfusion requirements in cardiothoracic (CT) surgery could improve blood inventory management and be used as a surrogate marker for assessing hemorrhage risk preoperatively. We developed a machine learning (ML) method to predict intraoperative RBC transfusions in CT surgery. A detailed database containing time-stamped clinical variables for all CT surgeries from 5/2014-6/2019 at a single center (n = 2410) was used for model development. After random forest feature selection, surviving features were inputs for ML algorithms using five-fold cross-validation. The dataset was updated with 437 additional cases from 8/2019-8/2020 for validation. We developed and validated a hybrid ML method given the skewed nature of the dataset. Our Gaussian Process (GP) regression ML algorithm accurately predicted RBC transfusion amounts of 0 and 1-3 units (root mean square error, RMSE 0.117 and 1.705, respectively) and our GP classification ML algorithm accurately predicted 4 + RBC units transfused (area under the curve, AUC = 0.826). The final prediction is the regression result if classification predicted < 4 units transfused, or the classification result if 4 + units were predicted. We developed and validated an ML method to accurately predict intraoperative RBC transfusions in CT surgery using local data.

A Knowledge Distillation Ensemble Framework for Predicting Short- and Long-Term Hospitalization Outcomes From Electronic Health Records Data

The ability to perform accurate prognosis is crucial for proactive clinical decision making, informed resource management and personalised care. Existing outcome prediction models suffer from a low recall of infrequent positive outcomes. We present a highly-scalable and robust machine learning framework to automatically predict adversity represented by mortality and ICU admission and readmission from time-series of vital signs and laboratory results obtained within the first 24 hours of hospital admission. The stacked ensemble platform comprises two components: a) an unsupervised LSTM Autoencoder that learns an optimal representation of the time-series, using it to differentiate the less frequent patterns which conclude with an adverse event from the majority patterns that do not, and b) a gradient boosting model, which relies on the constructed representation to refine prediction by incorporating static features. The model is used to assess a patient's risk of adversity and provides visual justifications of its prediction. Results of three case studies show that the model outperforms existing platforms in ICU and general ward settings, achieving average Precision-Recall Areas Under the Curve (PR-AUCs) of 0.891 (95% CI: 0.878-0.939) for mortality and 0.908 (95% CI: 0.870-0.935) in predicting ICU admission and readmission.

The criticality Index-mortality: A dynamic machine learning prediction algorithm for mortality prediction in children cared for in an ICU

BACKGROUND

The Criticality Index-Mortality uses physiology, therapy, and intensity of care to compute mortality risk for pediatric ICU patients. If the frequency of mortality risk computations were increased to every 3 h with model performance that could improve the assessment of severity of illness, it could be utilized to monitor patients for significant mortality risk change.

OBJECTIVES

To assess the performance of a dynamic method of updating mortality risk every 3 h using the Criticality Index-Mortality methodology and identify variables that are significant contributors to mortality risk predictions.

POPULATION

There were 8,399 pediatric ICU admissions with 312 (3.7%) deaths from January 1, 2018 to February 29, 2020. We randomly selected 75% of patients for training, 13% for validation, and 12% for testing.

MODEL

A neural network was trained to predict hospital survival or death during or following an ICU admission. Variables included age, gender, laboratory tests, vital signs, medications categories, and mechanical ventilation variables. The neural network was calibrated to mortality risk using nonparametric logistic regression.

RESULTS

Discrimination assessed across all time periods found an AUROC of 0.851 (0.841-0.862) and an AUPRC was 0.443 (0.417-0.467). When assessed for performance every 3 h, the AUROCs had a minimum value of 0.778 (0.689-0.867) and a maximum value of 0.885 (0.841,0.862); the AUPRCs had a minimum value 0.148 (0.058-0.328) and a maximum value of 0.499 (0.229-0.769). The calibration plot had an intercept of 0.011, a slope of 0.956, and the R² was 0.814. Comparison of observed vs. expected proportion of deaths revealed that 95.8% of the 543 risk intervals were not statistically significantly different. Construct validity assessed by death and survivor risk trajectories analyzed by mortality risk quartiles and 7 high and low risk diseases confirmed a priori clinical expectations about the trajectories of death and survivors.

CONCLUSIONS

The Criticality Index-Mortality computing mortality risk every 3 h for pediatric ICU patients has model performance that could enhance the clinical assessment of severity of illness. The overall Criticality Index-Mortality framework was effectively applied to develop an institutionally specific, and clinically relevant model for dynamic risk assessment of pediatric ICU patients.

Prediction across healthcare settings: a case study in predicting emergency department disposition

Several approaches exist today for developing predictive models across multiple clinical sites, yet there is a lack of comparative data on their performance, especially within the context of EHR-based prediction models. We set out to provide a framework for prediction across healthcare settings. As a case study, we examined an ED disposition prediction model across three geographically and demographically diverse sites. We conducted a 1-year retrospective study, including all visits in which the outcome was either discharge-to-home or hospitalization. Four modeling approaches were compared: a ready-made model trained at one site and validated at other sites, a centralized uniform model incorporating data from all sites, multiple site-specific models, and a hybrid approach of a ready-made model re-calibrated using site-specific data. Predictions were performed using XGBoost. The study included 288,962 visits with an overall admission rate of 16.8% (7.9–26.9%). Some risk factors for admission were prominent across all sites (e.g., high-acuity triage emergency severity index score, high prior admissions rate), while others were prominent at only some sites (multiple lab tests ordered at the pediatric sites, early use of ECG at the adult site). The XGBoost model achieved its best performance using the uniform and site-specific approaches (AUC = 0.9–0.93), followed by the calibrated-model approach (AUC = 0.87–0.92), and the ready-made approach (AUC = 0.62–0.85). Our results show that site-specific customization is a key driver of predictive model performance.

Machine learning techniques for mortality prediction in emergency departments: a systematic review

OBJECTIVES

This systematic review aimed to assess the performance and clinical feasibility of machine learning (ML) algorithms in prediction of in-hospital mortality for medical patients using vital signs at emergency departments (EDs).

DESIGN

A systematic review was performed.

SETTING

The databases including Medline (PubMed), Scopus and Embase (Ovid) were searched between 2010 and 2021, to extract published articles in English, describing ML-based models utilising vital sign variables to predict in-hospital mortality for patients admitted at EDs. Critical appraisal and data extraction for systematic reviews of prediction modelling studies checklist was used for study planning and data extraction. The risk of bias for included papers was assessed using the prediction risk of bias assessment tool.

PARTICIPANTS

Admitted patients to the ED.

MAIN OUTCOME MEASURE

In-hospital mortality.

RESULTS

Fifteen articles were included in the final review. We found that eight models including logistic regression, decision tree, K-nearest neighbours, support vector machine, gradient boosting, random forest, artificial neural networks and deep neural networks have been applied in this domain. Most studies failed to report essential main analysis steps such as data preprocessing and handling missing values. Fourteen included studies had a high risk of bias in the statistical analysis part, which could lead to poor performance in practice. Although the main aim of all studies was developing a predictive model for mortality, nine articles did not provide a time horizon for the prediction.

CONCLUSION

This review provided an updated overview of the state-of-the-art and revealed research gaps; based on these, we provide eight recommendations for future studies to make the use of ML more feasible in practice. By following these recommendations, we expect to see more robust ML models applied in the future to help clinicians identify patient deterioration earlier.

Key Contributions in Clinical Research Informatics

Objectives: To summarize key contributions to current research in the field of Clinical Research Informatics (CRI) and to select best papers published in 2020.

Method: A bibliographic search using a combination of Medical Subject Headings (MeSH) descriptors and free-text terms on CRI was performed using PubMed, followed by a double-blind review in order to select a list of candidate best papers to be then peer-reviewed by external reviewers. After peer-review ranking, a consensus meeting between two section editors and the editorial team was organized to finally conclude on the selected four best papers.

Results: Among the 877 papers published in 2020 and returned by the search, there were four best papers selected. The first best paper describes a method for mining temporal sequences from clinical documents to infer disease trajectories and enhancing high-throughput phenotyping. The authors of the second best paper demonstrate that the generation of synthetic Electronic Health Record (EHR) data through Generative Adversarial Networks (GANs) could be substantially improved by more appropriate training and evaluation criteria. The third best paper offers an efficient advance on methods to detect adverse drug events by computer-assisting expert reviewers with annotated candidate mentions in clinical documents. The large-scale data quality assessment study reported by the fourth best paper has clinical research informatics implications, in terms of the trustworthiness of inferences made from analysing electronic health records.

Conclusions: The most significant research efforts in the CRI field are currently focusing on data science with active research in the development and evaluation of Artificial Intelligence/Machine Learning (AI/ML) algorithms based on ever more intensive use of real-world data and especially EHR real or synthetic data. A major lesson that the coronavirus disease 2019 (COVID-19) pandemic has already taught the scientific CRI community is that timely international high-quality data-sharing and collaborative data analysis is absolutely vital to inform policy decisions.

Improving the In-Hospital Mortality Prediction of Diabetes ICU Patients Using a Process Mining/Deep Learning Architecture

Diabetes intensive care unit (ICU) patients are at increased risk of complications leading to in-hospital mortality. Assessing the likelihood of death is a challenging and time consuming task due to a large number of influencing factors. Healthcare providers are interested in the detection of ICU patients at higher risk, such that risk factors can possibly be mitigated. While such severity scoring methods exist, they are commonly based on a snapshot of the health conditions of a patient during the ICU stay and do not specifically consider a patient's prior medical history. In this paper, a process mining/deep learning architecture is proposed to improve established severity scoring methods by incorporating the medical history of diabetes patients. First, health records of past hospital encounters are converted to event logs suitable for process mining. The event logs are then used to discover a process model that describes the past hospital encounters of patients. An adaptation of Decay Replay Mining is proposed to combine medical and demographic information with established severity scores to predict the in hospital mortality of diabetes ICU patients. Significant performance improvements are demonstrated compared to established risk severity scoring methods and machine learning approaches using the Medical Information Mart for Intensive Care III dataset.