Machine Learning Versus Usual Care for Diagnostic and Prognostic Prediction in the Emergency Department: A Systematic Review

Evaluating accuracy and fairness of clinical decision support algorithms when health care resources are limited

OBJECTIVE

Guidance on how to evaluate accuracy and algorithmic fairness across subgroups is missing for clinical models that flag patients for an intervention but when health care resources to administer that intervention are limited. We aimed to propose a framework of metrics that would fit this specific use case.

METHODS

We evaluated the following metrics and applied them to a Veterans Health Administration clinical model that flags patients for intervention who are at risk of overdose or a suicidal event among outpatients who were prescribed opioids (N = 405,817): Receiver - Operating Characteristic and area under the curve, precision - recall curve, calibration - reliability curve, false positive rate, false negative rate, and false omission rate. In addition, we developed a new approach to visualize false positives and false negatives that we named 'per true positive bars.' We demonstrate the utility of these metrics to our use case for three cohorts of patients at the highest risk (top 0.5 %, 1.0 %, and 5.0 %) by evaluating algorithmic fairness across the following age groups: <=30, 31-50, 51-65, and >65 years old.

RESULTS

Metrics that allowed us to assess group differences more clearly were the false positive rate, false negative rate, false omission rate, and the new 'per true positive bars'. Metrics with limited utility to our use case were the Receiver - Operating Characteristic and area under the curve, the calibration - reliability curve, and the precision - recall curve.

CONCLUSION

There is no "one size fits all" approach to model performance monitoring and bias analysis. Our work informs future researchers and clinicians who seek to evaluate accuracy and fairness of predictive models that identify patients to intervene on in the context of limited health care resources. In terms of ease of interpretation and utility for our use case, the new 'per true positive bars' may be the most intuitive to a range of stakeholders and facilitates choosing a threshold that allows weighing false positives against false negatives, which is especially important when predicting severe adverse events.

Precision emergency medicine

BACKGROUND

Precision health is a burgeoning scientific discipline that aims to incorporate individual variability in biological, behavioral, and social factors to develop personalized health solutions. To date, emergency medicine has not deeply engaged in the precision health movement. However, rapid advances in health technology, data science, and medical informatics offer new opportunities for emergency medicine to realize the promises of precision health.

METHODS

In this article, we conceptualize precision emergency medicine as an emerging paradigm and identify key drivers of its implementation into current and future clinical practice. We acknowledge important obstacles to the specialty-wide adoption of precision emergency medicine and offer solutions that conceive a successful path forward.

RESULTS

Precision emergency medicine is defined as the use of information and technology to deliver acute care effectively, efficiently, and authentically to individual patients and their communities. Key drivers and opportunities include leveraging human data, capitalizing on technology and digital tools, providing deliberate access to care, advancing population health, and reimagining provider education and roles. Overcoming challenges in equity, privacy, and cost is essential for success. We close with a call to action to proactively incorporate precision health into the clinical practice of emergency medicine, the training of future emergency physicians, and the research agenda of the specialty.

CONCLUSIONS

Precision emergency medicine leverages new technology and data-driven artificial intelligence to advance diagnostic testing, individualize patient care plans and therapeutics, and strategically refine the convergence of the health system and the community.

Role of Artificial Intelligence in Improving Syncope Management

Syncope is common in the general population and a common presenting symptom in acute care settings. Substantial costs are attributed to the care of patients with syncope. Current challenges include differentiating syncope from its mimickers, identifying serious underlying conditions that caused the syncope, and wide variations in current management. Although validated risk tools exist, especially for short-term prognosis, there is inconsistent application, and the current approach does not meet patient needs and expectations. Artificial intelligence (AI) techniques, such as machine learning methods including natural language processing, can potentially address the current challenges in syncope management. Preliminary evidence from published studies indicates that it is possible to accurately differentiate syncope from its mimickers and predict short-term prognosis and hospitalisation. More recently, AI analysis of electrocardiograms has shown promise in detection of serious structural and functional cardiac abnormalities, which has the potential to improve syncope care. Future AI studies have the potential to address current issues in syncope management. AI can automatically prognosticate risk in real time by accessing traditional and nontraditional data. However, steps to mitigate known problems such as generalisability, patient privacy, data protection, and liability will be needed. In the past AI has had limited impact due to underdeveloped analytical methods, lack of computing power, poor access to powerful computing systems, and availability of reliable high-quality data. All impediments except data have been solved. AI will live up to its promise to transform syncope care if the health care system can satisfy AI requirement of large scale, robust, accurate, and reliable data.

Cherry on Top or Real Need? A Review of Explainable Machine Learning in Kidney Transplantation

Research on solid organ transplantation has taken advantage of the substantial acquisition of medical data and the use of artificial intelligence (AI) and machine learning (ML) to answer diagnostic, prognostic, and therapeutic questions for many years. Nevertheless, despite the question of whether AI models add value to traditional modeling approaches, such as regression models, their "black box" nature is one of the factors that have hindered the translation from research to clinical practice. Several techniques that make such models understandable to humans were developed with the promise of increasing transparency in the support of medical decision-making. These techniques should help AI to close the gap between theory and practice by yielding trust in the model by doctors and patients, allowing model auditing, and facilitating compliance with emergent AI regulations. But is this also happening in the field of kidney transplantation? This review reports the use and explanation of "black box" models to diagnose and predict kidney allograft rejection, delayed graft function, graft failure, and other related outcomes after kidney transplantation. In particular, we emphasize the discussion on the need (or not) to explain ML models for biological discovery and clinical implementation in kidney transplantation. We also discuss promising future research paths for these computational tools.

Applications and Performance of Machine Learning Algorithms in Emergency Medical Services: A Scoping Review

OBJECTIVE

The aim of this study was to summarize the literature on the applications of machine learning (ML) and their performance in Emergency Medical Services (EMS).

METHODS

Four relevant electronic databases were searched (from inception through January 2024) for all original studies that employed EMS-guided ML algorithms to enhance the clinical and operational performance of EMS. Two reviewers screened the retrieved studies and extracted relevant data from the included studies. The characteristics of included studies, employed ML algorithms, and their performance were quantitively described across primary domains and subdomains.

RESULTS

This review included a total of 164 studies published from 2005 through 2024. Of those, 125 were clinical domain focused and 39 were operational. The characteristics of ML algorithms such as sample size, number and type of input features, and performance varied between and within domains and subdomains of applications. Clinical applications of ML algorithms involved triage or diagnosis classification (n = 62), treatment prediction (n = 12), or clinical outcome prediction (n = 50), mainly for out-of-hospital cardiac arrest/OHCA (n = 62), cardiovascular diseases/CVDs (n = 19), and trauma (n = 24). The performance of these ML algorithms varied, with a median area under the receiver operating characteristic curve (AUC) of 85.6%, accuracy of 88.1%, sensitivity of 86.05%, and specificity of 86.5%. Within the operational studies, the operational task of most ML algorithms was ambulance allocation (n = 21), followed by ambulance detection (n = 5), ambulance deployment (n = 5), route optimization (n = 5), and quality assurance (n = 3). The performance of all operational ML algorithms varied and had a median AUC of 96.1%, accuracy of 90.0%, sensitivity of 94.4%, and specificity of 87.7%. Generally, neural network and ensemble algorithms, to some degree, out-performed other ML algorithms.

CONCLUSION

Triaging and managing different prehospital medical conditions and augmenting ambulance performance can be improved by ML algorithms. Future reports should focus on a specific clinical condition or operational task to improve the precision of the performance metrics of ML models.

Systematic reviews of machine learning in healthcare: a literature review

INTRODUCTION

The increasing availability of data and computing power has made machine learning (ML) a viable approach to faster, more efficient healthcare delivery.

METHODS

A systematic literature review (SLR) of published SLRs evaluating ML applications in healthcare settings published between1 January 2010 and 27 March 2023 was conducted.

RESULTS

In total 220 SLRs covering 10,462 ML algorithms were reviewed. The main application of AI in medicine related to the clinical prediction and disease prognosis in oncology and neurology with the use of imaging data. Accuracy, specificity, and sensitivity were provided in 56%, 28%, and 25% SLRs respectively. Internal and external validation was reported in 53% and less than 1% of the cases respectively. The most common modeling approach was neural networks (2,454 ML algorithms), followed by support vector machine and random forest/decision trees (1,578 and 1,522 ML algorithms, respectively).

EXPERT OPINION

The review indicated considerable reporting gaps in terms of the ML's performance, both internal and external validation. Greater accessibility to healthcare data for developers can ensure the faster adoption of ML algorithms into clinical practice.

Applications of natural language processing at emergency department triage: A narrative review

INTRODUCTION

Natural language processing (NLP) uses various computational methods to analyse and understand human language, and has been applied to data acquired at Emergency Department (ED) triage to predict various outcomes. The objective of this scoping review is to evaluate how NLP has been applied to data acquired at ED triage, assess if NLP based models outperform humans or current risk stratification techniques when predicting outcomes, and assess if incorporating free-text improve predictive performance of models when compared to predictive models that use only structured data.

METHODS

All English language peer-reviewed research that applied an NLP technique to free-text obtained at ED triage was eligible for inclusion. We excluded studies focusing solely on disease surveillance, and studies that used information obtained after triage. We searched the electronic databases MEDLINE, Embase, Cochrane Database of Systematic Reviews, Web of Science, and Scopus for medical subject headings and text keywords related to NLP and triage. Databases were last searched on 01/01/2022. Risk of bias in studies was assessed using the Prediction model Risk of Bias Assessment Tool (PROBAST). Due to the high level of heterogeneity between studies and high risk of bias, a metanalysis was not conducted. Instead, a narrative synthesis is provided.

RESULTS

In total, 3730 studies were screened, and 20 studies were included. The population size varied greatly between studies ranging from 1.8 million patients to 598 triage notes. The most common outcomes assessed were prediction of triage score, prediction of admission, and prediction of critical illness. NLP models achieved high accuracy in predicting need for admission, triage score, critical illness, and mapping free-text chief complaints to structured fields. Incorporating both structured data and free-text data improved results when compared to models that used only structured data. However, the majority of studies (80%) were assessed to have a high risk of bias, and only one study reported the deployment of an NLP model into clinical practice.

CONCLUSION

Unstructured free-text triage notes have been used by NLP models to predict clinically relevant outcomes. However, the majority of studies have a high risk of bias, most research is retrospective, and there are few examples of implementation into clinical practice. Future work is needed to prospectively assess if applying NLP to data acquired at ED triage improves ED outcomes when compared to usual clinical practice.

Supporting clinical decision making in the emergency department for paediatric patients using machine learning: A scoping review protocol

INTRODUCTION

Machine learning as a clinical decision support system tool has the potential to assist clinicians who must make complex and accurate medical decisions in fast paced environments such as the emergency department. This paper presents a protocol for a scoping review, with the objective of summarising the existing research on machine learning clinical decision support system tools in the emergency department, focusing on models that can be used for paediatric patients, where a knowledge gap exists.

MATERIALS AND METHODS

The methodology used will follow the scoping study framework of Arksey and O'Malley, along with other guidelines. Machine learning clinical decision support system tools for any outcome and population (paediatric/adult/mixed) for use in the emergency department will be included. Articles such as grey literature, letters, pre-prints, editorials, scoping/literature/narrative reviews, non-English full text papers, protocols, surveys, abstract or full text not available and models based on synthesised data will be excluded. Articles from the last five years will be included. Four databases will be searched: Medline (EBSCO), CINAHL (EBSCO), EMBASE and Cochrane Central. Independent reviewers will perform the screening in two sequential stages (stage 1: clinician expertise and stage 2: computer science expertise), disagreements will be resolved by discussion. Data relevant to the research question will be collected. Quantitative analysis will be performed to generate the results.

DISCUSSION

The study results will summarise the existing research on machine learning clinical decision support tools in the emergency department, focusing on models that can be used for paediatric patients. This holds the promise to identify opportunities to both incorporate models in clinical practice and to develop future models by utilising reviewers from diverse backgrounds and relevant expertise.

Machine learning for the diagnosis of acute coronary syndrome using a 12-lead ECG: a systematic review

OBJECTIVES

Prompt diagnosis of acute coronary syndrome (ACS) using a 12-lead electrocardiogram (ECG) is a critical task for emergency physicians. While computerized algorithms for ECG interpretation are limited in their accuracy, machine learning (ML) models have shown promise in several areas of clinical medicine. We performed a systematic review to compare the performance of ML-based ECG analysis to clinician or non-ML computerized ECG interpretation in the diagnosis of ACS for emergency department (ED) or prehospital patients.

METHODS

We searched Medline, Embase, Cochrane Central, and CINAHL databases from inception to May 18, 2022. We included studies that compared ML algorithms to either clinicians or non-ML based software in their ability to diagnose ACS using only a 12-lead ECG, in adult patients experiencing chest pain or symptoms concerning for ACS in the ED or prehospital setting. We used QUADAS-2 for risk of bias assessment. Prospero registration CRD42021264765.

RESULTS

Our search yielded 1062 abstracts. 10 studies met inclusion criteria. Five model types were tested, including neural networks, random forest, and gradient boosting. In five studies with complete performance data, ML models were more sensitive but less specific (sensitivity range 0.59-0.98, specificity range 0.44-0.95) than clinicians (sensitivity range 0.22-0.93, specificity range 0.63-0.98) in diagnosing ACS. In four studies that reported it, ML models had better discrimination (area under ROC curve range 0.79-0.98) than clinicians (area under ROC curve 0.67-0.78). Heterogeneity in both methodology and reporting methods precluded a meta-analysis. Several studies had high risk of bias due to patient selection, lack of external validation, and unreliable reference standards for ACS diagnosis.

CONCLUSIONS

ML models have overall higher discrimination and sensitivity but lower specificity than clinicians and non-ML software in ECG interpretation for the diagnosis of ACS. ML-based ECG interpretation could potentially serve a role as a "safety net", alerting emergency care providers to a missed acute MI when it has not been diagnosed. More rigorous primary research is needed to definitively demonstrate the ability of ML to outperform clinicians at ECG interpretation.

The Utility of Predictive Modeling and a Systems Process Approach to Reduce Emergency Department Crowding: A Position Paper

Emergency department (ED) crowding and its main causes, exit block and boarding, continue to threaten the quality and safety of ED care. Most interventions to reduce crowding have not been comprehensive or system solutions, only focusing on part of the care procession and not directly affecting boarding reduction. This position paper proposes that the ED crowding problem can be optimally addressed by applying a systems approach using predictive modeling to identify patients at risk of being admitted to the hospital and uses that information to initiate the time-consuming bed management process earlier in the care continuum, shortening the time during which patients wait in the ED for an inpatient bed assignment, thus removing the exit block that causes boarding and subsequently reducing crowding.

Implementation of Prediction Models in the Emergency Department from an Implementation Science Perspective-Determinants, Outcomes, and Real-World Impact: A Scoping Review

STUDY OBJECTIVE

Prediction models offer a promising form of clinical decision support in the complex and fast-paced environment of the emergency department (ED). Despite significant advancements in model development and validation, implementation of such models in routine clinical practice remains elusive. This scoping review aims to survey the current state of prediction model implementation in the ED and to provide insights on contributing factors and outcomes from an implementation science perspective.

METHODS

We searched 4 databases from their inception to May 20, 2022: MEDLINE (through PubMed), Embase, Scopus, and CINAHL. Articles that reported implementation outcomes and/or contextual determinants under the Reach, Effectiveness, Adoption, Implementation Maintenance (RE-AIM)/Practical, Robust, Implementation, and Sustainability Model (PRISM) framework were included. Characteristics of studies, models, and results of the RE-AIM/PRISM domains were summarized narratively.

RESULTS

Thirty-six reports on 31 implementations were included. The most common prediction models implemented were early warning scores. The most common implementation strategies used were training stakeholders, infrastructural changes, and using evaluative or iterative strategies. Only one report examined ED patients' perspectives, whereas the rest were focused on the experience of health care workers or organizational stakeholders. Key determinants of successful implementation include strong stakeholder engagement, codevelopment of workflows and implementation strategies, education, and usability.

CONCLUSION

Examining ED prediction models from an implementation science perspective can provide valuable insights and help guide future implementations.

Application of machine learning approach in emergency department to support clinical decision making for SARS-CoV-2 infected patients

To support physicians in clinical decision process on patients affected by Coronavirus Disease 2019 (COVID-19) in areas with a low vaccination rate, we devised and evaluated the performances of several machine learning (ML) classifiers fed with readily available clinical and laboratory data. Our observational retrospective study collected data from a cohort of 779 COVID-19 patients presenting to three hospitals of the Lazio-Abruzzo area (Italy). Based on a different selection of clinical and respiratory (ROX index and PaO2/FiO2 ratio) variables, we devised an AI-driven tool to predict safe discharge from ED, disease severity and mortality during hospitalization. To predict safe discharge our best classifier is an RF integrated with ROX index that reached AUC of 0.96. To predict disease severity the best classifier was an RF integrated with ROX index that reached an AUC of 0.91. For mortality prediction the best classifier was an RF integrated with ROX index, that reached an AUC of 0.91. The results obtained thanks to our algorithms are consistent with the scientific literature an accomplish significant performances to forecast safe discharge from ED and severe clinical course of COVID-19.

Overinterpretation of findings in machine learning prediction model studies in oncology: a systematic review

OBJECTIVES

In biomedical research, spin is the overinterpretation of findings, and it is a growing concern. To date, the presence of spin has not been evaluated in prognostic model research in oncology, including studies developing and validating models for individualized risk prediction.

STUDY DESIGN AND SETTING

We conducted a systematic review, searching MEDLINE and EMBASE for oncology-related studies that developed and validated a prognostic model using machine learning published between 1st January, 2019, and 5th September, 2019. We used existing spin frameworks and described areas of highly suggestive spin practices.

RESULTS

We included 62 publications (including 152 developed models; 37 validated models). Reporting was inconsistent between methods and the results in 27% of studies due to additional analysis and selective reporting. Thirty-two studies (out of 36 applicable studies) reported comparisons between developed models in their discussion and predominantly used discrimination measures to support their claims (78%). Thirty-five studies (56%) used an overly strong or leading word in their title, abstract, results, discussion, or conclusion.

CONCLUSION

The potential for spin needs to be considered when reading, interpreting, and using studies that developed and validated prognostic models in oncology. Researchers should carefully report their prognostic model research using words that reflect their actual results and strength of evidence.

Machine learning does not outperform traditional statistical modelling for kidney allograft failure prediction

Machine learning (ML) models have recently shown potential for predicting kidney allograft outcomes. However, their ability to outperform traditional approaches remains poorly investigated. Therefore, using large cohorts of kidney transplant recipients from 14 centers worldwide, we developed ML-based prediction models for kidney allograft survival and compared their prediction performances to those achieved by a validated Cox-Based Prognostication System (CBPS). In a French derivation cohort of 4000 patients, candidate determinants of allograft failure including donor, recipient and transplant-related parameters were used as predictors to develop tree-based models (RSF, RSF-ERT, CIF), Support Vector Machine models (LK-SVM, AK-SVM) and a gradient boosting model (XGBoost). Models were externally validated with cohorts of 2214 patients from Europe, 1537 from North America, and 671 from South America. Among these 8422 kidney transplant recipients, 1081 (12.84%) lost their grafts after a median post-transplant follow-up time of 6.25 years (Inter Quartile Range 4.33-8.73). At seven years post-risk evaluation, the ML models achieved a C-index of 0.788 (95% bootstrap percentile confidence interval 0.736-0.833), 0.779 (0.724-0.825), 0.786 (0.735-0.832), 0.527 (0.456-0.602), 0.704 (0.648-0.759) and 0.767 (0.711-0.815) for RSF, RSF-ERT, CIF, LK-SVM, AK-SVM and XGBoost respectively, compared with 0.808 (0.792-0.829) for the CBPS. In validation cohorts, ML models' discrimination performances were in a similar range of those of the CBPS. Calibrations of the ML models were similar or less accurate than those of the CBPS. Thus, when using a transparent methodological pipeline in validated international cohorts, ML models, despite overall good performances, do not outperform a traditional CBPS in predicting kidney allograft failure. Hence, our current study supports the continued use of traditional statistical approaches for kidney graft prognostication.

Machine learning models predicting undertriage in telephone triage

BACKGROUND

Undertriaged patients have worse outcomes than appropriately triaged patients. Machine learning provides better triage prediction than conventional triage in emergency departments, but no machine learning-based undertriage prediction models have yet been developed for prehospital telephone triage. We developed and validated machine learning models for telephone triage.

MATERIALS AND METHODS

We conducted a retrospective cohort study with the largest after-hour house-call (AHHC) service dataset in Japan. Participants were ≥16 years and used the AHHC service between 1 November 2018 and 31 January 2021. We developed five prediction models based on support vector machine (SVM), lasso regression (LR), random forest (RF), gradient-boosted decision tree (XGB), and deep neural network (DNN). The primary outcome was undertriage, and predictors were telephone triage level and routinely available telephone-based data, including age, sex, 80 chief complaint categories and 10 comorbidities. We measured the area under the receiver operating characteristic curve (AUROC) for all the models.

RESULTS

We identified 15,442 eligible patients (age: 38.4 ± 16.6, male: 57.2%), including 298 (1.9%; age: 58.2 ± 23.9, male: 55.0%) undertriaged patients. RF and XGB outperformed the other models, with the AUROC values (95% confidence interval; 95% CI) of the SVM, LR, RF, XGB and DNN for undertriage being 0.62 (0.55-0.69), 0.79 (0.74-0.83), 0.81 (0.76-0.86), 0.80 (0.75-0.84) and 0.77 (0.73-0.82), respectively.

CONCLUSIONS

We found that RF and XGB outperformed other models. Our findings suggest that machine learning models can facilitate the early detection of undertriage and early intervention to yield substantially improved patient outcomes.KEY MESSAGESUndertriaged patients experience worse outcomes than appropriately triaged patients; thus, we developed machine learning models for predicting undertriage in the prehospital setting. In addition, we identified the predictors of risk factors associated with undertriage.Random forest and gradient-boosted decision tree models demonstrated better prediction performance, and the models identified the risk factors associated with undertriage.Machine learning models aid in the early detection of undertriage, leading to significantly improved patient outcomes and identifying undertriage-associated risk factors, including chief complaint categories, could help prioritize conventional telephone triage protocol revision.

Machine learning-based classification of circadian rhythm characteristics for mild cognitive impairment in the elderly

Introduction

Using wrist-wearable sensors to ecological transient assessment may provide a more valid assessment of physical activity, sedentary time, sleep and circadian rhythm than self-reported questionnaires, but has not been used widely to study the association with mild cognitive impairment and their characteristics.

Methods

31 normal cognitive ability participants and 68 MCI participants were monitored with tri-axial accelerometer and nocturnal photo volumetric pulse wave signals for 14 days. Two machine learning algorithms: gradient boosting decision tree and eXtreme gradient boosting were constructed using data on daytime physical activity, sedentary time and nighttime physiological functions, including heart rate, heart rate variability, respiratory rate and oxygen saturation, combined with subjective scale features. The accuracy, precision, recall, F1 value, and AUC of the different models are compared, and the training and model effectiveness are validated by the subject-based leave-one-out method.

Results

The low physical activity state was higher in the MCI group than in the cognitively normal group between 8:00 and 11:00 (P < 0.05), the daily rhythm trend of the high physical activity state was generally lower in the MCI group than in the cognitively normal group (P < 0.05). The peak rhythms in the sedentary state appeared at 12:00-15:00 and 20:00. The peak rhythms of rMSSD, HRV high frequency output power, and HRV low frequency output power in the 6h HRV parameters at night in the MCI group disappeared at 3:00 a.m., and the amplitude of fluctuations decreased; the amplitude of fluctuations of LHratio nocturnal rhythm increased and the phase was disturbed; the oxygen saturation was between 90 and 95% and less than 90% were increased in all time periods (P < 0.05). The F1 value of the two machine learning algorithms for MCI classification of multi-feature data combined with subjective scales were XGBoost (78.02) and GBDT (84.04).

Conclusion

By collecting PSQI Scale data combined with circadian rhythm characteristics monitored by wrist-wearable sensors, we are able to construct XGBoost and GBDT machine learning models with good discrimination, thus providing an early warning solution for identifying family and community members with high risk of MCI.

Artificial intelligence and its impact on the domains of universal health coverage, health emergencies and health promotion: An overview of systematic reviews

•

An overview of systematic reviews on the application of AI including 129 studies.

•

AI use is prominent in Universal Health Coverage, featuring image analysis in neoplasms.

•

Half of the reviews did not evaluate validation procedures nor reporting guidelines.

•

Risk of bias was only included un a third of the reviews.

•

There is not sufficient evidence to transfer AI to actual healthcare delivery.

Background

Artificial intelligence is fueling a new revolution in medicine and in the healthcare sector. Despite the growing evidence on the benefits of artificial intelligence there are several aspects that limit the measure of its impact in people’s health. It is necessary to assess the current status on the application of AI towards the improvement of people’s health in the domains defined by WHO’s Thirteenth General Programme of Work (GPW13) and the European Programme of Work (EPW), to inform about trends, gaps, opportunities, and challenges.

Objective

To perform a systematic overview of systematic reviews on the application of artificial intelligence in the people’s health domains as defined in the GPW13 and provide a comprehensive and updated map on the application specialties of artificial intelligence in terms of methodologies, algorithms, data sources, outcomes, predictors, performance, and methodological quality.

Methods

A systematic search in MEDLINE, EMBASE, Cochrane and IEEEXplore was conducted between January 2015 and June 2021 to collect systematic reviews using a combination of keywords related to the domains of universal health coverage, health emergencies protection, and better health and wellbeing as defined by the WHO’s PGW13 and EPW. Eligibility criteria was based on methodological quality and the inclusion of practical implementation of artificial intelligence. Records were classified and labeled using ICD-11 categories into the domains of the GPW13. Descriptors related to the area of implementation, type of modeling, data entities, outcomes and implementation on care delivery were extracted using a structured form and methodological aspects of the included reviews studies was assessed using the AMSTAR checklist.

Results

The search strategy resulted in the screening of 815 systematic reviews from which 203 were assessed for eligibility and 129 were included in the review. The most predominant domain for artificial intelligence applications was Universal Health Coverage (N = 98) followed by Health Emergencies (N = 16) and Better Health and Wellbeing (N = 15). Neoplasms area on Universal Health Coverage was the disease area featuring most of the applications (21.7 %, N = 28). The reviews featured analytics primarily over both public and private data sources (67.44 %, N = 87). The most used type of data was medical imaging (31.8 %, N = 41) and predictors based on regions of interest and clinical data. The most prominent subdomain of Artificial Intelligence was Machine Learning (43.4 %, N = 56), in which Support Vector Machine method was predominant (20.9 %, N = 27). Regarding the purpose, the application of Artificial Intelligence I is focused on the prediction of the diseases (36.4 %, N = 47). With respect to the validation, more than a half of the reviews (54.3 %, N = 70) did not report a validation procedure and, whenever available, the main performance indicator was the accuracy (28.7 %, N = 37). According to the methodological quality assessment, a third of the reviews (34.9 %, N = 45) implemented methods for analysis the risk of bias and the overall AMSTAR score below was 5 (4.01 ± 1.93) on all the included systematic reviews.

Conclusion

Artificial intelligence is being used for disease modelling, diagnose, classification and prediction in the three domains of GPW13. However, the evidence is often limited to laboratory and the level of adoption is largely unbalanced between ICD-11 categoriesand diseases. Data availability is a determinant factor on the developmental stage of artificial intelligence applications. Most of the reviewed studies show a poor methodological quality and are at high risk of bias, which limits the reproducibility of the results and the reliability of translating these applications to real clinical scenarios. The analyzed papers show results only in laboratory and testing scenarios and not in clinical trials nor case studies, limiting the supporting evidence to transfer artificial intelligence to actual care delivery.

A case study to investigate the impact of overcrowding indices in emergency departments

Background

Emergency department (ED) overcrowding is widespread in hospitals in many countries, causing severe consequences to patient outcomes, staff work and the system, with an overall increase in costs. Therefore, health managers are constantly looking for new preventive and corrective measures to counter this phenomenon. To do this, however, it is necessary to be able to characterize the problem objectively. For this reason, various indices are used in the literature to assess ED crowding. In this work, we explore the use of two of the most widespread crowding indices in an ED of an Italian national hospital, investigate their relationships and discuss their effectiveness.

Methods

In this study, two of the most widely used indices in the literature, the National Emergency Department Overcrowding Scale (NEDOCS) and the Emergency Department Working Index (EDWIN), were analysed to characterize overcrowding in the ED of A.O.R.N. “A. Cardarelli” of Naples, which included 1678 clinical cases. The measurement was taken every 15 minutes for a period of 7 days.

Results

The results showed consistency in the use of EDWIN and NEDOCS indices as measures of overcrowding, especially in severe overcrowding conditions. Indeed, in the examined case study, both EDWIN and NEDOCS showed very low rates of occurrence of severe overcrowding (2–3%). In contrast, regarding differences in the estimation of busy to overcrowded ED rates, the EDWIN index proved to be less sensitive in distinguishing these variations in the occupancy of the ED. Furthermore, within the target week considered in the study, the results show that, according to both EDWIN and NEDOCS, higher overcrowding rates occurred during the middle week rather than during the weekend. Finally, a low degree of correlation between the two indices was found.

Conclusions

The effectiveness of both EDWIN and NEDOCS in measuring ED crowding and overcrowding was investigated, and the main differences and relationships in the use of the indices are highlighted. While both indices are useful ED performance metrics, they are not always interchangeable, and their combined use could provide more details in understanding ED dynamics and possibly predicting future critical conditions, thus enhancing ED management.

Machine Learning in the Prediction of Trauma Outcomes: A Systematic Review

STUDY OBJECTIVE

Machine learning models carry unique potential as decision-making aids and prediction tools for improving patient care. Traumatically injured patients provide a uniquely heterogeneous population with severe injuries that can be difficult to predict. Given the relative infancy of machine learning applications in medicine, this systematic review aimed to better understand the current state of machine learning development and implementation to help create a basis for future research.

METHODS

We conducted a systematic review from inception to May 2021, using Embase, MEDLINE through Ovid, Web of Science, Google Scholar, and relevant gray literature, for uses of machine learning in predicting the outcomes of trauma patients. The screening and data extraction were performed by 2 independent reviewers.

RESULTS

Of the 14,694 identified articles screened, 67 were included for data extraction. Artificial neural networks comprised the most commonly used model, and mortality was the most prevalent outcome of interest. In terms of machine learning model development, there was a lack of studies that employed external validation, feature selection methods, and performed formal calibration testing. Significant heterogeneity in reporting was also observed between the machine learning models employed, patient populations, performance metrics, and features employed.

CONCLUSION

This review highlights the heterogeneity in the development and reporting of machine learning models for the prediction of trauma outcomes. While these models present an area of opportunity as an ancillary to clinical decision-making, we recommend more standardization and rigorous guidelines for the development of future models.

Suboptimal prehospital decision- making for referral to alternative levels of care - frequency, measurement, acceptance rate and room for improvement

Background

The emergency medical services (EMS) have undergone dramatic changes during the past few decades. Increased utilisation, changes in care-seeking behaviour and competence among EMS clinicians have given rise to a shift in EMS strategies in many countries. From transport to the emergency department to at the scene deciding on the most appropriate level of care and mode of transport. Among the non-conveyed patients some may suffer from “time-sensitive conditions” delaying diagnosis and treatment. Thus, four questions arise:

How often are time-sensitive cases referred to primary care or self-care advice?

How can we measure and define the level of inappropriate clinical decision-making?

What is acceptable?

How to increase patient safety?

Main text

To what extent time-sensitive cases are non-conveyed varies. About 5–25% of referred patients visit the emergency department within 72 hours, 5% are hospitalised, 1–3% are reported to have a time-sensitive condition and seven-day mortality rates range from 0.3 to 6%.

The level of inappropriate clinical decision-making can be measured using surrogate measures such as emergency department attendances, hospitalisation and short-term mortality. These measures do not reveal time-sensitive conditions. Defining a scoring system may be one alternative, where misclassifications of time-sensitive cases are rated based on how severely they affected patient outcome.

In terms of what is acceptable there is no general agreement. Although a zero-vision approach does not seem to be realistic unless under-triage is split into different levels of severity with zero-vision in the most severe categories.

There are several ways to reduce the risk of misclassifications. Implementation of support systems for decision-making using machine learning to improve the initial assessment is one approach. Using a trigger tool to identify adverse events is another.

Conclusion

A substantial number of patients are non-conveyed, including a small portion with time-sensitive conditions. This poses a threat to patient safety. No general agreement on how to define and measure the extent of such EMS referrals and no agreement of what is acceptable exists, but we conclude an overall zero-vision is not realistic. Developing specific tools supporting decision making regarding EMS referral may be one way to reduce misclassification rates.

Building artificial intelligence and machine learning models : a primer for emergency physicians

There has been a rise in the number of studies relating to the role of artificial intelligence (AI) in healthcare. Its potential in Emergency Medicine (EM) has been explored in recent years with operational, predictive, diagnostic and prognostic emergency department (ED) implementations being developed. For EM researchers building models de novo, collaborative working with data scientists is invaluable throughout the process. Synergism and understanding between domain (EM) and data experts increases the likelihood of realising a successful real-world model. Our linked manuscript provided a conceptual framework (including a glossary of AI terms) to support clinicians in interpreting AI research. The aim of this paper is to supplement that framework by exploring the key issues for clinicians and researchers to consider in the process of developing an AI model.

Can Prehospital Data Improve Early Identification of Sepsis in Emergency Department? An Integrative Review of Machine Learning Approaches

BACKGROUND

Sepsis is associated with high mortality, especially during the novel coronavirus disease 2019 (COVID-19) pandemic. Along with high monetary health care costs for sepsis treatment, there is a lasting impact on lives of sepsis survivors and their caregivers. Early identification is necessary to reduce the negative impact of sepsis and to improve patient outcomes. Prehospital data are among the earliest information collected by health care systems. Using these untapped sources of data in machine learning (ML)-based approaches can identify patients with sepsis earlier in emergency department (ED).

OBJECTIVES

This integrative literature review aims to discuss the importance of utilizing prehospital data elements in ED, summarize their current use in developing ML-based prediction models, and specifically identify those data elements that can potentially contribute to early identification of sepsis in ED when used in ML-based approaches.

METHOD

Literature search strategy includes following two separate searches: (1) use of prehospital data in ML models in ED; and (2) ML models that are developed specifically to predict/detect sepsis in ED. In total, 24 articles are used in this review.

RESULTS

A summary of prehospital data used to identify time-sensitive conditions earlier in ED is provided. Literature related to use of ML models for early identification of sepsis in ED is limited and no studies were found related to ML models using prehospital data in prediction/early identification of sepsis in ED. Among those using ED data, ML models outperform traditional statistical models. In addition, the use of the free-text elements and natural language processing (NLP) methods could result in better prediction of sepsis in ED.

CONCLUSION

This study reviews the use of prehospital data in early decision-making in ED and suggests that researchers utilize such data elements for prediction/early identification of sepsis in ML-based approaches.

Machine learning in vascular surgery: a systematic review and critical appraisal

Machine learning (ML) is a rapidly advancing field with increasing utility in health care. We conducted a systematic review and critical appraisal of ML applications in vascular surgery. MEDLINE, Embase, and Cochrane CENTRAL were searched from inception to March 1, 2021. Study screening, data extraction, and quality assessment were performed by two independent reviewers, with a third author resolving discrepancies. All original studies reporting ML applications in vascular surgery were included. Publication trends, disease conditions, methodologies, and outcomes were summarized. Critical appraisal was conducted using the PROBAST risk-of-bias and TRIPOD reporting adherence tools. We included 212 studies from a pool of 2235 unique articles. ML techniques were used for diagnosis, prognosis, and image segmentation in carotid stenosis, aortic aneurysm/dissection, peripheral artery disease, diabetic foot ulcer, venous disease, and renal artery stenosis. The number of publications on ML in vascular surgery increased from 1 (1991-1996) to 118 (2016-2021). Most studies were retrospective and single center, with no randomized controlled trials. The median area under the receiver operating characteristic curve (AUROC) was 0.88 (range 0.61-1.00), with 79.5% [62/78] studies reporting AUROC ≥ 0.80. Out of 22 studies comparing ML techniques to existing prediction tools, clinicians, or traditional regression models, 20 performed better and 2 performed similarly. Overall, 94.8% (201/212) studies had high risk-of-bias and adherence to reporting standards was poor with a rate of 41.4%. Despite improvements over time, study quality and reporting remain inadequate. Future studies should consider standardized tools such as PROBAST and TRIPOD to improve study quality and clinical applicability.

Development and validation of the SARICA score to predict survival after return of spontaneous circulation in out of hospital cardiac arrest using an interpretable machine learning framework

BACKGROUND

Accurate and timely prognostication of patients with out-of-hospital cardiac arrest (OHCA) who achieved the return of spontaneous circulation (ROSC) is crucial in clinical decision-making, resource allocation, and communications with next-of-kins. We aimed to develop the Survival After ROSC in Cardiac Arrest (SARICA), a practical clinical decision tool to predict survival in OHCA patients who attained ROSC.

METHODS

We utilized real-world Singapore data from the population-based Pan-Asian Resuscitation Outcomes Study between 2010-2018. We excluded patients without ROSC. The dataset was segmented into training (60%), validation (20%) and testing (20%) cohorts. The primary endpoint was survival (to 30-days or hospital discharge). AutoScore, an interpretable machine-learning based clinical score generation algorithm, was used to develop SARICA. Candidate factors were chosen based on objective demographic and clinical factors commonly available at the time of admission. Performance of SARICA was evaluated based on receiver-operating curve (ROC) analyses.

RESULTS

5970 patients were included, of which 855 (14.3%) survived. A three-variable model was determined to be most parsimonious. Prehospital ROSC, age, and initial heart rhythm were identified for inclusion via random forest selection. Finally, SARICA consisted of these 3 variables and ranged from 0 to 10 points, achieving an area under the ROC (AUC) of 0.87 (95% confidence interval: 0.84-0.90) within the testing cohort.

CONCLUSION

We developed and internally validated the SARICA score to accurately predict survival of OHCA patients with ROSC at the time of admission. SARICA is clinically practical and developed using an interpretable machine-learning framework. SARICA has unknown generalizability pending external validation studies.

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.

Risk of bias in studies on prediction models developed using supervised machine learning techniques: systematic review

OBJECTIVE

To assess the methodological quality of studies on prediction models developed using machine learning techniques across all medical specialties.

DESIGN

Systematic review.

DATA SOURCES

PubMed from 1 January 2018 to 31 December 2019.

ELIGIBILITY CRITERIA

Articles reporting on the development, with or without external validation, of a multivariable prediction model (diagnostic or prognostic) developed using supervised machine learning for individualised predictions. No restrictions applied for study design, data source, or predicted patient related health outcomes.

REVIEW METHODS

Methodological quality of the studies was determined and risk of bias evaluated using the prediction risk of bias assessment tool (PROBAST). This tool contains 21 signalling questions tailored to identify potential biases in four domains. Risk of bias was measured for each domain (participants, predictors, outcome, and analysis) and each study (overall).

RESULTS

152 studies were included: 58 (38%) included a diagnostic prediction model and 94 (62%) a prognostic prediction model. PROBAST was applied to 152 developed models and 19 external validations. Of these 171 analyses, 148 (87%, 95% confidence interval 81% to 91%) were rated at high risk of bias. The analysis domain was most frequently rated at high risk of bias. Of the 152 models, 85 (56%, 48% to 64%) were developed with an inadequate number of events per candidate predictor, 62 handled missing data inadequately (41%, 33% to 49%), and 59 assessed overfitting improperly (39%, 31% to 47%). Most models used appropriate data sources to develop (73%, 66% to 79%) and externally validate the machine learning based prediction models (74%, 51% to 88%). Information about blinding of outcome and blinding of predictors was, however, absent in 60 (40%, 32% to 47%) and 79 (52%, 44% to 60%) of the developed models, respectively.

CONCLUSION

Most studies on machine learning based prediction models show poor methodological quality and are at high risk of bias. Factors contributing to risk of bias include small study size, poor handling of missing data, and failure to deal with overfitting. Efforts to improve the design, conduct, reporting, and validation of such studies are necessary to boost the application of machine learning based prediction models in clinical practice.

SYSTEMATIC REVIEW REGISTRATION

PROSPERO CRD42019161764.

Models Predicting Hospital Admission of Adult Patients Utilizing Prehospital Data: Systematic Review Using PROBAST and CHARMS

Background

Emergency department boarding and hospital exit block are primary causes of emergency department crowding and have been conclusively associated with poor patient outcomes and major threats to patient safety. Boarding occurs when a patient is delayed or blocked from transitioning out of the emergency department because of dysfunctional transition or bed assignment processes. Predictive models for estimating the probability of an occurrence of this type could be useful in reducing or preventing emergency department boarding and hospital exit block, to reduce emergency department crowding.

Objective

The aim of this study was to identify and appraise the predictive performance, predictor utility, model application, and model utility of hospital admission prediction models that utilized prehospital, adult patient data and aimed to address emergency department crowding.

Methods

We searched multiple databases for studies, from inception to September 30, 2019, that evaluated models predicting adult patients’ imminent hospital admission, with prehospital patient data and regression analysis. We used PROBAST (Prediction Model Risk of Bias Assessment Tool) and CHARMS (Checklist for Critical Appraisal and Data Extraction for Systematic Reviews of Prediction Modeling Studies) to critically assess studies.

Results

Potential biases were found in most studies, which suggested that each model’s predictive performance required further investigation. We found that select prehospital patient data contribute to the identification of patients requiring hospital admission. Biomarker predictors may add superior value and advantages to models. It is, however, important to note that no models had been integrated with an information system or workflow, operated independently as electronic devices, or operated in real time within the care environment. Several models could be used at the site-of-care in real time without digital devices, which would make them suitable for low-technology or no-electricity environments.

Conclusions

There is incredible potential for prehospital admission prediction models to improve patient care and hospital operations. Patient data can be utilized to act as predictors and as data-driven, actionable tools to identify patients likely to require imminent hospital admission and reduce patient boarding and crowding in emergency departments. Prediction models can be used to justify earlier patient admission and care, to lower morbidity and mortality, and models that utilize biomarker predictors offer additional advantages.

Use of Time-to-Event Analysis to Develop On-Scene Return of Spontaneous Circulation Prediction for Out-of-Hospital Cardiac Arrest Patients

STUDY OBJECTIVE

We aimed to train and validate the time to on-scene return of spontaneous circulation prediction models using time-to-event analysis among out-of-hospital cardiac arrest patients.

METHODS

Using a Korean population-based out-of-hospital cardiac arrest registry, we selected a total of 105,215 adults with presumed cardiac etiologies between 2013 and 2018. Patients from 2013 to 2017 and from 2018 were analyzed for training and test, respectively. We developed 4 time-to-event analyzing models (Cox proportional hazard [Cox], random survival forest, extreme gradient boosting survival, and DeepHit) and 4 classification models (logistic regression, random forest, extreme gradient boosting, and feedforward neural network). Patient characteristics and Utstein elements collected at the scene were used as predictors. Discrimination and calibration were evaluated by Harrell's C-index and integrated Brier score.

RESULTS

Among the 105,215 patients (mean age 70 years and 64% men), 86,314 and 18,901 patients belonged to the training and test sets, respectively. On-scene return of spontaneous circulation was achieved in 5,240 (6.1%) patients in the former set and 1,709 (9.0%) patients in the latter. The proportion of emergency medical services (EMS) management was higher and scene time interval longer in the latter. Median time from EMS scene arrival to on-scene return of spontaneous circulation was 8 minutes for both datasets. Classification models showed similar discrimination and poor calibration power compared to survival models; Cox showed high discrimination with the best calibration (C-index [95% confidence interval]: 0.873 [0.865 to 0.882]; integrated Brier score at 30 minutes: 0.060).

CONCLUSION

Incorporating time-to-event analysis could lead to improved performance in prediction models and contribute to personalized field EMS resuscitation decisions.