Artificial intelligence for early diagnosis in emergency department

Abstract

In recent years, artificial intelligence (AI) has become an increasingly prominent player in emergency medicine, offering innovative tools to enhance the early diagnosis of acute conditions. This systematic review explores how AI, particularly through machine learning (ML) and deep learning (DL), is transforming the way physicians and healthcare professionals respond to high-stakes clinical scenarios. The evidence gathered shows that smart algorithms are capable of detecting complex patterns in clinical, diagnostic, and laboratory data, patterns that may even elude expert clinicians, especially under the high-pressure environment of the emergency room. From acute coronary syndrome to stroke, from sepsis to respiratory failure, AI has demonstrated impressive predictive power and provides real, practical support in risk stratification, triage optimization, and faster diagnosis. Equally important is its role in automated medical image analysis, which enables quicker and more accurate diagnostic decisions, offering real-time support for clinicians. However, the widespread adoption of these technologies also brings significant challenges: the need for algorithmic transparency, the necessity of earning the trust of healthcare providers, and the sensitive ethical issues related to patient data privacy. To overcome these barriers, it is essential to involve healthcare professionals in the development and implementation of AI technologies—ensuring their clinical expertise complements the analytical power of these new tools. Targeted training programs and large-scale validation studies are critical steps for ensuring the safe and effective use of AI. Ultimately, this review confirms that AI holds great promise as a catalyst for a more efficient, timely, and patient-centered approach to emergency medicine.

Introduction

As part of the modernization of diagnostic processes in medicine, artificial intelligence (AI) plays a fundamental role—especially in the early diagnosis of acute conditions in emergency and urgent care settings. The advent of AI and machine learning (ML) is truly revolutionizing the medical field, with a particularly strong impact on emergency medicine. In response to the growing demand from emergency departments (EDs) for accurate and efficient diagnostic tools, AI and ML have become essential resources that help physicians diagnose and treat acute medical conditions early, ultimately improving patient outcomes.

This is made possible by deep learning technologies, which are capable of analyzing large volumes of medical data and identifying complex patterns that the human mind might either miss or take too long to recognize—especially within the time-critical environment of the emergency department.

Numerous studies have shown that AI programs can significantly improve diagnostic speed, sensitivity, and specificity [1]. In particular, AI has already been successfully used to diagnose acute conditions like appendicitis and to recognize complex multisystem syndromes [2, 3], thanks to experimental models.

The use of machine learning algorithms has also been shown to enhance clinicians’ diagnostic capabilities in high-complexity cases [4]. In the fast-paced and dynamic setting of the emergency room, physicians often must make life-or-death decisions quickly, even with incomplete information. AI has been proven to significantly improve this clinical decision-making process [2].

For instance, convolutional neural networks (CNNs) have been employed to analyze advanced diagnostic images, such as CT scans and chest X-rays, providing crucial support for timely and accurate diagnoses [4, 5]. Likewise, automated image analysis powered by AI has greatly improved the detection of acute conditions such as ascites in emergency patients [6].

Advanced ML algorithms have also been successfully implemented for the early detection of critical time-sensitive conditions like stroke and myocardial infarction—where time is a decisive factor in patient survival [7].

AI applications in the diagnosis of acute illnesses go far beyond imaging analysis; they also encompass predictive analytics based on clinical data. In fact, AI not only interprets images from diagnostic tests but also enables the development of predictive models that can forecast both diagnoses and outcomes based on patient data.

Through data mining techniques and predictive modeling, clinicians can more effectively identify individual patient profiles and implement personalized, timely, and targeted diagnostic and therapeutic strategies [4]. For example, in the early detection of sepsis—a condition with high mortality risk—AI systems analyzing vital signs and laboratory results have already shown promising outcomes [5].

However, integrating these technologies into medical practice presents significant challenges. Issues such as algorithm transparency, result interpretability, and clinician acceptance will be critical to address moving forward.

Many physicians remain hesitant to rely on AI-based tools, largely due to a lack of trust in the results generated by technologies that are still in development or undergoing testing [6]. Recent research shows that involving healthcare professionals in the design and development of AI tools can improve both clinical effectiveness and acceptance [8].

Overcoming the barriers to the use of advanced technologies therefore requires targeted education and training for medical professionals [9]. Equally important are the ethical concerns associated with AI use, including algorithmic bias and the protection of sensitive health data.

The scientific community must strive to design and improve AI systems that are fair, transparent, and compliant with existing legal and ethical standards.

Ultimately, we are witnessing an opportunity to reform and significantly improve healthcare delivery through the application of AI to the early diagnosis of diseases in emergency settings. A careful evaluation of the capabilities and limitations of these technologies is essential to support their future clinical implementation.

These innovations show great promise in increasing hospital efficiency, optimizing diagnostic and treatment pathways, and reducing costs—but they also call for the upskilling and requalification of medical personnel.

A radical transformation of the diagnostic approach—one that promotes greater proficiency in interpreting data generated by these technologies [3]—could be achieved by integrating AI tools into medical education and training programs.

There is also strong interest within the scientific community in conducting further clinical research aimed at validating the safety, standardization, and reliability of AI systems to improve patient outcomes [9].

Still, fully understanding and addressing the technical and ethical limitations of these systems will be essential for ensuring an effective transition to a future in which AI plays a central role in the early diagnosis of acute conditions in emergency care [8].

This systematic review aims to synthesize and critically examine the current literature on the use of AI in the early diagnosis of diseases within emergency departments, assessing its impact on patient outcomes and the responsiveness of urgent care systems.

Materials and methods

The present systematic review was carried out according to the Preferred Reporting Items for Systematic Review (PRISMA) standards [10]. A systematic literature search and critical review of the collected studies were conducted. An electronic search of PubMed, Web of Science, Scopus, and IEEE Xplore from database inception until May 2025 was performed. The search terms were “artificial intelligence”, “machine learning”, “emergency medicine”, “early diagnosis” and “acute diseases”, and these terms were searched in the title, abstract, and keywords. The bibliographies of all located papers were examined and cross-referenced to further identify relevant literature. A methodological appraisal of each study was conducted according to the PRISMA standards, including an evaluation of bias. The data collection process included study selection and data extraction. The following inclusion criteria were used: (1) original research articles, (2) reviews and mini-reviews, (3) case reports/series, (4) and only papers written in English. Non-English papers were excluded.

Papers concerning the use of artificial intelligence on chronic diseases or elective surgical interventions were excluded.

Disagreements concerning eligibility among the researchers were resolved by consensus.

Results

The initial literature search identified a total of 3128 articles, of which 1245 were considered potentially relevant. After removing duplicates and carefully selecting based on titles and abstracts, 467 articles underwent full-text evaluation. Ultimately, 30 studies met the predefined inclusion criteria and were included in the systematic review (see Fig. 1).

Fig. 1.

PRISMA flow diagram of this systematic review

Overview of included studies

The included studies encompass a range of AI applications tailored to various diseases encountered in emergency settings (for every subheading see Table 1), including

Table 1.

Papers analyzed in this systematic review

S. N°	Studio	Pathology	AI model	Study objective	Results
	Xu et al.	ACS	ML	Risk stratification of patients with chest pain	More efficient triage process; reduced time to diagnosis
	Bishop et al.	ACS	ANN	Evaluation of ANN accuracy compared to computerized algorithms and physician interpretations	Greater diagnostic accuracy of ANN compared to computerized algorithms and physician interpretations
	McCord J et al.	ACS	ML (MI3)	Early diagnosis of AMI	100% sensitivity in identifying AMI within 30 min
	Wu CC et al.	ACS	ANN	Differential diagnosis between NSTEMI and unstable angina	High diagnostic accuracy (92.9%) and AUROC (98.4%) in rapid NSTEMI diagnosis in emergency settings
	Green et al.	ACS	ANN	Prediction of ACS in patients with chest pain	Greater PPV in diagnosing ACS with the ANN model compared to logistic regression
	Olsson et al.	ACS	ANN	Evaluation of ANN effectiveness in supporting ECG interpretation in patients with suspected ACS	Improved diagnostic sensitivity for ACS among trainees with 95% sensitivity and 88% specificity
	Xiao et al.	ACS	ML	Improve early diagnosis of AMI by integrating ECG tracings and demographic data	Superior performance with AUC 92.1%, accuracy 87.4% of the model compared to models based solely on ECG
	de Capretz et al.	ACS	ML (rete neurale convoluzionale)	Prediction of AMI or death at 30 days in a patient with chest pain	Rapid risk stratification (safe rule-out in 55% of cases, rule-in in 5.3% of cases) without the need for serial tests
	Choi et al.	ACS	ML	Prediction of significant coronary lesions in patients with chest pain and dyspnea undergoing coronary angiography	Greater support for clinicians in the differential diagnosis to distinguish ACS from other conditions
	O'Connell et al.	Ischemic stroke	CNN	Evaluating the effectiveness of an ANN model in identifying patients at risk of acute ischemic stroke based on imaging data	AUCROC 0.87 in the capacity for early diagnosis and in the treatment decision-making process
	Murray NM et al.	Ischemic stroke	ML; CNN	Evaluating the sensitivity of ML algorithms in the early diagnosis and intervention for patients with large-vessel occlusion ischemic strokes	ML models improve sensitivity in detecting occlusions
	Rajpurkar et al.	Pneumonia	DL	Assessment of sensitivity in pneumonia diagnosis through application of DL algorithms on chest radiographs	Sensitivity up to 92% in pneumonia detection in adult patients
	Sadegh-Zatel et al.	Pulmonary embolis	ML	Assessment of early mortality risk prediction in patients with pulmonary embolism using 5 ML algorithms	The ML models correctly predict the mortality of patients with pulmonary embolism
	Han et al	Acute respiratory failure	CNN	Assessment of early identification capacity in patients at risk of acute respiratory failure admitted to the ER using a CNN AI model	The CNN model outperformed predictive models based only on clinical parameters
	Desautels et al.	Sepsis	ML	Evaluation of the effectiveness of AI algorithms in the early diagnosis of sepsis	Predictive accuracy of 91% of dynamic algorithms using vital signs and lab test results
	Liu et al.	Sepsis	ML	Evaluation by 3 ML models in risk stratification of sepsis in patients with acute pancreatitis	The GBDT model achieved the best performance with an AUC of 0.985, outperforming the logistic regression model and the main clinical scoring systems in the early identification of sepsis
	Niemantsverdriet et al.	Sepsis	ML	Evaluation of the predictive capability of ML models compared to clinical tools in the diagnosis of sepsis	The ML models identified less conventional variables, such as eosinophils and platelet distribution, as potential biomarkers. This paves the way for new, potentially more reliable diagnostic models for early sepsis detection in clinical practice
	Nederpelt et al.	Acute injuries	ML	Application of AI models in patients with gunshot wounds for the prediction (and management) of shock, as well as the need for massive transfusion and major surgery	The model, trained on data from the Trauma Quality Improvement Program, achieved good performance: AUC of 0.89 for shock, 0.86 for massive transfusion, and 0.82 for major surgery
	Hu Z et al.	Cervical fractures	CNN	Evaluation of the diagnostic accuracy of seven AI models in detecting cervical fractures on over 1,800 CT scans	The models maintained good accuracy (AUC ~ 0.88–0.89), also detecting fractures missed by radiologists
	Lee SH et al.	Pelvic fractures	DL	Evaluation of the automatic classification capability of pelvic fractures on X-rays (according to AO/OTA system) by a DL model	The algorithm achieved high performance in bone segmentation and good results in fracture type classification
	Li K et al.	Acute traumatic coagulopathy (ATC)	ML	Prediction capability of ATC by ML model in patients with major trauma	The model based on the Random Forest algorithm showed an accuracy of 94%, precision of 93.3%, F1 score of 93.4%, and recall score of 94%, outperforming the logistic regression model
	Kui B et al.	Acute pancreatitis	ML	Prediction capability of the severity of acute pancreatitis by an AI model based on ML	EASY-APP is a practical, accurate, and readily applicable tool that allows early identification of patients at risk of severe progression of acute pancreatitis in the first few hours, improving early clinical intervention
	Andersson B et al.	Acute pancreatitis	ANN	Prediction capability of the severity of acute pancreatitis by a predictive model based on ANN	The neural network proved to be particularly effective in correctly classifying at-risk patients, proving to be a valuable early decision-making tool to identify subjects destined for a severe course
	Lee et al.	AKI due to iodinated contrast medium	ML	Capability of five ML models to predict contrast-induced AKI using clinical and biochemical parameters	. The LGB model showed the best performance (AUROC = 0.731), allowing physicians to administer nephroprotective treatment promptly
	Yamao et al.	AKI	ML	Predictive capacity for oliguria in ICU patients, using an ML model	This ML model, utilizing clinical and biochemical parameters, can predict the onset of oliguria at 6 and 72 h with high accuracy, obtaining AUCs of 0.964 and 0.916, respectively
	Wei et al.	AKI	ML	Predictive capacity of AKI in patients with acute respiratory distress syndrome (ARDS), using an ML model	Among 11 tested models, XGBoost showed the best performance, with an AUC of 0.865. The use of ML can improve the early identification of AKI in patients with ARDS, supporting more timely and effective clinical decisions
	Schipper et al.	Acute appendicitis	ML	Development of two ML models, with the aim of increasing the diagnostic accuracy of appendicitis in patients with acute abdominal pain. Comparison of the predictive accuracy of these models with the Alvarado score and that of three Emergency Department doctors	A greater predictive accuracy was observed in both ML models compared to the Alvarado score. Meanwhile, compared to the three Emergency Department doctors, the ML models showed greater performance gain without the integration of laboratory data, and matched or exceeded the doctors’ performance when using biochemical parameters
	Su et al.	Acute appendicitis	NLP	Evaluation of the predictive capacity in the diagnosis of acute appendicitis using an NLP predictive model	The integration of structured and unstructured data significantly improved the predictive accuracy of the diagnosis of acute appendicitis, both in adults and children
1. 3	Shung D	Acute gastrointestinal hemorrhage	NLP	Development of NLP-based algorithms and decision rules to identify patients with acute gastrointestinal bleeding early in the Emergency Department	An automated decision rule based on key clinical terms was found to be more accurate than standard tools (SNM) and can support the automated real-time activation of risk stratification systems

Legend: ACS acute coronary syndrome, ARI acute renal injury, ML machine learning, ANN artificial neural networks, CNN convolutional neural networks, DL deep learning, NLP natural language processing, PPV positive predictive value

AI hierarchy

Artificial intelligence (AI)

∣ ── Machine learning (ML)

∣ └── Deep learning (DL)

└── Natural language processing (NLP)

Acute coronary syndrome (ACS)

The study by Xu et al. [11] employed machine learning (ML) models to stratify the risk of patients presenting with chest pain, demonstrating superior predictive performance compared to traditional scoring systems. The use of ML algorithms led to a more efficient triage process, significantly reducing time to diagnosis.

The integration of artificial neural networks (ANNs) in ACS diagnosis was explored by Bishop et al. [12]. This scoping review, analyzing 14 studies, highlighted greater diagnostic accuracy using ANN compared to computerized algorithms and physician interpretation, regardless of specialization or years of experience. The review emphasizes the potential for ANN interpretation to be implemented in daily clinical practice in Emergency Departments, particularly for patients at risk of obstructive coronary artery disease, facilitating faster and more accurate diagnoses and timely reperfusion therapies.

Emakhu et al. [13] used algorithms to identify patients with ACS based on presenting symptoms, compared to alternative diagnoses. Variables used to assess ACS risk included systolic blood pressure, brain natriuretic peptide, chronic heart disease, coronary artery disease, creatinine, glucose, prior myocardial infarction, heart rate, nephrotic syndrome, red cell distribution width, and troponin levels. Their dynamic algorithms achieved a sensitivity of 86.3% and an AUROC of 93.3%. This framework could serve as a useful tool to support emergency physicians in the early identification of ACS patients, reducing misdiagnosis and improving clinical management.

The study by McCord J et al. [7] evaluated the use of a machine learning algorithm (MI3 – Myocardial Ischemic Injury Index) for early diagnosis of acute myocardial infarction (AMI) in patients presenting to the emergency department. MI3 combines age, sex, and high-sensitivity cardiac troponin I levels at time 0 and 30 min to generate a score from 0 to 100 indicating the likelihood of AMI. The study involved 529 patients, of whom 42 (7.9%) were confirmed to have an infarction. Patients were stratified into low-, intermediate-, and high-risk groups based on the MI3 score.

The algorithm achieved 100% sensitivity for identifying AMI within 30 min, allowing rapid rule-out in 67% of patients (low-risk group, MI3 ≤ 3.13). A 30–45-day follow-up revealed a very low incidence of major cardiac events (0.6%) in this group, suggesting the algorithm may support safe early discharge decisions.

Wu CC et al. [14]—Comput Methods Programs Biomed—developed an artificial neural network-based AI model to predict non-ST-elevation myocardial infarction (NSTEMI) in patients with chest pain, distinguishing it from unstable angina. The model, trained on clinical data from 268 patients, identified predictive variables such as blood pressure, QTc interval, liver enzymes, and troponin. It demonstrated high accuracy (92.9%) and AUROC (98.4), with excellent sensitivity and specificity, suggesting strong potential to support rapid NSTEMI diagnosis and triage in emergency settings.

Green et al. [15] compared the use of artificial neural networks (ANN) and logistic regression for predicting ACS in patients presenting with chest pain in the ED. Their findings show that ANN, using only ECG data, was effective in identifying patients without ACS. Adding clinical data did not improve performance. ANN outperformed logistic regression and may be a useful tool for the safe discharge of low-risk patients.

Olsson et al. [16] developed an artificial neural network to improve ECG interpretation in patients with ACS. When used by trainees, the network improved sensitivity from 68 to 93% without reducing specificity. The network itself achieved 95% sensitivity and 88% specificity, suggesting it may enhance diagnostic accuracy for trainees to levels approaching that of expert cardiologists.

Xiao et al. [17] developed a machine learning model that integrates ECG tracings with demographic data to improve early AMI diagnosis. Compared to ECG-only models, the multimodal model showed better performance (AUC 92.1%, accuracy 87.4%) even in the presence of confounding cardiac conditions. These results support the utility of multimodal approaches for real-world clinical application.

The study by de Capretz et al. [18] assessed the effectiveness of ML models in predicting AMI or 30-day mortality in patients with chest pain in the ED. Using initial clinical data (age, sex, ECG, lab tests), the best-performing model—a convolutional neural network—enabled safe rule-out in 55% and rule-in in 5.3% of cases, outperforming the ESC 0 h algorithm. The results suggest that AI can facilitate rapid risk stratification without requiring serial testing.Choi et al. [19] evaluated the use of ML to predict significant coronary lesions in patients presenting with chest pain or dyspnea who underwent coronary angiography. ML models were tested using progressively comprehensive clinical data (medical history, ECG, echocardiogram). The best ML model achieved an AUC of 0.83 in internal validation and 0.79 in external validation. While some predictive variables overlapped with traditional methods, ML identified additional discriminators, showing its potential to help clinicians distinguish true coronary syndromes from mimicking conditions—reducing unnecessary coronary angiographies.

Stroke

A large-scale study by O'Connell et al. [20] used convolutional neural networks (CNNs) to identify patients at risk of acute ischemic stroke based on imaging data. The algorithm demonstrated an area under the receiver operating characteristic curve (AUC-ROC) of 0.87, highlighting its potential to assist in early diagnosis and treatment decision-making.

The study by Murray NM et al. [21] presents a systematic review on the use of artificial intelligence (AI) for the diagnosis of acute ischemic stroke due to large vessel occlusion. Analyzing 20 studies published between 2014 and 2019, the authors describe the use of machine learning algorithms such as random forest learning (RFL) and convolutional neural networks (CNNs), respectively applied for assessments like ASPECTS scoring and for the automatic detection of large vessel occlusions. AI enhances sensitivity in detecting occlusions and may accelerate triage and therapeutic intervention.

Lung diseases

The application of deep learning algorithms for the diagnosis of pneumonia from chest X-rays was explored by Rajpurkar et al. [22]. This study reported sensitivities of up to 92% in detecting pneumonia in adult patients. These innovations mark critical advancements in diagnostic radiology, significantly influencing patient management in the Emergency Department.

Sadegh-Zatel et al. [23] assessed how the use of machine learning (ML) algorithms can enable accurate prediction of early mortality risk in patients with pulmonary embolism (PE). Specifically, the study employed five ML models, various oversampling techniques tailored to each model, and clinical and laboratory parameters. The random forest (RF) model with random oversampling demonstrated superior predictive performance compared to the other algorithms. As such, this ML technique may reliably predict mortality in PE patients and assist healthcare professionals in making appropriate therapeutic decisions.

A further study by Han et al. [24] investigated the ability of an AI model to identify patients at risk of acute respiratory failure (defined as the need for advanced ventilatory support) within 72 h of admission to the Emergency Department. The model demonstrated high predictive accuracy (AUROC of 0.840 and 0.743), significant associations with clinical outcomes, and strong potential to assist in triage decision-making. It outperformed predictive models such as MEWS and XGBoost, which rely solely on clinical parameters.

Sepsis

Several noteworthy studies, including that by Desautels et al. [25], have demonstrated the effectiveness of AI algorithms in the early identification of sepsis. Their dynamic algorithms, which analyze vital signs and laboratory results, achieved a predictive accuracy of 91%, showing significant potential to improve mortality rates among septic patients.

Another study by Liu et al. [26] developed machine learning (ML)-based predictive models to identify the risk of sepsis in patients with acute pancreatitis. Among six tested models, the Gradient Boosting Decision Tree (GBDT) delivered the best performance, with an AUC of 0.985, significantly outperforming logistic regression and major clinical scoring systems (SOFA, qSOFA, SIRS, BISAP, SAPS II). The GBDT model proved to be an effective tool for early sepsis identification and clinical decision support.

The study by Niemantsverdriet et al. [27] explored the use of machine learning to improve sepsis diagnosis in the emergency department, using diagnostic labels assigned by an independent Endpoint Adjudication Committee rather than relying on traditional clinical tools or often-inaccurate ICD codes. By analyzing 95 routine clinical variables from 375 ED visits, the ML models identified both well-known predictors and less conventional variables, such as eosinophil count and platelet distribution width—suggesting novel potential biomarkers. This approach could lead to more reliable and clinically useful diagnostic models.

Trauma assessment

The study conducted by Salimi et al. [28] applied AI models to predict the need for surgical intervention in trauma patients based on initial clinical assessment data. These models not only improved diagnostic accuracy but also reduced decision-making time, enhancing the overall efficiency of emergency care.

In the study by Nederpelt et al. [29], the authors developed an artificial intelligence algorithm for field triage of patients with torso gunshot wounds, predicting the risk of shock, the need for massive transfusion, and major surgery. Trained using data from the Trauma Quality Improvement Program, the model demonstrated strong performance: AUC of 0.89 for shock, 0.86 for massive transfusion, and 0.82 for major surgery. The algorithm showed high accuracy and confidence in predictions, with the potential to improve both civilian and military field triage, with further iterations planned for refinement.

Hu Z et al. [30] evaluated the performance of the top seven AI models from the RSNA Challenge 2022 for detecting cervical spine fractures on over 1800 clinical CT scans acquired at a level I trauma center. While the models showed a slight decline in performance compared to the competition setting, they maintained good accuracy (AUC ~ 0.88–0.89) and were able to detect fractures missed by radiologists. The models performed better on non-contrast CT scans, while errors were often related to osseous degeneration. These AI models appear promising as diagnostic support tools and warrant further clinical validation.

Lee SH et al. [31], in a 2024 study published in Scientific Reports and conducted at the Gachon University Gil Medical Center, developed and validated a deep learning algorithm for the automatic classification of pelvic fractures on radiographs according to the AO/OTA system. Retrospective analysis of radiographs from 773 patients with fractures and 167 without showed high performance in bone segmentation and good results in fracture type classification. The findings suggest that AI can support rapid classification of pelvic fractures in acute care settings, potentially benefiting emergency contexts.

Li K et al. [32], in a 2020 study, developed a machine learning model to predict acute traumatic coagulopathy (ATC)—a condition that occurs early after trauma and affects approximately 30% of severely injured patients admitted to the emergency department. Using data from 1385 patients, the Random Forest-based model achieved 94% accuracy, 93.3% precision, an F1 score of 93.4%, and a recall score of 94%, outperforming logistic regression models. This tool can rapidly identify patients at risk of ATC, enabling earlier intervention and improved resource management.

Acute pancreatitis

The study published by Kui B et al. [33] in Clinical and Translational Medicine (2022) presents EASY-APP, an artificial intelligence model developed for the early prediction of acute pancreatitis severity. Based on data from over 4700 patients and using machine learning algorithms (XGBoost), the system relies on six simple clinical variables available at admission (respiratory rate, temperature, abdominal reflex, sex, age, and blood glucose). EASY-APP is a practical, accurate, and immediately applicable tool that enables the identification—within the first few hours—of patients at risk of developing severe acute pancreatitis, thereby improving early clinical intervention.

Andersson B et al. [34], in a study published in Pancreatology, developed a predictive model based on artificial neural networks to assess the severity of acute pancreatitis at the time of hospital admission. Two patient groups (208 individuals) were analyzed, and six relevant clinical variables were selected: duration of pain before arrival, creatinine, hemoglobin, ALT, heart rate, and white blood cell count. The neural network proved particularly effective in correctly classifying at-risk patients, demonstrating strong potential as an early decision-support tool to identify those likely to experience a severe disease course.

Acute kidney injury (AKI)

The study published in Scientific Reports (2022) by Ang Y, Li S et al. [35] developed and validated AKI-RiSc, an interpretable clinical score for the early identification of acute kidney injury in patients admitted from the emergency department. Utilizing a large retrospective dataset (119,468 admissions), the model is based on six simple and easily accessible variables: serum creatinine, serum bicarbonate, pulse, systolic blood pressure, diastolic blood pressure, and age. AKI-RiSc is a simple and sensitive clinical tool (82.6%) that outperforms existing models and can be easily implemented in emergency settings to allow timely intervention in patients at risk for AKI, with potential for international adoption.

The study by Lee et al. [36] developed five machine learning models to predict contrast-induced acute kidney injury (CI-AKI), a common complication associated with the use of iodinated contrast media in emergency radiologic investigations. The clinical and laboratory parameters used in these models included serum creatinine, systolic blood pressure, serum albumin, estimated glomerular filtration rate (eGFR), blood urea nitrogen, body weight, uric acid, hemoglobin, triglycerides, and body temperature. The Light Gradient Boosting Machine (LGB) model achieved the best performance (AUROC = 0.731), enabling clinicians to recognize AKI early and initiate nephroprotective treatment promptly.

The development of a machine learning (ML) model to predict oliguria—a precursor of AKI—in ICU patients was explored in a retrospective cohort study by Yamao et al. The study demonstrated that this ML algorithm, using clinical and biochemical parameters, could predict the onset of oliguria at 6 and 72 h with high precision, yielding AUC values of 0.964 and 0.916, respectively. Given the close link between oliguria and AKI, this algorithm holds significant clinical and therapeutic potential for the early management of this complication in emergency medicine.

A study by Wei et al. [37] employed machine learning techniques to predict the onset of AKI in patients with acute respiratory distress syndrome (ARDS). Among the 11 models tested, XGBoost showed the best performance, with an optimized AUC of 0.865. A compact model using ten variables maintained high predictive accuracy. These results indicate that ML can enhance early identification of AKI in ARDS patients, supporting more timely and effective clinical decision-making.

Acute abdominal pain

A study by Schipper et al. [38] developed two machine learning (ML) models aimed at improving diagnostic accuracy for appendicitis in patients presenting to the emergency department with acute abdominal pain. One ML model used only clinical data, while the other combined clinical and laboratory data. The predictive accuracy of these models was compared with that of the Alvarado score (a widely used clinical scoring system for appendicitis) and the diagnostic performance of three emergency physicians. Both ML models demonstrated higher predictive accuracy than the Alvarado score. Compared to the emergency physicians, the ML models showed greater performance gains when only clinical data were used and matched or outperformed the physicians when laboratory data were included.

The study by Su et al. [39] evaluated predictive models for diagnosing acute appendicitis (AA) in patients presenting to the emergency department with nonspecific abdominal symptoms, using both structured data and free-text clinical notes from a large U.S. national database. The integration of structured and unstructured data significantly improved the predictive accuracy of AA diagnosis in both adults and children. The results suggest that natural language processing (NLP) techniques can enhance decision-support tools, enabling faster and more accurate diagnosis in emergency settings.

Gastrointestinal bleeding

The study by Shung D. et al. [40] developed and compared algorithms based on natural language processing (NLP) and decision rules to enable early identification of patients with acute gastrointestinal bleeding in the emergency department, using data from electronic health records. An automated decision rule based on key clinical terms proved more accurate than standard tools (such as SNM) and could support the real-time activation of automated risk stratification systems, thereby improving triage and clinical management.

Discussion

The study by Shung D. et al. [40] developed and compared algorithms based on natural language processing (NLP) and decision rules to enable early identification of patients with acute gastrointestinal bleeding in the emergency department, using data from electronic health records. An automated decision rule based on key clinical terms proved more accurate than standard tools (such as SNM) and could support the real-time activation of automated risk stratification systems, thereby improving triage and clinical management.

The integration of AI applications in the early diagnosis of diseases within emergency departments offers compelling advantages, including increased diagnostic accuracy, reduced time to intervention, and overall improvement in patient care. However, challenges remain, such as the need for robust validation across diverse populations, addressing ethical concerns related to AI usage, and integrating AI systems into existing clinical workflows.

Furthermore, this systematic review highlights the importance of clinician involvement in the development of AI systems, ensuring that these tools provide complementary support to healthcare providers rather than replacing their expertise. The engagement of professionals in interpreting and integrating AI results can enhance trust and practicality in real-world applications.

Modern cardiology can particularly benefit from AI in diagnosing myocardial infarction and acute coronary syndrome (ACS), promising significant improvements in timeliness and diagnostic accuracy. Early diagnosis of ACS is crucial for prognosis, given its high global morbidity and mortality rates [41, 42].

There is considerable attention on enhancing diagnostic processes to reduce myocardial damage and increase the chances of recovering ischemic heart tissue [41]. The first step involves faster risk stratification during triage, which reduces diagnosis time. The second step relates to improved diagnostic accuracy, as identified by Bishop et al. [12]. Another key aspect is the potential to rule out differential diagnoses.

AI can also integrate patient parameters and blood biomarkers, such as ischemia-modified albumin (IMA) and troponin, to further enhance ACS diagnosis. Common parameters used in AI and ML models include: sex, age, systolic blood pressure, B-type natriuretic peptide, chronic heart disease, coronary artery disease, creatinine, glucose, myocardial infarction, heart rate, nephrotic syndrome, red blood cell distribution width, and troponin level.

These biomarkers, when combined with AI-developed predictive models, can provide critical and timely diagnostic insights, improving the sensitivity and specificity of ACS diagnosis [43]. Studies have shown that IMA’s diagnostic value increases when combined with other myocardial injury markers, further enhancing reliability in patients with suspected acute ischemia [43].

AI also shows promise in cardiovascular imaging analysis. Techniques such as coronary angiography and multidetector computed tomography (MDCT) can identify coronary anomalies and atherosclerotic lesions more sensitively and specifically than human interpretation [44]. While not replacing human analysis, these technologies enable more accurate and timely diagnoses, aiding in the evaluation of both STEMI and NSTEMI [44].

Algorithms can also accurately identify patients safe for discharge—those at low risk of recurrence or short-term complications [7]. However, these algorithms should be stratified for different populations with varying cardiovascular risk factors [45].

Like ACS, stroke is a time-dependent condition where rapid diagnosis is critical to minimize permanent brain damage. AI-based approaches have shown promise in improving diagnostic sensitivity and specificity and in supporting swift clinical decisions [46–48].

The most impactful use of AI in stroke care lies in interpreting imaging data. CNN-based software, like e-ASPECTS, can identify high-risk stroke patients (O’Connell et al.) [20]. These technologies have proven non-inferior in sensitivity to physicians in the early stages of cerebral ischemia diagnosis [49]. Such software also allows for accurate patient stratification and shorter response times.

AI systems greatly aid emergency physicians by enabling early recognition of major vascular occlusions. These algorithms analyze diagnostic images in real-time, assisting initial triage and improving patient selection for treatments such as mechanical thrombectomy [46]. Physician validation of algorithmic findings has shown high accuracy rates, enhancing targeted patient management [48].

Machine learning techniques have also been used to predict clinical outcomes following intra-arterial therapy for ischemic stroke. Perfusion and diffusion analysis allows precise assessment of salvageable brain tissue, facilitating personalized therapeutic decisions [50, 51]. However, it is important to assess factors like reliability and generalizability across different populations [52], and to ensure appropriate training for healthcare providers using these technologies [36].

Respiratory diseases, another major cause of emergency department morbidity and mortality [53, 54], are also seeing benefits from AI in early diagnosis and treatment. Pulmonary embolism (PE), a condition with high mortality, demands early diagnosis to improve outcomes. Specific predictive scores for PE risk can be generated by ML algorithms. For example, the “Pulmonary Embolism Result Forecast Model” (PERFORM), validated in multicenter studies, serves as a clinical decision support tool [53], showing significant diagnostic accuracy.

D-dimer levels have also been used to stratify PE risk [55], allowing for earlier identification and timely intervention. Acute respiratory failure (ARF) is another potentially fatal condition where early diagnosis is vital. Integrated analysis of clinical and imaging data has improved ARDS diagnostic accuracy and supported more informed clinical decisions [54].

Recent developments also include analyzing digitized breath sounds via deep learning, identifying and classifying abnormal respiratory sounds such as crackles and wheezes. These systems enable rapid, remote diagnosis, particularly valuable in telemedicine.

Sepsis is another critical condition requiring rapid detection and treatment. Machine learning applications are improving sepsis prediction quality. Wang et al. demonstrated that AI can aid clinical decision-making, especially in ICUs [56], improving clinical practice and patient outcomes.

Le et al. found that algorithms like the Artificial Intelligence Sepsis Expert (AISE) can predict sepsis onset 4–12 h before clinical recognition [57, 58]. ML techniques can also predict sepsis-related cardiac events [59], and text mining of electronic health records enables early sepsis detection [60].

Emerging diagnostic biomarkers, identified through ML analysis of genomic data, are opening new paths for sepsis diagnosis [61]. Combining AI-based approaches with human clinical expertise represents the future of early sepsis diagnosis [62].

AI is revolutionizing trauma care by quickly and accurately identifying signs of injury, improving clinical outcomes. According to Hunter et al. [63], AI systems can support decision-making throughout the entire trauma care process, especially prognosis. ML analysis of clinical records optimizes patient triage in emergency departments, improving outcome prediction accuracy [64].

ML models assess trauma-related risks, supporting therapeutic decisions. Early diagnosis of trauma-induced hemorrhage is crucial; AI and ML models can analyze vital signs and biometric data to predict complications [65]. Harvin et al. show that ML already assists in determining surgical intervention for trauma patients [65]. AI-enhanced radiology enables early detection of abdominal trauma, even when subtle signs elude experts [66].

One specific application is the early detection of pancreatic injury. Huang et al. developed a clinical decision support model to identify pancreatic damage in abdominal trauma patients [67]. In acute pancreatitis diagnosis, AI combines advanced imaging techniques (MRI, ultrasound) [68, 69] to identify morphological changes not easily detected by traditional methods.

Laboratory data analysis, including markers like amylase and lipase, also shows promise in identifying patients at risk for pancreatitis and its complications [70]. Moreover, AI enables accurate monitoring of recurrent pancreatitis patients and early pancreatic cancer detection [71, 72], as recurrent pancreatitis significantly increases cancer risk [70, 73]. AI-based predictive models can stratify patients for timely, targeted intervention [74].

Early diagnosis of acute kidney injury (AKI) is essential, enabling immediate intervention and better outcomes [75]. AI uses clinical and biometric variables—including urine output and serum creatinine levels—to develop predictive models. Studies suggest AI can detect abnormalities 48 h before clinical signs of AKI appear [76, 77], particularly helpful in ICU patients, who are at high risk [78].

AI also analyzes emerging renal biomarkers like KIM-1, NGAL, and IL-18, which offer earlier and more sensitive indicators of kidney damage than creatinine [79]. As KIM-1 and NGAL increase in early AKI, AI can automatically process these markers [80–82].

In imaging-based AKI diagnostics, ultrasound evaluates renal morphology changes due to ischemia or injury [83]. This multimodal approach integrates AI’s predictive capacity with traditional verification, improving diagnostic accuracy and creating more reliable clinical flowcharts [84].

AI also benefits acute abdominal diagnostics in emergency settings. It combines large datasets with vital signs, physical findings, and lab results, enhancing diagnostic accuracy and reducing time to appropriate diagnosis—vital in conditions like acute appendicitis or cholecystitis [85, 86]. AI also interprets imaging such as X-rays, critical for acute abdominal evaluation.

Conclusions

While this systematic review highlights AI’s promising applications in emergency medicine, methodological disparities, small sample sizes, and study result variability may limit generalizability. Future research should focus on large-scale, multicenter studies using standardized metrics to enable comparative analyses across clinical settings.

Additionally, addressing regulatory challenges, ethical considerations, and potential algorithmic bias is crucial for responsible AI advancement in healthcare. Researchers and stakeholders must jointly develop strategies to ensure equitable access and ethical deployment of AI diagnostics, especially for marginalized populations.

In conclusion, applying AI and ML technologies in emergency departments has the potential to transform early disease diagnosis, improve patient outcomes, and enhance the efficiency of emergency care delivery. This systematic review provides key insights and paves the way for future studies aimed at optimizing AI integration into clinical practice. Continued progress in AI technologies promises further innovations in emergency medicine [87].

This expanded systematic review summarizes both the potential and the challenges of integrating AI into emergency department diagnostics, emphasizing its transformative role in improving emergency care delivery. Further exploration and validation of these technologies are essential for seamless integration into clinical practice, ultimately aiming to improve patient health outcomes.

Authors’ contributions

NDF: conceptualization; YL, CZ: data curation; GB AV GV: writing; RB RR: formal analysis, investigation; SM SC: formal analysis; PB AM: investigation; PR: methodology; PF: review and editing; RLR: review and editing.

Funding

The author(s) declare that no financial support was received for the research, authorship, and/or publication of this article.

Data availability

Not applicable.

Declarations

Competing interests

The authors declare no competing interests.