Role of artificial intelligence and point of care ultrasound in management of critically ill patients

Abstract

Point of care ultrasound (POCUS) has become an invaluable tool in the management of critically ill patients, offering real-time diagnostic insights into cardiovascular, pulmonary, and abdominal pathophysiology. In the critical care setting, timely diagnosis is essential to manage life-threatening conditions. The integration of artificial intelligence (AI) into POCUS has been a transformative technological advancement in the field. AI-incorporated POCUS can help clinicians of varying experience levels overcome limitations associated with operator dependency and varied image quality. These innovations are valuable, not only in resource-constrained settings, but also during time-sensitive clinical scenarios, profoundly impacting patient outcomes. AI-driven platforms can provide prompt feedback of protocolized exams such as rapid ultrasound in shock and bedside lung ultrasound in emergency. This is especially relevant in situations, where basic imaging such as transthoracic echocardiography is often performed by non-specialized personnel. Most AI tools remain investigational, and the need for robust validation of machine learning in clinical workflows remains a burning question. Despite a promising role in simulation, the effectiveness of these tools in real-world clinical scenarios depends heavily on the quality of training datasets. The integration of AI and POCUS, while reducing diagnostic errors, is also revolutionizing diagnostics with deep learning models, which demonstrate high accuracy. Beyond improving diagnostic precision, AI is optimizing workflows, reducing processing times, and enabling real-time interpretation. Future advancements in AI, integrating imaging with clinical data and predictive modeling, have the potential to significantly enhance prognostic accuracy and patient outcomes, particularly in critical care settings. This narrative review aims to explore the current applications, advancements, and future directions of AI-assisted POCUS in the intensive care unit, with a particular focus on machine learning in critically ill patients.

Key Words: Artificial intelligence; Point of care ultrasound; Intensive care unit; Automation; Machine learning

Core Tip: By synthesizing the latest evidence, our article provides critical care practitioners and researchers with a comprehensive overview of how artificial intelligence-enhanced point of care ultrasound can transform bedside decision-making, improve patient outcomes, and expand access to high-quality care worldwide.

INTRODUCTION

Point of care ultrasound (POCUS) has revolutionized critical care medicine, offering rapid, bedside diagnostics that enhance clinical decision in real time[1]. Ultrasound machines for POCUS are now readily available in critical care units. The applications of POCUS are both diagnostic and therapeutic, allowing physicians to manage life-threatening conditions. Moreover, POCUS is a standard tool for rapid assessment of the cardiac function, the volume status, and the causes of shock, with demonstrated impact on diagnosis and management in intensive care unit (ICU) settings[2]. Despite its proven utility, the widespread adoption of POCUS in the ICU is hampered by challenges such as steep learning curves and accurate image interpretation. Thus, the clinical effectiveness of POCUS in critical care remains operator-dependent and highly variable.

Artificial intelligence (AI) has emerged as a powerful tool in medicine over the last two decades, particularly in diagnostic imaging starting from era of the LeNet-5 architecture in 1989[3]. The integration of different AI applications with POCUS has drastically improved real-time image quality assessment, automated diagnostics, and individualized user guidance[4]. Nevertheless, performance of AI integrated POCUS may degrade in the presence of poor-quality images or with severity of illnesses. Ultrasound has unique challenges for AI integration since it gathers real-time images and increases the variability in training datasets based on operator dependency[5]. Despite challenges, AI-POCUS has expanded access to timely, high quality critical care imaging, bridging gaps in expertise and optimizing bedside diagnostics.

As the intensive care management becomes increasingly complex and time sensitive, AI-POCUS will emerge as a powerful innovation with the potential to redefine rapid bedside diagnostics and a tool for novice users. This review explores the role of AI and POCUS in critical care, highlighting advances in technology, current usage, and the growing evidence and future directions supporting their use.

EVOLUTION OF POCUS IN CRITICAL CARE

The rapid evolution of POCUS has transformed it from a procedural adjunct into a frontline diagnostic modality. Landmark work by Lichtenstein demonstrated its utility in evaluating pneumothorax, atelectasis, and interstitial syndromes, proving against the misconception that lungs were “invisible” to ultrasound[6,7]. Subsequently, researchers suggested that POCUS may offer potential advantage in reducing inter-operator variability and shorten learning curves, especially for novice learners and in resource limited settings.

Considering the acuity of ICU settings, the adoption of cardiac POCUS was driven by the need for immediate hemodynamic assessment in cases of shock, cardiac arrest, or acute myocardial infarction, which is now recommended by the American Society of Echocardiography and the Society of Critical Care Medicine[8,9]. Cardiac POCUS can improve critical care outcomes by enhancing bedside evaluation of ventricular function, preload, pericardial effusions, and valvular abnormalities[10]. Lu et al[11] reported that cardiac bedside POCUS led to immediate changes in clinical management or interventions in at least 75% of cases. Pulmonary POCUS has also become useful to diagnose acute respiratory failure. One parallel-group trial showed that pulmonary POCUS increased the accuracy of a presumptive diagnoses of acute dyspnea within 4 hours of admission (88% vs 64%, P < 0.0001)[12]. Findings like B-lines (i.e., vertical artefacts seen in lung ultrasound), consolidations, pleural effusions, and pneumothorax, pulmonary POCUS allows clinicians with real-time insights into the etiology of respiratory compromise, outperforming chest radiography in both sensitivity and specificity[13]. Finally, critical care clinicians can employ the Focused Assessment with Sonography in Trauma (FAST) protocol to evaluate diagnose life-threatening conditions for unexplained hypotension. This eliminates the delays and risks of transferring unstable patients for more advanced imaging or relying only on clinical signs and symptoms[14]. Altogether, POCUS has evolved into a powerful tool for comprehensive hemodynamic and respiratory assessment, thereby expediting clinical decisions.

ROLE OF AI IN POCUS

Broadly defined, AI are computational systems capable of performing tasks that blend human intelligence, such as recognition, prediction, and decision making. In healthcare, AI applications are largely driven by machine learning (ML), algorithms that improve predictions from data, and deep learning (DL), which uses multilayered neural networks to model complex, non-linear relationships. Several architectures exist under DL, including artificial neural networks for general data modeling, recurrent neural networks for sequential or time-series inputs such as ECG or physiologic monitoring, and convolutional neural networks (CNNs), which are uniquely suited for image analysis[15]. CNNs apply multiple layers of filters to automatically extract and recognize visual features, such as edges, textures, and patterns, allowing them to detect subtle abnormalities that are imperceptible to the human eye. CNN’s capacity to recognize complex image characteristics has revolutionized critical care imaging, especially in echocardiography and lung ultrasound, where rapid, high-dimensional data interpretation is crucial[3,16]. CNN has advanced medical imaging by detecting subtle visual features and recognizing patterns with remarkable accuracy in a short time[17]. Recent studies have increasingly demonstrated their clinical utility, particularly in image guided analyses and through predictive modeling. Among the most notable applications are echocardiography and lung imaging, where CNN driven interpretation shows significant promise in ICUs and emergency departments[17].

Early adaptation of AI in POCUS

Initial studies explored the feasibility of adapting existing AI architectures for POCUS applications. Boice et al[18] employed ShrapML, a CNN originally designed for shrapnel detection, to identify pneumothorax using phantom and swine models. The model achieved only 50% accuracy without image augmentation, which improved to 90%-97% following augmentation and model refinement. Combining image augmentation, ShrapML, and a more realistic phantom, Hernandez-Torres et al[19] then showed a 90%-97% accuracy in diagnosing pathologies in E-FAST scan. While encouraging technical feasibility, these studies were limited to preclinical models, denoting the need for human validation before clinical implementation.

Pulmonary applications

In pulmonary imaging, AI-POCUS has shown comparable accuracy to chest radiography for pneumothorax diagnosis while offering superior efficiency[20]. In pediatric patients with pneumonia, DL analysis of chest ultrasounds demonstrated efficacy in diagnosing consolidation [sensitivity of 88%, specificity of 89%, positive predictive value of 89%, and negative predictive value (NPV) of 87%][21]. Notably, AI developments in bedside lung ultrasound in ICU has gained significant improvements, where recent prospective studies demonstrate DL models that can achieve up to 95% accuracy in distinguishing normal (A-line) from abnormal (B-line) patterns, comparable to expert interpretation[22]. Such findings highlight AI’s ability to augment bedside diagnostic accuracy and standardize interpretation across operators.

AI-assisted lung ultrasound facilitates standardized scoring and monitoring of disease progression, as seen in coronavirus disease 2019 pneumonia and other acute respiratory disorders[23]. Subsequently, Roy et al[24], in a multicenter study demonstrated a DL (spatial transformer network model) that could identify the degree of disease severity at a frame-level, video-level, and pixel-level using abnormal artefacts such as B-lines in patients with lung pathologies. Despite this, AI-assisted POCUS is limited by issues of generalizability and the need for optimal acoustic windows, for which extensive validation continues to be necessary.

Cardiac applications

The integration of AI in cardiac POCUS has focused on automating quantitative measurements such as left ventricular ejection fraction (LVEF), chamber dimensions, as well as the assessment of wall motion abnormalities. When used by novice operators in the ICU, inter-rater reliability for detecting LVEF was comparable to expert sonographers (intraclass correlated coefficient of 0.88-0.94) and high specificity, especially if its < 40%[25].

Training and workflow optimization

One of the most promising uses of AI-POCUS is in education and procedural guidance. This approach has shown measurable improvements in diagnostic accuracy and speed. Modern literature has shown that AI-POCUS can be used to help bridge talent, competence, and resource gaps in resource poor medical centers. Nhat et al[27] showed that that AI-POCUS improved the accuracy of novice users from 68.9% (95%CI: 65.6%-73.9%) to 82.9% (95%CI: 79.1%-86.7%) (P < 0.001). In a randomized controlled trial, Baum et al[28] showed that the AI group had faster scan times of A4C views [57 seconds, interquartile range (IQR): 32-75 seconds vs 85 seconds, IQR: 50-172 seconds; P = 0.01], higher image quality scores [4.5 (IQR: 2-5.5) vs 2 (IQR: 1-3); P < 0.01], and more accurately identified reduced systolic function (85% vs 50%; P = 0.02) when compared to the non-AI group. Moreover, some obstacles for widespread adaptation of AI-POCUS include the steep learning curve for novice users for ultrasound and the subjective interpretation of results, to which software solutions have adapted by providing procedural instruction and/or guidance. For instance, iScanHelper (Mindray Inc, France) can evaluate an ultrasound image in real-time and educate learners on how to modify their technique to optimize image acquisition. This technology allows for more accurate and precise exams while training the user[26].

APPLICATIONS OF AI-ASSISTED POCUS IN ICU

AI POCUS is emerging as a transformative bedside tool for critically ill patients, offering automated image acquisition, quantitative analysis, and real-time interpretation. Among all organ systems, cardiac applications have been most extensively explored, given the central role of hemodynamic assessment in ICU management. Rapid and accurate evaluation of cardiac function and fluid status is essential for guiding therapy and predicting outcomes, and AI integration has shown strong potential to enhance both the accuracy and efficiency of POCUS-based assessments.

Cardiac POCUS in ICU

LVEF: AI-POCUS has demonstrated reliable performance in quantifying key cardiac parameters such as LVEF and left ventricular outflow tract velocity-time integral, which are crucial for assessing stroke volume and cardiac output[29]. Gallant et al[25] showed that regardless of expertise, the sensitivity for measuring LVEF still remained low (56%-70%) in acute care settings, mainly by using parasternal long axis and A4C views (95%CI: -0.10 to 1.0). Their study did not incorporate other practical bedside parameters such as E-point septal separation, which is a well-recognized and reliable adjunct for rapid LVEF assessment in critically ill patients[30]. In a multicenter study of healthy patients (n = 200), Kagiyama et al[31] showed that intraclass correlated coefficient between conventional LVEF and by AI-POCUS (0.81) (P = 0.008) and very low mean bias -1.5% and a higher sensitivity 85% (95%CI: 76%-91%) by using left ventricular end diastolic volumes as a primary parameter. The American Heart Association also stresses that AI in echocardiography now supports automated chamber quantification, valvular assessment, disease detection, and workflow optimization, with improved cardiac imaging aspects[32].

Considering the automated advancements in cardiac POCUS, Zhang et al[33] developed a CNN capable of accurately identifying 23 cardiac views and segmenting chambers across five standard positions, showing 96% accuracy, and identifying partially obscured chambers. Automated calculation of LVEF was extensively studied by Asch et al[34] (n = 55000 echo images), where authors showed that auto-LVEF values showed high consistency (mean absolute deviation = 2.9%) and consistent agreement (r = 0.95, bias = 1.0%), with > 90% sensitivity and specificity for detecting LVEF of ≤ 35%. Motazedian et al[35] also reinforced this idea and showed that LVEF can be reproducible by AI, regardless of the severity, with an area under the curve (AUC) of 0.98 for identifying an abnormal LVEF. For a threshold of less than 50%, the NPV of 97%, and AUC of 0.99 and NPV of 98% in identifying severe dysfunction of < 30%[35]. In addition, Malins et al[36] showed that a DL model, using single frames from multiple videos for LVEF estimation with a ResNet 2D CNN backbone, achieved an AUC > 0.90 in distinguishing between LVEF ≤ 40% and LVEF > 40%.

Inferior vena cava assessment: Beyond developing models trained for identifying LV function, researchers have expanded AI-POCUS to assess volume status and right heart parameters and focused on inferior vena cava (IVC) assessment, right atrial pressure as a measure of volume status. IVC diameter is an important hemodynamic measurement in which AI-POCUS has shown promise and by using a novel active circle algorithm, precise measurements of IVC could be recorded, leading to a better understanding of fluid status and right heart function[37]. One study by Zamzmi et al[38] showed that a lightweight-DL could distinguish IVC from other echocardiographic views including artefacts and different morphology of IVC, with an accuracy of 0.97, a precision 0.96, a sensitivity of 0.97, and harmonic mean of precision (f-1) of 0.96. Blaivas et al[39], a major contributor to automated models in AI-POCUS, studied image interpretation among 220 subjects using a long short-term memory, a type of DL model and found excellent agreement between the algorithm and POCUS experts (κ = 0.45, 95%CI: 0.33-0.56). Additionally, Yildiz Potter et al[40] studied the application of the YoloV3 algorithm for diagnosing pericardial effusion, achieving 89% sensitivity and 92% specificity, enabling faster diagnosis. The evidence illustrates how AI-POCUS can reliably quantify cardiac function, assess volume status, and detect structural abnormalities at the bedside. While LVEF, IVC, and right atrial pressure are now well-characterized, there remains significant opportunity to extend automation to other critical parameters, such as tricuspid annular plane systolic excursion and mitral annular plane systolic excursion, for a more comprehensive Focused Assessed Transthoracic Echocardiography evaluation.

Outcomes prediction: Recently studies have demonstrated that ultrasound ML multi-models can enhance the prediction of critical outcomes like death, cardiac arrest, or cardiogenic shock in ICU patients, based on transesophageal echocardiography. Using parameters such as left ventricular outflow tract and velocity-time integral, systolic pulmonary artery pressure, tricuspid annular plane systolic excursion, and LVEF, ML algorithms have outperformed traditional methods in prognostic value, achieving a concordance index of 0.80 compared to 0.73 (P = 0.012)[41]. Supporting this, Laumer et al[42] demonstrated a specific DL algorithm trained to detect subtle differences in cardiac wall motion abnormalities, (i.e., distinguishing between acute myocardial infarction and stress cardiomyopathy), which surpassed cardiologists, where the model achieved a mean AUC of 0.79 and an overall accuracy of 74.8%, compared to a mean AUC of 0.71 and accuracy of 64.4% for cardiologists. Blavias, a pioneer in DL algorithms, used a simplified DL (VGG16) for visually estimating LVEF, used Stanford’s Echonet dynamic database and reported a mean absolute error of 8.08% (95%CI: 7.60%-8.55%), indicating strong performance compared to skilled clinicians, whose root mean square error was 11.98% with a correlation of 0.348[43]. These findings suggest that, when properly trained, DL and ML models can match or even surpass clinicians in certain diagnostic capabilities, reinforcing their potential in critical care decision making.

Pulmonary POCUS in ICU

AI reliably identifies pulmonary B-lines, a characteristic ultrasonographic feature of pulmonary edema[44]. Zech et al[45] showed that CNNs can learn site-specific biases rather than true disease features, achieving an AUC of 0.861 (95%CI: 0.855-0.866) when trained on cohorts stratified with a 10-fold difference in pneumonia prevalence. This shows the need for external validation of disease prevalence to ensure reliable AI deployment across institutions. Moore et al[46] developed an automated model for detecting and quantification of B lines using handheld ultrasound with interrelated correlation between the manual and reviewer-adjusted counts of 0.94 (95%CI: 0.90-0.96). Baloescu et al[47], in an effort to establish a foundation for standardizing AI data screening and analytics, developed a DL model with a sensitivity of 93% (95%CI: 81%-98%) for identifying B-lines. Their model also demonstrated a high level of agreement with expert interpretations, achieving a kappa of 0.88 (95%CI: 0.79-0.97). Similarly, Ienghong et al[48] demonstrated that automated detection of B-lines was associated with a 12.9% increase in sensitivity (95%CI: 0.92-0.98) and a 14% improvement in specificity (95%CI: 0.74-0.80) for diagnosing pulmonary edema compared to manual readings. While most AI research focused on B-line detection, future studies should also emphasize additional diagnostic features, including lung sliding and pleural or parenchymal fluid accumulation, which can be trained using tissue density-based DL models.

Abdominal and musculoskeletal POCUS in ICU

In abdominal FAST, AI-POCUS has shown strong diagnostic performance in detecting and localizing abdominal free fluid[49]. DL models such as YOLOv3, U-Net, ResNet50 (Residual Network), DenseNet121, InceptionV3, and Vgg11bn have been trained and validated in this area, with sensitivities of 88%-98% and specificities of 68%-100% in internal validation cohorts[49,50]. While these algorithms can provide real-time recognition of hemoperitoneum, Leo et al[51], showed that YOLOv3 (n = 94) demonstrated superior image-processing speeds (57 milliseconds) using right upper quadrant imaging that make them suitable for bedside use, (97% AUC over the controls). Similarly, Cheng et al[52] showed that ResNet V50 DL has 96% accuracy in diagnosing free fluid in Morrisons pouch. One of the advantages of AI-POCUS is its application in critical care musculoskeletal scanning. Prolonged ICU stays often lead to muscle wasting, complicating patients’ recovery and return to baseline after discharge. AI-POCUS has proven to be an efficient method for assessing muscle wasting compared to traditional manual tracing. The average time required for assessment was significantly reduced from 19.6 to 9.4 minutes (IQR: 7.2-11.7), compared to the 16.9-21.7-minute range with manual methods[53].

As AI-assisted POCUS improves, models may be better at accommodating anatomical variations, changes in the breathing pattern variation, or unique pathology can lead to improper interpretation by the AI algorithms[54]. Given the high dependency of AI on large datasets, conclusions from studies with the sample sizes described above should be considered preliminary. Moreover, current models lack the clinical context that experienced clinicians apply, risking errors when training a DL with extensive data do not account for complex scenarios. These gaps signify the need for further refinement of ML and DL models to improve accuracy and context-sensitive diagnosis.

CHALLENGES IN CLINICAL IMPLEMENTATION

While AI-POCUS has been extensively discussed, Park[5] highlighted several key obstacles in employing AI-POCUS, which includes operator dependency, inconsistent outcomes, and the importance of prospective studies that involve direct interactions between human examiners and AI models, rather than relying solely on retrospective image analysis. Studies have addressed potential limitations and compared traditional methods with AI-assisted techniques, with the goal of establishing a more effective and coherent diagnostic approach. For instance, Filipiak-Strzecka et al[55] found no significant difference in echocardiogram measurements between AI-integrated software (LVIvo) and handheld ultrasound (mean difference: -0.61%, 95%CI: -1.89 to 0.68, P = 0.31). The study also revealed that in over 50% of patients (n = 63), the AI model could not reproduce the LVEF measurement, emphasizing the need for refinement AI-assisted diagnostics. This reality is often overlooked, which needs further clarification in future studies.

The reliability of AI models largely depends on image quality. A meta-analysis by Kuo et al[56] showed no significant difference in diagnostic accuracy between clinicians and AI for emergency radiograph interpretations. In external validation tests, AI showed a pooled sensitivity of 91% (95%CI: 84%-95%), compared to 94% (95%CI: 90%-96%) for clinicians, whereas the pooled specificity was 91% (95%CI: 81%-95%) for AI and 94% (95%CI: 91%-95%) for clinicians. Blaivas et al’s extensive work[57] on DL model included a “do-it-yourself” approach for classifying POCUS images across various anatomical areas (pelvis, heart, lung, abdomen, musculoskeletal, ocular, and central vascular access) and improved quality workflows for POCUS protocols. Blaivas et al[57] assessing DL algorithms with 21 videos from two novel POCUS machines, performance was notably lower than that of a more commonly used POCUS system. However, Della Ripa et al’s mixed-method study[58] raised concerns regarding AI-POCUS in the obstetric population undergoing antenatal care, revealing that AI use led to a 15% reduction in provider confidence in diagnosis, citing concerns over algorithmic accuracy and the potential for diminished clinical judgment due to over-reliance on AI. Regulatory, ethical, and data privacy considerations further complicate AI integration in healthcare, as the use of large datasets for training and validation must comply with current and evolving standards for patient confidentiality and AI usage transparency[59].

DL models may often underperform in real world settings compared to the training environment due to a lack of generalizability. In one meta-analysis of 77 studies, Kelly et al[60] showed that model performance decreased (mean 6%) when externally validated, whilst having better accuracy (AUC of 0.903). Disease severity can also introduce bias. Training a model on imaging data that represent the extremes of a disease spectrum may have limited generalizability in mild or moderate forms of the disease[61]. Errors in labeling training data can also lead to inaccuracies, as small measurement mistakes may cause significant misclassifications. Ultimately, the lack of standardization in developing AI algorithms makes them prone to bias and variability across algorithms[62].

Another challenge in implementing AI in POCUS is the resource-intensive nature of AI infrastructures[56]. Moreover, while AI can automate measurements and aid interpretation, it may underperform in complex cases and cannot replace expert judgment. Prospective clinical trials are essential to validate these tools before widespread adoption in ICU and non-ICU settings.

PREDICTIVE MODELING AND FUTURE DIRECTION

AI-assisted POCUS is a rapidly evolving field, with the most recent advancement being diagnostic and predictive tasks related to outcomes. Recently, Holste et al[63] developed PanEcho, a DL model, across 18 diagnostic tasks and estimated 21 parameters, using 1.2 million echocardiographic videos from 24405 patients, that achieved a median AUC of 0.91. The model reliably identified major pathologies including LVEF (AUC: 0.98-0.99; mean absolute error: 4.2%-4.5%), right ventricular systolic dysfunction (AUC: 0.93-0.94), and severe aortic stenosis (AUC: 0.98-1.00). Evolution of ML in imaging extends beyond diagnostic accuracy to optimizing workflows, including scheduling image acquisition, reducing processing times, and enhancing automated interpretation. To standardize research and reporting, Weikert et al[64] proposed consensus guidelines using an 11-point framework that emphasizes critical aspects such as defining the research question, selecting appropriate ML models, ensuring adequate sample size and study design, specifying training, validation, and test datasets, and transparently reporting data labeling reliability.

Furthermore, robotic AI-POCUS also have been under study and has been trained to perform extensive cardiology imaging and analysis. Lin et al[65] proposed an automated robotic-arm ultrasound scanning system, which has the potential to reduce operator-dependent variability. While still experimental, such systems could be transformative in critical care by enabling faster, more reliable image acquisition. In parallel, DL based real-time interpretation is advancing, Dave et al[22] validated a DL model for automated B-line detection in lung ultrasound, achieving performance equivalent to expert interpretation (accuracy of 95%, sensitivity of 93%, specificity of 96%).

A particularly exciting direction is the development of multi-models, which integrate imaging with electronic medical record (EMR) data to improve prognostic accuracy. For example, Khader et al[66] showed that combining chest radiographs with EMR data significantly improved survival prediction (area under the receiver operating characteristic of 0.863) compared to imaging or clinical data alone. Similarly, Lin et al[67] demonstrated that multimodal models integrating chest X-rays with clinical variables (PrismICU) achieved superior predictive performance (AUC of 0.95, F1 score = 0.95) compared to single-modality approaches. Translating this to POCUS, multi-integrated AI models could greatly enhance prognostication in ICU patients if combined with ultrasound-derived parameters and clinical data (Table 1)[20-23,25,27,35,38,40,47,48,51-53,55,56,63,66,67].

Table 1 Demonstrating the different artificial intelligence integrated point of care ultrasound models employed in intensive care units.

Ref.	Software utilized	Parameter studied	Result
Yang et al[20], 2024	Intelligent lung ultrasound	Diagnosis of pneumothorax in critically ill patients using intelligent lung ultrasound vs CXR	AI-ultrasound: Sensitivity of 79.4%, specificity of 85.4%. CXR: Sensitivity of 82.4%, specificity of 80.5%
Kessler et al[21], 2024	AI algorithm for detecting features of pulmonary consolidation via POCUS	Identification and localization of lung consolidation using previously confirmed positive and negative images	Overall accuracy of AI algorithm: 88.5%, sensitivity of 88%, specificity of 89%, positive predictive value of 89%, NPV of 87%
Gallant et al[25], 2025	ExoAI LVEF analysis package	LVEF measured by novices with AI assistance vs expert sonographers	Inter-rater reliability between experts and novices was very high (intraclass correlated coefficient of 0.88-0.94)
Dave et al[22], 2023	Previously trained LUS DL model	Comparing real time DL classification to expert classification of normal (A-line pattern) and abnormal (B-line pattern) lung patterns	Real time DL model demonstrated 95% accuracy, 93% sensitivity, and 96% specificity for identification of the B-line pattern
Mento et al[23], 2021	AI-assisted ultrasound DL model	Stratification of patients to low risk or high risk groups for disease progression	High agreement (85.96%) between lung ultrasound interpretation by experts (gold standard), and DL model
Nhat et al[27], 2023	Bespoke AI tool for LUS interpretation	Accuracy of non-clinical examiners before and after AI assistance	Performance improved from 68.9% (95%CI: 65.6%-73.9%) to 82.9% (95%CI: 79.1%-86.7%) (P < 0.001) after use of AI tool
Motazedian et al[35], 2023	FoCUS with AI	FoCUS vs transthoracic echocardiography measurement of LVEF	AUC of 0.98 for abnormal LVEF (< 50%), NPV of 0.97; AUC of 0.99 for severe dysfunction (< 30%)
Zamzmi et al[38], 2023	Multi-stage AI system that utilizes lightweight and open-world machine learning architecture	Automatically computed vs expert calculation of right atrial pressure from inferior vena cava measurement	Excellent agreement (P < 0.01) between groups. Macro accuracy = 0.85
Yildiz Potter et al[40], 2023	YoloV3 object detection algorithm	Pericardial effusion detection in scans confirmed by expert’s vs negative controls	Algorithm had 92% specificity and 89% sensitivity in pericardial effusion identification
Baloescu et al[47], 2020	DL algorithm based on deep convolutional neural networks	Identification of B-lines	DL model had a sensitivity of 93% (95%CI: 81%-98%) for identifying B-lines. High level of agreement with expert interpretations, achieving a kappa of 0.88 (95%CI: 0.79-0.97)
Ienghong et al[48], 2025	Auto B-lines	Detection of pulmonary edema	AI usage resulted in 12.9% increase in sensitivity (95%CI: 0.92-0.98) and a 14% improvement in specificity (95%CI: 0.74-0.80) for diagnosing pulmonary edema compared to manual readings
Leo et al[51], 2023	YoloV3 object detection algorithm	Identification of the presence and location of hemoperitoneum on POCUS by algorithm vs expert	Algorithm had 95% sensitivity, 94% specificity, 95% accuracy, and 97% AUC compared to gold standard expert diagnosis
Cheng et al[52], 2021	CNN based ResNet50-V2	Free-fluid detection in Focused Assessment with Sonography in Trauma examinations	AI model correctly classified positive and negative images with an accuracy of 0.941
Nhat et al[53], 2024	AI tool measuring rectus femoris cross-sectional area	Time spent on scanning	Scanning time reduced from 19.6 minutes (IQR: 16.9-21.7) to 9.4 minutes (IQR: 7.2-11.7) when using AI tool (P < 0.001)
Filipiak-Strzecka et al[55], 2021	LVivo	LVEF	No significant difference between handheld ultrasound and LVivo AI software measurements between (mean difference: -0.61%, 95%CI: -1.89 to 0.68, P = 0.31)
Kuo et al[56], 2022	Meta-analysis with a hierarchical model to calculate pooled sensitivity and specificity was used	Fracture detection: AI model vs clinician	Pooled sensitivity: 92% (95%CI: 88%-93%) for AI and 91% (95%CI: 85%-95%) for clinicians. Pooled specificity: 91% (95%CI: 88%-93%) for AI and 92% (95%CI: 89%-92%) for clinicians. No significant difference found between clinician and AI performance
Holste et al[63], 2025	PanEcho	Successful detection of a variety of cardiac pathology	Successful detection of the following: LVEF (AUC: 0.98-0.99; mean absolute error: 4.2%-4.5%), right ventricular systolic dysfunction (AUC: 0.93-0.94), and severe aortic stenosis (AUC: 0.98-1.00)
Khader et al[66], 2023	Combined medical transformer for multimodal survival prediction	Intensive care unit Survival	CXR model: (AUROC = 0.811), clinical data model (AUROC = 0.785), combined model (AUROC = 0.863, P < 0.001)
Lin et al[67], 2024	PrismICU	Accurate prediction of 30-day mortality	Multimodal approach (AUC = 0.95, F1 score = 0.95) was more accurate than CXR ((AUC = 0.71, F1 score = 0.36) and clinical parameters (AUC = 0.72, F1 score = 0.50) individually

AI: Artificial intelligence; AUC: Area under the curve; AUROC: Area under the receiver operating characteristic; CXR: Chest radiography; DL: Deep learning; FoCUS: Focused cardiac ultrasound; IQR: Interquartile range; LUS: Lung ultrasound; LVEF: Left ventricular ejection fraction; NPV: Negative predictive value; POCUS: Point of care ultrasound.

The growing suitability of ML in the ICU is evident from increasing research, supported by publicly available databases such as MIMIC, eICU, and AmsterdamUMCdb, which have enabled advances in predictive modeling and clinical applications[68]. However, differences in data variety, frequency, and clinical practices across multicenter databases require careful selection and tailored for specific research questions[69]. Despite rapid progress, the field of AI-assisted POCUS still requires robust evaluation to confirm whether technological advances translate into measurable improvements in patient care and outcomes. McCague et al[70] emphasized the importance of involving frontline healthcare professionals in the development and testing of these tools to ensure clinical relevance and usability. Future studies should adopt specific endpoints and follow standardized frameworks such as the CLAIM 2024 criteria, which provide transparent guidance across the research process[71]. Such approaches will enhance the applicability of AI-assisted imaging in clinical care.

CONCLUSION

AI-POCUS is not intended to replace comprehensive echocardiography, lung ultrasound, or formal FAST examinations, but rather to serve as a powerful adjunct, particularly when expert interpretation is limited. While applications of CNN and DL remain largely within the research domain, access to large and diverse datasets, and validation through prospective studies, ML and DL can be incorporated into routine critical care practice. The future direction is multi-modal AI, integrating POCUS with EMR and physiologic data to enable predictive, context-aware diagnostics, which could transform bedside care into a data driven, adaptive process, bridging the gap between automation and clinical expertise.