AF diagnosis with EKG is one of the most widely studied implementations of AI in AIS/neurocardiology diagnosis and identification to stroke risk stratification and treatment options (Figure 3). This includes several FDA-approved programs such as Cardiomatics and AccurKardia (89, 90). Certain AI/ML models have been shown to outperform validated tools such as CHA₂DS₂-VASc in stroke prediction (91, 92).

While normal sinus rhythm on initial EKG may point clinicians away from a diagnosis of AF, Attia et al. recently used over 20 years of EKG data from the Mayo Clinic to assess the ability of AI to identify AF from a normal sinus rhythm on EKG (93). This study demonstrated area under the curve (AUC) of 0.87 for correctly diagnosing AF among patients who had previously recorded AF rhythms. However, this AUC increased to 0.90 when predicting AF in undiagnosed patients, notably if EKG was obtained up to 31 days prior to the first EKG-recorded AF episode (93). This AI apparatus was then compared to the validated CHARGE-AF score for AF risk prediction and performed with a similar degree of accuracy (94). These measures have also been used in the assessment and diagnosis of AF in the setting of cryptogenic stroke (95). Additionally, Weil et al. investigated the predictive value of an AI-generated EKG score of AF risk, noting that a higher AI-EKG-AF score correlated with the subsequent development of cerebral infarcts on MRI, therefore demonstrating the value of such AI-algorithms in predicting cerebrovascular disease even with EKG data alone (96).

Based on these findings, Raghunath et al. subsequently developed their own deep neural network (DNN), while also including a 1-year predictive factor for future-onset AF for a sample of 430,000 patients over 30 years. This DNN demonstrated a final AUC of 0.85 and a sensitivity of 69% in a simulated deployment scenario (97). The authors analyzed their study cohort and noted that 62% of patients with predicted AF subsequently suffered from stroke within 5 years of the study, a greater accuracy than currently approved protocols (97). Kim et al. also followed a similar 12-lead EKG model mainly using ResNet blocks, but also incorporated an attention analysis for the model using Grad-CAM with convoluted neural network (CNN) feature map gradients (98). The authors noted that sinus rhythm EKGs with prolonged QRS compared to a specific point between the T and P waves correlated with an AI-diagnosis of AF (98).

Most of these studies also found that monitoring close to an AF event provided better results and eliminated the indication for long-term monitoring in this population (97, 98). New models are continually being created to improve diagnostic efficacy and risk prediction (99, 100).

Photoplethysmography (PPGs) are light-based sensors for blood flow measurement utilized both in the inpatient and outpatient settings in wearable devices such as smartwatches. The WATCH AF trial was among the first studies demonstrating the feasibility of AF diagnosis and observation in PPG-enabled devices, citing an accuracy of 96.1% (118). A recent systematic literature review found 24 cohort studies mostly assessing AF detected by smartwatches showing its significant growth and popularity in recent years, with several studies pointing to the easy implantation of these tools (119–122). PPG technology may also be used as a mobile phone application, such as with FibriCheck and SMARTBEATS, applications which have yielded high diagnostic accuracy, with sensitivity and specificity rates of >95% when compared to the diagnostic ability of single-lead EKGs (123–125).

AI-enabled electrocardiogram models have been developed, utilizing CNNs to assist in the early detection of significant aortic stenosis (AS) (127). Using a sample of over 258,000 of EKGs to train, validate and test the model, researchers at Mayo Clinic, Rochester developed an AI-EKG system capable of predicting moderate to severe AS with impressive sensitivity, specificity and accuracy (127).

Since echocardiography is the primary modality for diagnosis and evaluation of VHD, the primary focus of AI research in valvular heart disease remains on the review and interpretation of echocardiograms. AI can assist in the segmentation of valves and cardiac structures for automated analysis. Using image recognition algorithms, aortic and mitral valve disease states have been directly detected from the images themselves. Several AI-enabled algorithms have been developed to diagnosis mitral valvular disease via quantitative analysis of 3-dimensional (3-D) echocardiography. These include Mitral Valve Quantification (Philips Medical Imaging, Andover, MA), eSie Valve (Siemens Healthcare, Mountain View, CA, USA), Mitral Valve Navigator (Philips Medical Systems), Auto Valve Analysis (Siemens; California, USA), and eSie PISA Volume Analysis (Siemens Medical Solutions USA, Inc., Mountain View, CA) (128). Measurements obtained during echocardiographic valvular assessment have been integrated with other clinical data to identify novel aortic valve disease subgroups and describe new predictors of aortic valve disease progression (129).

Similarly, another FDA-approved AI-apparatus is iCardio.ai, which has been cleared for use in the detection of AS This apparatus was developed through training and testing CNNs to use two-dimensional (2D) TTE to diagnose AS, with an AUC of 0.96 (130). Since severe aortic stenosis is a known risk factor for ischemic stroke, screening of at-risk populations with timely and efficient AI-enabled tools can be valuable.

AI applications can provide a more rapid analysis of cardiac MRI data, automated and quantitative evaluations of the valvular structures, and a proper appraisal of patient prognosis in the setting of disease (131, 132). This information is essential in gauging disease severity and stroke risk.

DL algorithms have also been utilized in digital stethoscope platforms capable of automated cardiac auscultation and murmur detection. AI-assisted cardiac auscultation in practice refers to the auto-interpretation of phonocardiograms, within the domain of signal processing. Various algorithms can diagnose severe AS with sensitivity and specificity of over 90% (139). Chorba et al. developed and tested a deep CNN model that classifies phonocardiograms in 3 categories: heart murmur, no murmur, or inadequate signal. To determine the accuracy of the signal quality and murmur detection performance, the algorithm output was compared with annotations from three cardiologists. This algorithm yielded a sensitivity and specificity for detecting murmurs of 76.3% and 91.4%, respectively. The study also used the gold standard echocardiogram to detect clinically significant (deemed as moderate or greater) AS and MR, with moderate success. Detection of AS was 93.2% sensitive and 86.0% specific, and detection of MR was 66.2% sensitive and 94.6% specific (139).

AI-platforms can be useful in screening for valvular heart disease through auscultation. One such AI-apparatus recently approved by the FDA is the Eko Murmur Analysis Software (EMAS), a “smart” stethoscope which can detect and characterize even asymptomatic murmurs in both pediatric and adult patients (140). This valuable tool which can be used on a daily basis during physical exams, can assist with early diagnosis of valvular heart disease such as aortic and mitral valve stenosis which are among known risk factors for ischemic stroke.

Artificial intelligence and machine learning have emerged as powerful tools to aid clinicians in predicting stroke and related outcomes. These models have demonstrated equal or superior performance to traditional statistical methods, prompting further investigation into the potential for AI/ML integration into risk assessment and prognostic prediction for stroke across various cardiovascular conditions.

Zhou et al. utilized machine learning algorithms to predict stroke and mortality in patients with mitral regurgitation (141). They found that a gradient boosting machine (GBM) method significantly outperformed many traditional statistical methods in predicting adverse outcomes of TIA/stroke and all-cause mortality. For prediction of TIA/stroke, the GBM model achieved an AUC of 0.8084 in comparison to 0.4128 for logistic regression (LR), 0.5528 for decision tree (DT), 0.7202 for random forest (RF), 0.7429 for support vector machine (SVM), and 0.7533 for artificial neural network (ANN). For prediction of all-cause mortality, the GBM model achieved an area under the receiver operating characteristic curve (AUC) of 0.7962 in comparison to 0.4063 for LR, 0.5490 for DT, 0.7132 for RF, 0.7354 for SVM, and 0.7702 for ANN (141).

Extending this research, Rauf et al. applied a similar GBM model to patients with mitral stenosis with atrial flutter (142). Their results further confirm the GBM model's efficacy, with an AUC of 0.8037 for TIA/stroke risk prediction and 0.8388 for all-cause mortality risk prediction, significantly outperforming traditional statistical methods with AUC ranging from 0.4392 to 0.6976 for TIA/stroke and 0.4719 to 0.6174 for all-cause mortality (142).

In context of AS, Shimoni et al. trained a random survival forest (RSF) model using data from an AS registry to assess prognostic characteristics with an AUC of 0.83 (0.80–0.86) for 1-year outcome and 0.83 (0.81–0.84) for 5-year outcome (143).

Several predictive scales have been developed to assess the risk associated with congenital abnormalities, particularly PFO, and their subsequent risk for AIS. The Risk of Paradoxical Embolism (RoPE) score and the PFO-associated Stroke Causal Likelihood (PASCAL) risk stratification system are the most widely utilized risk scales (148, 150). The RoPE score consists of age, hypertension, diabetes mellitus, smoking history, previous stroke, or transient ischemic attack (TIA), and neuroimaging findings to estimate the likelihood of a stroke occurring due to a PFO. Higher RoPE scores indicate a greater probability of a stroke occurring due to a pathologic PFO, but conversely, are associated with a lower risk of recurrent stroke (148, 149, 151). The PASCAL risk stratification system categorizes patients into subgroups based on the presence of high-risk PFO features, such as a large shunt or atrial septal aneurysm, in conjunction with the RoPE score (148, 149).

There have been several studies utilizing ML algorithms for stroke risk in patients with documented CHD. Bai et al. constructed a RF ML algorithm in which they established the relationship between cryptogenic stroke and predictor variables, including morphologic and functional characteristics of PFO, as detected by transthoracic and transesophageal echocardiogram (TTE/TEE) (146). Traditional algorithmic approaches such as SVM and ANN have been considered. However, the high dependence on database accuracy and time efficiency rendered these approaches suboptimal.

Luo et al. constructed their random forest survival (RFS) model under supervised machine learning and used random forest variable importance (VIMP) to select variables such as fasting blood glucose, thickness of interventricular septum, ratio of mitral peak early (E) to late (A) diastolic filling velocity, left ventricular end-systolic dimension, body mass index, systolic blood pressure, and thickness of the posterior wall from a population of patients with post-closure PFO diagnosed by TTE/TEE (148). Compared to the concordance index (C-index) of 0.54 in the traditional Cox proportional hazard regression model, the RFS model had a C-index of 0.87. RFS analysis supported traditional risk factors and identified the ratio of mitral peak early to late diastolic filling velocity and the thickness of the interventricular or posterior wall in post-closure PFO patients as predictive values to prevent stroke recurrence.

In another study in adult patients with CHD by Chiriac et al., two ML models (RegCOX and XGBoost) were constructed as prediction models for stroke and systemic embolism (SSE) (144). Both models were compared against and significantly outperformed the traditional CHA₂DS₂-VASC score among the ACHD population with AUC at 5 years of 0.80 for RegCOX, 0.79 for extreme gradient boosting (XGBoost), and 0.74 for CHA₂DS₂-VASC. Both ML models included the CHA₂DS₂-VASC score as part of their regression in addition to history of cerebrovascular disease, ASD, and the Charlson comorbidity index as the other high-risk variables.

PFO size has been implicated with increased mortality rates in AIS patients and clinical studies have shown varying efficacy for PFO closure vs. medical therapy for secondary stroke prevention with high-risk PFO. Factors such as age and size of PFO have been considered in past clinical trials to support the closure of high-risk PFO and increasing data has supported PFO closure to significantly reduce the risk of stroke or TIA. The analysis by RFS in Luo et al. supported continued targeting of high-risk groups even post-PFO closure (148). ASD, as identified by the RegCOX and XGBoost models of Chiriac et al. has been identified as a significant risk factor for late morbidity and mortality with normal survival reported following early ASD closure. The PASCAL risk stratification has shown in a recent meta-analysis that high-risk PFO features and RoPE score together yield strong predictive value of recurrent ischemic stroke in closure with a 90% relative risk reduction in the PROBABLE group, 62% in the POSSIBLE group, and no significant difference in the UNLIKELY group (149). While the closure of PFO holds promise as a therapeutic intervention for recurrent ischemic stroke, challenges and questions regarding optimal timing of PFO closure, prophylactic PFO closure in adults with CHD, long-term durability of closure devices, potential risks and patient safety factors are still yet to be studied (149).

The cardiac remodeling seen in HFpEF often manifests as EKG changes, a modality critical to the assessment of HF (161, 162). However, these subtle changes are often difficult to detect, which is why EKGs had not previously been an optimal screening mechanism for HF. Various studies have generated AI networks based on waveforms from EKGs to detect cardiac pathology ranging from ejection fraction to valvular disease, each implicated either directly or indirectly to HF (160, 163). The rationale is such that deep learning algorithms and neural networks could be “trained” with enough data to detect the otherwise subtle EKG abnormalities associated with the pathological cardiac remodeling seen in HF.

Akbilgic et al. developed an EKG-AI model using a deep residual CNN based on standard 12-lead EKGs (164). The authors noted that this model could predict HF with moderately high accuracy, with the authors noting an AUC comparable to AUCs from notable HF risk calculators from the Framingham Heart Study (FHS) and Atherosclerosis Risk in Communities (ARIC) study (164).

Other CNNs have been developed to detect HFrEF (with an EF of 40% or lower) with 12-lead EKGs data, with the capacity to “retrain” to utilize single-lead EKG input alone during routine examination, one of which is AI-EKG (165). Utilizing an EKG-enabled stethoscope, this deep learning system was able to detect HFrEF with an AUC of 0.81, a sensitivity of 91.9% and specificity of 80.2% (165). Kwon and colleagues also developed a deep learning model to predict HFpEF, based on an ensemble network using 12-lead, 6-lead, and single-lead EKGs. The authors noted a high level of predictive value (AUC 0.866) in detecting HFpEF using both 6-lead and 12-lead EKGs and noted that this model outperformed already-implemented screening-tests for other medical conditions, such as mammography for breast cancer (161).

In the HF care field, wearable sensors coupled with ML analytics may potentially improve clinical outcomes and reduce hospitalizations. One HF-specific, AI-powered platform which has recently gained approval from the FDA is the EchoGo® Heart Failure platform by Ultromics, based in the United Kingdom (UK). This platform applies AI via CNN to detect HFpEF using just a single 4-chamber TTE image. This apparatus was able to successfully reclassify HF among 73.5% of previously non-diagnostic results, with an accuracy of 90%, and a sensitivity and specificity of 87.8% and 83.0%, respectively (167). Moreover, Us2.ai, a Singapore-based medical technology firm has received FDA clearance for their automated decision support tool for echocardiography (168). LVEF is among the twenty-three validated parameters which can be used to assess cardiac systolic function. Such platforms allow for the accurate acquisition of cardiac measurements even without human input and may better optimize the screening and diagnosis of heart failure. There has also been an investigation done in the prediction of LAT using TEE vs. TTE images. Piezko et al. have done the most substantive investigation of AI in predicting LAT using TTE and clinical data compared to TEE prior to cardioversion or ablation (169). Their AI model accurately predicted LAT (AUC 0.85), outperforming other risk factors (LVEF and CHA2DS2-VASc), and would have avoided TEE in 40% of patients (169). While reducing TEE burdens, this also provides a means of significantly mitigating the risk of ischemic stroke in the future from said thrombi, particularly in an ever-growing cohort of patients with AF and atrial flutter.

With regard to myocardial strain analysis, motion estimation is the critical task to obtain GLS, which is a measure of LV function. This has conventionally been done using speckle tracking algorithms. Using echocardiograms, Salte and colleagues generated a deep learning AI pipeline consisting of 4 ANNs trained for motion estimation and compared this to semi-automated speckle-tracking (170). This pipeline was “taught” to detect patterns in consecutive images to obtain an optical flow vector field to predict local velocities, perform timing of cardiac events, and to visualize cardiac motion (170). Without any human input, this pipeline was able to successfully classify cardiac views, estimate motion, and ultimately measure GLS within 15 s, with accuracy comparable to conventional speckle-tracking methods (170). Therefore, the role of AI applications to assess cardiac function by performing myocardial strain analysis without human input, provides another means of expedited, more efficient detection of HF, and warrants further investigation.

One avenue of potential inquiry in the context of detecting cardiac tumors using AI is that of parallel computing to identify cardiac tumors. Parallel computing entails the use of computers to perform multiple tasks simultaneously to break a multi-faceted problem into individual tasks to be solved (173, 174). In the context of cardiac tumors, this would involve segmenting the various radiographic features of the tumor across multiple imaging modalities and incorporating this data to generate a single diagnosis (173). Computing may be supplemented and enhanced by AI and ML algorithms to improve the sensitivity and specificity of current imaging modalities for diagnosing cardiac tumors (174, 175). In the realm of neurocardiology, this is particularly pertinent with respect to myxomas which may embolize and lead to LVOs and infarcts (176).

AI models have increasingly been used to predict the risk of stroke amongst patients with specific comorbidities (177, 178). Despite the elevated risk of stroke with cardiac tumors such as myxomas and papillary fibroelastomas, there are currently no studies evaluating the risk of stroke in patients with any form of cardiac tumors. This represents an important gap in the literature and should be an area of focus in future studies evaluating the role of AI in stroke prediction.

18F-Fluorodeoxyglucose (FDG) PET/CT involves injecting a radiolabeled glucose analog into the patient and evaluating its uptake throughout the endocardium, myocardium, and other cardiac structures. Radiotracer uptake is higher in metabolically active tissues, as in the case of IE where FDG PET/CT can visualize infected valvular structures with high uptake (185). This technique is most useful to rule out IE in prosthetic heart valves, where other imaging modalities including echocardiography may be extremely limited. However, the lack of standardized metrics of uptake and the difficult distinction between infection and inflammation has limited its use (185, 186). The integration of AI enabled better differentiation among various cardiac pathologies such as valvular abscesses, intracardiac fistulae, and pseudoaneurysms seen in IE, as opposed to reactive inflammatory changes seen either postoperatively or in the setting of chronic valvular disease.

In 2016, Godefroy et al. assessed AI and ML algorithms with FDG PET/CT in the setting of IE (187). Using a sample of 68 patients with a known prior history of IE, the authors developed an AI and ML algorithm model that could be applied to 40 patients with suspected IE in their study (187). The model was combined with the European Society of Cardiology's 2015 criteria for suspected IE. The AI model demonstrated higher rates of accuracy, sensitivity, and specificity when compared to expert clinicians alone (183). Unfortunately, few studies analyzing AI and ML for imaging in IE have been published since, which has stymied the advancement of AI and ML as diagnostic tools for IE patients.

Resende et al. developed a log-linear ML model which was able to predict higher rates of inpatient mortality amongst IE patients undergoing surgery compared to patients who received conservative management (188). Another study by Luo et al. had similar results with a ML model that predicted early mortality and risk of stroke after surgery better than pre-existing logistic regression models and the well-established European System for Cardiac Operative Risk Evaluation model (189).

However, while there have been very few investigations focusing specifically on AI/ML applications for neurological sequelae, the prospective use of AI/ML serves as “proof of principle” that prevention is better than cure. Neurological conditions are among the most common manifestations of systemic IE, and therefore, the focus of using AI/ML technology in the realm of neurocardiology for IE patients is best suited to detection of infective cardiac pathology before the neurological pathologies manifest.

A four-year retrospective study conducted with patients diagnosed with cryptogenic stroke from 4 Korean tertiary hospitals found that an image-based AI model of EKG data had excellent diagnostic performance in AF prediction. The AI-AF risk score, derived from sinus rhythm EKGs, demonstrated high predictive accuracy for AF, with an AUC of 0.806. Incorporation of parameters such as left atrial volume index, age, sex, BMI, atrial ectopic burden, and PR interval further improved the model's performance (AUC 0.880, p = 0.002). There was also a significant temporal trend showing increased AI-AF score with increased AF duration and as recorded EKG time approached AF onset (37). This AI-based approach not only aids in early AF detection, but also offers potential timesaving and cost-effective advantages for timely interventions to prevent recurrent strokes.

These platforms can utilize EKG data to retrospectively detect undiagnosed AF in cryptogenic stroke patients. Even for patients who already have suffered from cryptogenic stroke, AI can detect AF as an etiology for cryptogenic stroke, thereby allowing for secondary prevention of future ischemic strokes. Choi et al. recently developed a transformer-based AI model trained to analyze EKGs from cryptogenic stroke patients both to detect and predict paroxysmal AF (37). This model demonstrated a robust performance, with an AUC of 0.880, and a greater accuracy in identifying longer duration AF episodes (37). Ho et al.'s meta-analysis found multiple new wearable technology, that includes those currently being used after cardioembolic stroke, to be feasible options for diagnosis of paroxysmal atrial fibrillation in cryptogenic stroke (194). However, they noted no current significant difference of these modalities compared to classic workup technology (such as Holter monitoring) and emphasized the need for further studies.

Comparing the efficacy of wearable technology for cardiac monitoring after stroke found there were promising technologies in EKG and PPG data as previously described for afib.

Investigators such as Garg et al. have demonstrated that ML models with natural language processing (NLP) may provide an avenue of automation for the classification of ischemic stroke subtypes based on data from patient electronic medical records (EMR), medical history, physical examinations, laboratory and imaging results (77). These authors tested various ML models such as K-nearest neighbors (KNN), random forests (RF), and XGBoost and conducted a stacking process, whereby model predictions are used to generate new models. They found an 80% agreement between a stroke neurologist and the ML model using stacking for logistic regression with kappa of 0.72 for TOAST classification, with excellent subtype discrimination in all ML models (c-statistic >0.8) aside from the KNN approach. This study suggests that ML models can not only enhance scalability of large-scale epidemiological studies and clinical trials, but also remove a point of error in manual adjudication of disease classification and investigator bias (77). For cryptogenic stroke, ML models may help in analyzing etiologic phenotypes leading to further subtyping for directed treatment.

Using AI tools such as the NLP machine learning model to automate subtype classification in ischemic stroke can significantly impact management strategy based on characteristics found in various clinical studies (77). Management of cryptogenic stroke currently focuses on secondary prevention to reduce the risk of recurrence. Risk stratification tools, including the AI-based models listed above, can guide physician decisions on anticoagulation vs. antiplatelet therapy for patients with or without identified sources of cardioembolism, such as AF or a PFO (190).

Additionally, the CHA₂DS₂-VASc score, traditionally used to predict stroke severity in patients with AF, is now being considered in patients with cryptogenic stroke and underlying congenital heart disease (CHD) (144).

Detection of congenital heart abnormalities in cryptogenic strokes include the use of modalities such as TTE, TEE, cardiac MRI, cardiac CT, and transcranial doppler ultrasound (TCD). TEE with agitated saline injection remains the gold standard for PFO diagnosis relative to intraoperative inspection, with a weighted sensitivity of 89% (149). The presence of PFO in cryptogenic stroke patients has been associated with various morphologic and functional characteristics, including the size of the PFO, mobility of the interatrial septum, and the presence of a right-to-left cardiac shunt.

The CHALLENGE ESUS/CS registry, a multicenter registry of cryptogenic stroke patients, utilized TEE to identify potential embolic sources (147). While there are advantages to cardiac MRI for detecting certain cardioembolic sources, a meta-analysis showed a 29.3% (95% CI, 23.6%–35.0%) detection yield for CMRI compared to TEE at 53.7% (95% CI, 47.4%–49.9%) (P < 0.001) (149).

For PFOs, Bai et al. found that, in cryptogenic stroke patients, PFO size was significantly greater in systole [2.0 (1.5, 2.9) mm vs. 1.6 (1.1, 2.0) mm, p < 0.001] and diastole [1.7 (1.4, 2.2) mm vs. 1.3 (1.1, 1.8) mm, p < 0.001] than in non-CS patients and that large PFO were also more common (p < 0.001) (146). ANN and SVM were performed as comparisons with goodness of fit testing and coefficient of certainty, with higher accuracy and reliability noted with the RF model. The final test set accuracy was 70%, with an AUC of 0.816 and a sensitivity and specificity of 73% and 65%, respectively. Additionally, the frequency of right-to-left shunt was significantly higher in CS patients (p < 0.001) (146). Supervised RF learning is insensitive to noise with good generalizability, making it suitable for the high-risk features of PFO.

AI may facilitate the use of ML to analyze diagnostic modalities for the early detection of atrial pathology, thereby mitigating the risk of AF and of subsequent cardioembolic and cryptogenic strokes. Additionally, it can allow for detection of undiagnosed AF to reduce the risk for future cardioembolic events even for patients who already suffered from strokes.

Studies into AI for atrial cardiomyopathy have recently been conducted with the rationale that early progression of the disease process can manifest as subtle EKG changes which can be detected via ML. Investigators at the Mayo Clinic developed an AI algorithm consisting of a CNN trained to screen for temporal and spatial features on 12-lead EKGs to identify and stratify atrial cardiomyopathy for patients with heart failure with preserved ejection fraction (HFpEF) (198). This algorithm was able to identify structural heart disease in patients with higher AI-probability for AF based on EKG findings, with these patients demonstrating LVH, increased atrial filling pressures, and reduced cardiac output (198).

Therefore, while AI is still in its relative infancy for use in atrial cardiopathy, these algorithms provide a means of primary, secondary, and tertiary prevention. The timely detection of subtle cardiac structural and hemodynamic changes can allow for clinicians to intervene before the manifestation of further cardiac pathologies such as AF and subsequent stroke.

Echocardiography is the primary imaging modality, used to diagnose TTC. However, initial differentiation between ACS/MI and TTC remains challenging. Therefore, investigators have sought to develop algorithms to utilize echocardiograms to differentiate TTC from MI. Laumer and colleagues developed and trained a DL algorithm using TTEs from 228 patients to differentiate TTC from acute MI (206). They utilized a temporal CNN to predict the probabilities of TTC and MI for the respective patients and compared these findings to echocardiographic evaluation by 4 cardiologists (206). This algorithm demonstrated a superior performance to the committee of cardiologists in differentiating TTC from acute MI, with an AUC of 0.79 and an overall accuracy of 74.8% (compared to an AUC 0.71 and an accuracy of 64.4% for the cardiologists) (206).

While TTE has traditionally been a first-line imaging modality for TTC, cardiac MRI is emerging as a gold standard, non-invasive modality to differentiate TTC from acute MI. For instance, T2-weighted imaging can detect myocardial edema, which manifests in the acute phase of TTC and is indicative of reversible myocardial injury (207). Other MRI criteria for TTC include wall motion abnormalities and the relative lack of contrast enhancement (208). Additionally, atrial and ventricular strain parameters are of growing relevance to diagnosing TTC, and MRI provides an ideal modality to conduct strain analyses. Authors have also utilized ML with cardiac MRI protocols to diagnose and differentiate TTC from other cardiac pathologies. Cau et al. created a cardiac MRI-based ML ensemble model using the Extremely Randomized Trees algorithm trained with MRI strain measurements and demographics to different TTC from myocarditis. The authors noted a high degree of sensitivity (92%) and specificity (86%) with an AUC of 0.94 in diagnosing TTC (209). The authors also noted that strain rate and left atrial conduit strain were the parameters that yielded the greatest value for TTC identification (209).

Investigators have developed AI algorithms for stroke risk prediction in TTC. De Filippo and colleagues analyzed patients from the International Takotsubo (InterTAK) registry in a multi-center study and designed their own ML model to predict in-hospital mortality (210). This was a logistic regression-based algorithm evaluated in internal and external validation cohorts (210). This InterTAK model demonstrated a strong performance in predicting mortality for the internal validation cohort, with an AUC of 0.89, a sensitivity of 85% and a specificity of 76% in the internal validation cohort (210). The model also yielded satisfactory performance on the external validation cohort, with an AUC of 0.82, a sensitivity of 74%, and a specificity of 79% (210). The authors also showed the further applicability of ML: AI for subsequent regression analyses to ascertain patient characteristics also associated with mortality. They noted that LVEF, atrial fibrillation, physical stress, leukocytosis, and age were among the most relevant features associated with mortality. From these ML-based analyses, the authors were able to cluster patients into six risk groups based on short-term prognoses (210).

There have been few validated scores developed to predict NDAF with a high degree of external validity, two of which include the CHASE-LESS and AS5F scores (220–224). However, various ML algorithms are currently being developed to provide a higher accuracy prognostic modality in this role (225–228). Sung et al. designed an ML algorithm to utilize stroke hospital admission data in the EMR to predict newly diagnosed AF early in patients’ hospital courses (227). This model utilized common EMR data fields and clinical free text information. It outperformed standard prediction methods (as shown by a higher C-index) (227). This type of automated risk predictor allows clinicians to treat cerebral ischemia while understanding the full clinical picture of the patient, encompassing the broader heart-brain axis.

AI models have also been generated using other markers, such as BNP, neuroimaging, and genetics. A 2022 study by Pang et al. developed and evaluated an integrated nomogram model for post-stroke AF based on variables such as prior TIA, BNP levels, CRP levels, and National Institute of Health Stroke (NIHSS) scores. This integrated model provided enhanced benefit predicting the incidence of post-stroke AF, identifying higher-risk patients who were candidates for extended monitoring (217).

Cho et al. developed a DL algorithm to use 6-lead EKGs to diagnose MIs, with reasonable accuracy (sensitivity 84.4%, specificity 88.5%) and high negative predictive value (97.5%) (233). This has undergone iterative development to extend to 12-lead EKGs. One Taiwanese retrospective study investigated a DL model trained detect STEMI through analysis of 100,000 EKGs and compared its performance to cardiologists. The authors noted a superior performance of the DL algorithm, with a high AUC of 0.997, sensitivity of 98.4%, and specificity of 96.9% (234).

Moreover, the Rule-Out Acute Myocardial Infarction Using Artificial Intelligence Electrocardiogram Analysis (ROMAIE) trial is among the major prospective clinical trials currently underway to validate a DL-based 12-lead EKG algorithm (235). This study will allow for real-time substantiation and further application of AI in a clinical setting for AMI.

There are relatively few studies to date which have analyzed AI algorithms in the detection and prognostication of CAD specifically in stroke patients. One such study was conducted by Heo et al., who obtained multidetector coronary CTAs for patients who suffered acute ischemic strokes without a known history of CAD (236). The authors obtained demographics, laboratory results, NIHSS scores, blood pressure, and carotid stenosis to generate an ML model to predict CAD (236). They noted a satisfactory performance in predicting CAD, with an AUC of 0.763 in predicting CAD of any degree, and an AUC of 0.714 to predict obstructive CAD (236). Other investigators have developed models to incorporate both DL and ML algorithms using a combination of EKG, clinical, and laboratory data to predict obstructive CAD, with similar results (AUC of 0.767) (237).

However, the results of AI investigations for broader patient cohorts have demonstrated the potential of AI as a diagnostic and prognostic tool for CAD. Kang et al. developed an SVM-based learning algorithm to utilize coronary CTAs to diagnose coronary artery stenosis and compared this algorithm to three expert readers of CTA imaging (238). This model showed a superior performance to the human readers with a high degree of accuracy (AUC 0.94, sensitivity of 93%, sensitivity of 95%) (238). While this model did not specifically focus on stroke patients, it demonstrates the potential of such algorithms as tools for diagnosis and prognostication of CAD which can be applied specifically for ischemic stroke patients in future studies.

There have already been substantive investigations conducted in assessing the use of AI/ML, DL, or CNN algorithms for the timely detection and prognostication of acute ischemic stroke, cardiovascular diseases, and neurological pathologies. These studies have demonstrated an iterative increase in accuracy, speed, and applicability of this technology within diverse settings. The recent algorithms discussed have also proven their strong statistical power in evaluating complex datasets and identifying trends that may otherwise elude traditional statistical analysis.

These advancements in clinical use originate from the ability to incorporate anatomic analysis, functional and hemodynamic parameters, clinical factors, and imaging modalities. Mobile devices and other widely sold technologies continue to provide greater ease of monitoring, more timely diagnosis, and therefore, more timely treatment. Simply having increased accessibility to continuous monitoring may also increase the predictive value of them. These uses thereby may improve patient outcomes and reduce the overall burden on the healthcare system.

As well as its application in standard clinical decision making, future directions using AI/ML include improving access to more equitable healthcare to underserved communities (82, 85, 86). This may be accomplished through increased digital access to diagnostic tools, improving efficient clinical decision making for disease escalation, and reducing physical or cultural barriers through access to high quality care with telemedicine (86, 244, 271). Some of these goals are already evident with common technology today, through systems such as broadened Medicare coverage for telestroke (272). Technology can also provide greater autonomy to patients with managing their own health needs, such as the tools above which allow for cardiac monitoring even at home or in clinic (273). Furthermore, substantive research has been conducted to address the limitations of bias evident in AI algorithms with suggestions of improvement including careful cohort selection, cultural awareness and expertise in developing algorithms, and joint multicenter data (271, 274–277).

There is a clear demand for expanded clinical use of these AI-enabled platforms and algorithms. Robust validation of these ML/AI models is required for future clinical use, as well as ensuring data privacy and security, addressing algorithm bias and interpretability, and future studies on multi-center populations. The widespread incorporation of these models may also offer an avenue to accelerate the development of better diagnostic and predictive technology.

There is an evident difference in the depth of AI/ML research conducted between certain neuroradiology pathologies. Of note, the investigation of this technology remains limited for cardiac tumors, the detection and prognostication of CAD in stroke patients, and in the neurological sequelae of infectious endocarditis, such as meningitis, abscesses, mycotic aneurysms, and intracerebral hemorrhage. Further studies are warranted to expand on the prospective benefits of AI/ML applications in these and other related cardiovascular and cerebrovascular conditions.

The future of AI/ML in neurocardiology is promising, particularly in the development of future investigations that may provide further diagnostic and prognostic value for stroke patients, better elucidate the sequelae of presenting and developing diseases, and improve patient care with its increased efficacy. As AI/ML becomes more commonly incorporated into clinical practice, the knowledge base, algorithmic training, and accuracy of models will continue to improve.