Transcatheter aortic valve implantation (TAVI) is a crucial procedure in treating severe aortic stenosis, which is characterised by the narrowing of the aortic valve (1). TAVI provides a minimally invasive alternative to open-heart surgery for older patients and those with numerous comorbidities (2) with faster recovery times, shorter hospital stays, and less procedural risks (3–6). In patients with severe aortic stenosis, TAVI has been found to reduce symptoms, as well as increase quality of life and overall survival rates (7, 8). The emphasis of this paper is to predict the risk for TAVI, given that it has already been proven to have multiple benefits.

The mortality rate after TAVI might vary depending on a number of factors, including patient characteristics, comorbidities, and intra-procedural concerns. For intermediate-risk patients undergoing TAVI, longer-term follow-up in the PARTNER 2 study revealed a death rate of 26.2% at 5 years (9, 10). The SURTAVI trial showed that all-cause mortality was 31.3% at 5 years using the CoreValve self-expanding prosthesis in intermediate-risk patients (11). In a similar vein, the CoreValve U.S. Pivotal High-Risk Study found that high-risk patients had a 5-year mortality rate of 47.9% (12), and a separate study found the mortality rate to be 58.8% (13). The NOTION trial demonstrated that in low-risk patients, the mortality rate of TAVI with the first-generation CoreValve self-expanding prosthesis was 27.6% after 5 years of follow up (14). The PARTNER 3 trial also showed that the 5-year all-cause mortality for TAVI was 10.0% (15). One of the more recent trials, Evolut trial, found that the 4-year mortality for TAVI using self-expanding CoreValve was 10.7% (16). A separate meta-analysis has also found the all-cause mortality to be higher in low-risk patients undergoing TAVI over surgical aortic valve replacement (27.5% vs. 17.3%) (17). Thus, given the high probability of fatality and differences in mortality depending on the risk levels of patients, prediction prior to the start of the procedure is increasingly necessary.

Improving risk assessment and patient outcomes may be possible with the application of artificial intelligence (AI) in predicting death after TAVI. In-depth patient data analysis by AI algorithms has the ability to uncover pertinent trends and risk factors that can help forecast post-TAVI death (18). These models can help identify high-risk individuals who might need more measures or more intensive postoperative surveillance. On the other hand, AI models could also identify patients who may not benefit from TAVI, given that the risk of mortality outweighs the benefits of undergoing the procedure.

Thus, our goal in analysing the accuracy of AI-generated death forecasts in TAVI procedures was to determine how well AI algorithms can predict mortality outcomes for patients undergoing TAVI. This evaluation aims to assess the effectiveness of AI models in predicting post-TAVI death rates and evaluate the accuracy of these forecasts.

We conducted a systematic review following guidelines from the Preferred Reporting Items for Systematic Reviews and Meta-analyses for systematic review (PRISMA) standard (19). In total, we performed our search on 4 databases, including PubMed, Medline, Embase and Cochrane, from the date of inception to 24 October 2023. Across all databases, combinations of different search terms, including Medical Subject Headings (MeSH) terms were generated. The following search strings were used: “Transcatheter aortic valve replacement AND artificial intelligence”, “Transcatheter aortic valve replacement AND machine learning”, “Transcatheter aortic valve replacement AND deep learning”, “Transcatheter aortic valve implantation AND artificial intelligence”, “Transcatheter aortic valve implantation AND machine learning”, “Transcatheter aortic valve implantation AND deep learning”, “Aortic stenosis AND artificial intelligence”, “Aortic Stenosis AND machine learning” and “Aortic stenosis AND deep learning”. The search was perfomed using MeSH terms only.

In total, our systematic search yielded 2,239 records. After 1,504 duplicate records were removed, 735 remained for title and abstract review. The high number of duplicates was likely due to our search strings throughout the various databases producing similar articles. Based on our exclusion criteria, we eliminated 669 studies and retrieved full-text articles for 66 articles. Finally, a further 56 articles were excluded based on full-text assessment, leaving 10 studies (29–38) for data abstraction (Figure 1).

To the best of our knowledge, this is the first meta-analysis in literature comparing AI models against traditional scores in predicting post-TAVI mortality. Overall, the results of our double-arm meta-analysis demonstrate that applying AI to predict death in TAVI cases has a higher predictive accuracy than conventional clinical scoring techniques. In particular, the 30-day (AUC: 0.75) and 1-year mortality (AUC: 0.79) with non-gradient-boosting algorithm subgroups both demonstrated high mean AUC values with low heterogeneity. In our combined analysis of traditional risk scores, the pooled mean AUC value for intra-hospital mortality, 30-day mortality, 1-year mortality with gradient boosting, 1-year mortality without gradient boosting, and 5-year mortality was 0.95 [0.90, 1.00] 0.75 [0.72, 0.79], 0.79 [0.72, 0.86], 0.68 [0.67, 0.69] and 0.79 [0.75, 0.83], respectively. Hence from gross comparison of short- and long-term AUC alone, AI-based models may fare better than current scoring methods at correctly predicting death outcomes in TAVI patients. This suggests that AI models have the potential to aid clinicians in predicting post-TAVI mortality in patients and act as an adjunct or alternative to traditional clinical scores. AI models are able to process large amounts of diverse patient data and can be modelled to continuously analyse new data in order to make accurate predictions in real time. This is a significant advantage over traditional scores, which can only utilise a limited number of variables to predict outcomes.

In the study by Kwiecinski et al. (29), the number of packed red blood cell units transfused, length of hospital stay and minimum estimated glomerular filtration rate provided the greatest contribution to the AI model, with specificities of 94 (92–97)%, 33 (29–37)% and 53 (49–58)%, respectively (29). No other studies reported the specificity values for the individual clinical variables used in their predictive models. As an illustration of the multiplicity of variables that AI models can evaluate, the main 30-day mortality predictive variables reported in the study by Lertsanguansinchai et al. (31) were height, chronic lung disease, STS score, preoperative left ventricular ejection fraction (LVEF), age, and preoperative left ventricular outflow tract velocity time integral (VOT VTI), while the main 1-year mortality variables were preoperative LVEF, STS score, hear rate, systolic blood pressure, home oxygen use, serum creatinine level, and preoperative LVOT Vmax. A more detailed summary of the variables used across the included studies can be found in Supplementary Table S1.

Researchers often compare the AI-generated death predictions with the actual mortality results seen in real-world TAVI patients to assess model accuracy. This evaluation aids in determining the degree to which the AI models are capable of making trustworthy and precise forecasts of post-operative mortality. Here, we provide a brief summary of the advantages and disadvantages associated with each AI algorithm featured in our study.

Firstly, the RF model is capable of analysing high-dimensional data with a huge number of diverse predictors, which can even exceed the number of observations (39).While a single decision tree is prone to noise and overfitting when grown on its training set, the RF model improves accuracy by averaging multiple decision trees. However, due to their complexity, RF models tend to have low intrinsic interpretability (40). Additionally, the presence of dependent observations in data may contribute to increased bias and inaccurate predictive variable selection (41). While the GB algorithm minimises errors and maximises predictive accuracy in the final model, GB algorithms, like the RF models, may also suffer from low interpretability. The GB model may also be prone to over-fitting if the additive process of gradient boosting is not regularised. ANN and MLP algorithms are suitable for analysing non-linear relationships between dependent and independent variables. However, these models are difficult to apply to real-time predictions and are also prone to overfitting (42). Finally, LR algorithms can only be used to predict discrete functions and cannot predict continuous outcomes. It also assumes a linear relationship between dependent and independent variables.

Overall, AI-based prediction models are able to take into account a myriad of patient data to generate more accurate predictions on post-TAVI mortality than traditional scores. However, such models require a large number of patient variables to generate predictions, some of which may not be readily available to clinicians in the immediate clinical setting. Hence, at present, AI-based prediction models may be less user-friendly than simple traditional risk scores. In future, further research into simpler AI-based models that are able to use easily-available clinical parameters in predictions is needed to increase the clinical utility of these models.

Firstly, the RF model constructs a series of decision trees and utilises bagging and randomisation of predictors in order to accurately predict outcomes (24). The original dataset is first divided into smaller subsets through random feature selection. Individual decision trees are then grown on the subsets, allowing for the construction of many decision trees with low correlation to one another. Finally, multiple decision trees are averaged, thereby minimising variance and improving the accuracy of the final predictive model (24). The RF model is thus capable of analysing high-dimensional data with a huge number of diverse predictors, which can even exceed the number of observations (39).

While a single decision tree is prone to noise and overfitting when grown on its training set, the RF model improves accuracy by averaging multiple decision trees. However, due to their complexity, RF models tend to have low intrinsic interpretability (40). Additionally, the presence of dependent observations in data may contribute to increased bias and inaccurate predictive variable selection (41).

The GB model involves an ensemble of weak learner decision trees. Each weak learner progresses through a stepwise progression, where previous iterations of weak learners are combined into a successive strong learner, correcting for the errors of the preceding learner each time (25, 26). Eventually, this process is repeated until errors are minimised and predictive accuracy maximised in the final model. However, similar to the RF model, GB algorithms may also suffer from low interpretability. The GB model may also be prone to over-fitting if the additive process of gradient boosting is not regularised.

The ANN algorithm is a mathematical model developed based on the transmission of signals amoung neural networks in biological nervous systems. An input signal is channelled through a collection of nodes, which are artificial neurons (27). Each node analyses the input signal and then transmits an output to each of its connected neurons, mimicking the transmission of action potentials across neurons and synapses in the human brain. The nodes are also organised into layers. The input signal travels from the first input layer to the last output layer, undergoing different transformations at each layer.

The MLP algorithm is a class of feedforward ANN that is comprised of an input layer, one or more hidden layers and an output layer. The input signal first passes sequentially through the layers in a feedforward process until an output is generated. Subsequently, the data is then backpropagated from the output layer back into the hidden layers. The error signal of each node that contributed to the overall error is determined and the weights of the network updated until the gradient of the mean squared error converges. While ANN and MLP are able to analyse non-linear relationships between dependent and independent variables, these models are difficult to apply to real-time predictions and are also prone to overfitting (42).

The LR algorithm fits a logistic function onto a dataset and predicts the probability of an independent variable, such as mortality, from a dependent variable. The maximimum likelihood estimation method is most commonly used to maximise the likelihood function and find the optimal fit of the model (28). LR algorithms can only be used to predict discrete functions and cannot predict continuous outcomes. It also assumes a linear relationship between dependent and independent variables.

Overall, AI-based prediction models are able to take into account a myriad of patient data to generate more accurate predictions on post-TAVI mortality than traditional scores. However, such models require a large number of patient variables to generate predictions, some of which may not be readily available to clinicians in the immediate clinical setting. Hence, at present, AI-based prediction models may be less user-friendly than simple traditional risk scores. In future, further research into simpler AI-based models that are able to use easily-available clinical parameters in predictions is needed to increase the clinical utility of these models.

The main limitations of this study were firstly, the high heterogeneity of the data, specifically in comparing the intra-hospital mortality subgroup, 1-year mortality with gradient boosting subgroup and the overall pooled AUC of all 10 studies. We postulated that this was due to significant differences in the training datasets used for each AI model. Fundamentally, there was large variability in the quantity of data, as well as in the type and number of parameters used to predict mortality in each model. In addition, due to the lack of studies comparing the same AI algorithms and the same control group, it was difficult to perform a more homogenous subgroup analysis on each AI algorithm. Hence, our results may not be generalisable to all AI algorithms or all datasets predicting mortality in patients post-TAVI.

Secondly, all the studies included in our analysis were of a retrospective nature. There was also a lack of data comparing AI model predictive performance to traditional scores. Therefore, further research, including randomized controlled trials, comparing the use of AI algorithms with traditional scores in predicting post-TAVI mortality will be needed in the future in order to determine the validity of our observations.

Finally, our study did not analyse other endpoints that are known post-TAVI complications, such as stroke, pacemaker need and heart failure, and focused solely on mortality. Hence, this may limit the applicability of our findings in clinical practice.

Personalized risk assessments are the ultimate goal of research into AI systems for predicting patient outcomes in TAVI procedures. The potential of AI-generated mortality forecasts to improve the precision and value of risk assessment in TAVI is highlighted in this systematic review. AI can help healthcare providers predict and monitor patient outcomes, which will hopefully result in better decision-making and more desirable post-TAVI outcomes in future.