The need for artificial intelligence literacy in cardiology
Evidence-based medicine is the systematic approach through which physicians use the best available scientific evidence from clinical research to guide clinical decision-making. However, the increasing use of artificial intelligence (AI), which in most current cardiology applications is operationalized primarily through machine learning (ML) models, is contributing to a research paradigm increasingly based on large and heterogeneous datasets and extending beyond traditional statistical approaches. The availability of these large datasets, together with the marked reduction in the cost of data processing, advances in computing power, and improvements in algorithms and ML techniques, has facilitated the expansion of this research approach in medicine and cardiology.
This evolution is crucial, as the volume of data continues to grow without a corresponding increase in human cognitive capacity.
The goal of AI/ML methods is to transform complex, high-volume data into clinically interpretable information. In cardiology, these tools may support diagnostic accuracy, risk stratification, precision-medicine research, clinical workflow, and hypothesis generation.1,2
In cardiology, this translates into detecting subtle ECG abnormalities, analysing cardiac imaging with greater accuracy, or predicting patient outcomes through the integration of multiple clinical variables.3–8
Despite significant interest in the application of AI in cardiology, the technical terminology and various ML methods used are often difficult to grasp for non-experts. Therefore, it is essential for cardiologists to develop a basic understanding of these concepts to effectively engage with this emerging field of cardiovascular research. This manuscript is a scientific statement intended to provide a pedagogical framework for clinicians.
It is not a technical manual for developers, but a guide to the conceptual depth required for safe clinical integration. This scientific statement aims to discuss AI in cardiology from an educational standpoint, explaining what cardiologists need to know, particularly to avoid exclusively relying on experts for choosing research methods in this field, and instead interact with them with greater awareness.
Defining the scope of artificial intelligence literacy for cardiologists
Artificial intelligence literacy for the cardiologist is not synonymous with technical proficiency in computer programming or mathematical derivation. Rather, it represents a functional competence required to navigate the modern evidence-based landscape. This literacy can be categorized into three levels of depth:
Decoding the key terms you must know
Fundamentals of artificial intelligence
In this statement, AI is used as the umbrella term for computational systems intended to perform tasks that usually require human intelligence; ML is used when referring specifically to models that learn patterns from data, which represent the main focus of current cardiovascular applications.
Artificial intelligence
Artificial intelligence is a field within computer science that develops systems designed to mimic human intelligence and to perform tasks that usually require human cognitive abilities. These tasks include, for example, recognizing patterns, learning from data to make decisions and understanding language. Artificial intelligence systems can function with different levels of independence, from basic automation, often referred to as ‘narrow AI’, to more complex and adaptive learning capabilities, approaching what is known as ‘general AI’.
Machine learning
Machine learning, a branch of AI, involves algorithms and models that enable computers to perform tasks by learning from data without explicit programming. It includes training on datasets to improve performance and encompasses various techniques: supervised learning using labelled data, unsupervised learning to identify patterns in unlabelled data, semi-supervised learning, which uses a small portion of labelled data (due to the high cost of human labelling for each case), and reinforcement learning, which optimizes a sequence of decisions to achieve maximum cumulative reward.
Algorithm
In general, an algorithm is a set of defined, sequential instructions designed to perform a specific task or solve a problem, fundamental in computer science for software development and data operations. In ML, an algorithm refers to a set of mathematical procedures or rules used to learn patterns from data. These algorithms adjust their parameters based on input data to improve their performance on tasks such as classification, prediction, and decision-making over time.
Dataset, structured and unstructured data
In AI or ML, a dataset is a collection of data used to train and test algorithms, essential for model accuracy and performance. Data are considered structured when they are organized in a predefined format, typically arranged in rows and columns and stored in databases or spreadsheets. Such datasets follow a consistent order and are easily searchable, making them ideal for algorithms that require direct access to specific numerical or categorical variables.
‘Unstructured data’, on the other hand, refers to data that lack a predefined format or structure. This includes formats such as text, images, audio, and video. Processing unstructured data often requires more complex techniques to extract useful information and patterns, since it does not conform neatly to traditional database fields.
Labelling
Labelling is the process of attaching a specific tag or description to input data, such as identifying a picture as ‘a cat’ or an email as ‘spam’. This step is essential for training supervised learning models, as it provides the examples the models use to learn and make predictions on new, unlabelled data.
Feature engineering: features, selection, and extraction
A feature is a measurable property of data used as an input for algorithms (Figure 1A). Feature engineering encompasses two distinct processes: feature selection, which involves choosing the most relevant subset of variables from existing data to reduce complexity (Figure 1B); and feature extraction, which creates new features by transforming raw, high-dimensional data (such as the raw ECG waveform) into a more compact representation (Figure 1C).
In ECG analysis, feature extraction is critical. It mathematically transforms the raw signal—often using time-frequency analysis or deep learning layers—to generate numerical descriptors that capture signal morphology and dynamics. Unlike simple visual markers, these extracted features can reveal subtle, non-linear patterns invisible to the human eye, translating the raw ECG into a format the algorithm can efficiently process.
Learning techniques and methods
Four major ML techniques exist depending on whether the outcome variable is used during the training process.
Supervised learning
Supervised learning refers to algorithms that leverage labelled data to optimize model parameters, aiming to minimize the discrepancy between predictions and actual outcome labels. Depending on the data type of the outcome label, we refer to classification tasks, where the outcome variable is categorical, and regression tasks where the outcome is continuous.
Unsupervised learning
Unsupervised learning refers to methods which identify patterns in the data without outcome labels. These algorithms, for example, can be used to cluster data and identify previously unknown relationships between parameters, or to reduce model complexity through dimensionality reduction.
Semi-supervised
Semi-supervised or weakly supervised learning refers to learning strategies where only a portion of the data are labelled (for example due to the high cost of humans labelling each case), but models are trained using both the labelled and unlabelled data to utilize all available data.
Reinforcement learning
Reinforcement learning is a learning strategy best for dynamic environments where the given model or agent conducts actions and is rewarded depending on whether it got closer to the desired outcome. This strategy allows a more general adaptation to changing environments as it learns the optimal behavioural patterns through trial and error. In many cases, trained models are publicly available for a given problem (e.g. trained CNNs for image recognition).
Instead of initializing the training process using randomly selected model weights, we can utilize these pretrained models as starting points for our model setting and fine-tune these pretrained models for our specific task. This process is referred to as transfer learning.
Specific structures and models
Tree-based methods
Decision trees (DTs) and random forests (RFs) are types of ML models, specifically referred to as tree-based methods, and they fall into the category of supervised learning.9 Decision trees where the target variables are continuous values (typically real numbers) are called regression trees. Decision trees where the predicted variable is categorical (typically class labels) are called classification trees.
Within the tree, there are ‘non-leaf nodes’ and ‘leaf nodes’. The non-leaf nodes are the questions that guide you through the tree: depending on the answer, you follow a different branch. These non-leaf nodes help narrow down the options as you move closer to the final decision.
In contrast, leaf nodes are found at the end of the tree and do not pose further questions. They represent the final outcomes you arrive at after answering all the questions.
In the context of supervised learning, each non-leaf node corresponds to a split in the training data based on an input feature. A single split of the training data is unlikely to result in a highly accurate prediction; however, by creating a set of hierarchically organized splits (in the form of a tree), a better prediction can be achieved.
While DTs are simple and effective models, their performance can be limited in real-world situations. To improve predictions, multiple DTs can be combined into a ‘RF’ which is part of a class of techniques called ensemble methods.10
This process of building trees in parallel allows each tree to function independently, which helps reduce variance and increase the model’s stability. In RF, multiple DTs are trained using a method called bagging, where bootstrap samples are generated by sampling the training data with replacement, so that each tree is built on a different dataset. The final prediction is made by averaging or voting across the trees.
Another approach is the gradient boosting machine (GBM), where each new model focuses on correcting the errors made by the previous ones, making it particularly powerful for both classification and regression problems. Whereas RF consists of trees built in parallel, without dependence on one another, GBM consists of trees built sequentially, with each tree attempting to correct the errors of the previous ones (Figure 2).
Supplementary material online, File S1 (related to Figure 2).
Example of decision tree
Imagine you have a patient who experienced syncope, and you need to determine the reason. Initially, you might ask: ‘Is the patient young or elderly?’. If the answer is ‘young’, you proceed in one direction, but if the answer is ‘elderly’, you take another path.
Then, if the patient is young, you may ask: ‘Did the syncope occur at rest or during exertion?’. If the answer is ‘at rest’, you might ask another question, while if it occurred during exertion, you might decide to perform, for example, a diagnostic test. This is how a DT works: you ask a series of questions, and with each answer, you follow a branch that leads you to a decision.
Now, an RF works differently. Imagine you have many colleagues (representing multiple individual trees) helping you determine the cause of the syncope. Each of them has access to the same patient data and may ask similar questions, but at different moments (i.e. at different points in their tree) or may weigh information differently.
This built-in diversity leads to different choices each time. In the end, each colleague provides their opinion on the cause of the syncope, and you follow the suggestion of the majority to arrive at a more robust decision.
In contrast, GBM works in a sequential manner. Imagine that, instead of asking all your colleagues at once, you consult them one by one. The first colleague makes a prediction, and you note any errors they make.
Then, the second colleague looks specifically at those errors and refines the prediction, focusing on where the first colleague made mistakes. The process continues with each subsequent colleague improving on the errors of the previous ones, acting as an expert in their own specific domain, gradually refining the final diagnosis.
Neural network
A way to implement ML is the use of neural networks (NNs). These systems rely on artificial ‘neurones’ that, to some extent, replicate the behaviour of neurones in the human brain. Neurones are organized into interconnected layers, forming a network that takes an input stimulus and generates a corresponding output.10
Neural networks are composed of linear operations followed by nonlinear activation functions, which enable them to model complex, non-linear relationships in data. While linear regression aims to fit a simple linear relationship between variables to make predictions, NNs extend this framework by stacking multiple layers of nonlinear transformations, allowing the learning of hierarchical representations and complex patterns that cannot be captured by linear models alone.11
Neural networks are organized into layers of nodes. Each node functions like an artificial neurone and connects to many other nodes across multiple layers. There are three main types of layers: the input layer, hidden layers, and the output layer (Figure 3).
During training, the NN models non-linear relationships and complex feature interactions across nodes. The network ‘learns’ by adjusting its connection weights to minimize prediction error, thereby improving its ability to recognize complex patterns.
Supplementary material online, File S2 (related to Figure 3)
Example of neural network
Let us explore how a deep NN (DNN) functions using the task of determining whether an ECG indicates sinus rhythm (SR) or atrial fibrillation (AF). As illustrated in Figure 3, the network analyses the data in a progressive, hierarchical manner, moving from raw signal data to an abstract diagnosis.
Unlike traditional diagnostic methods where a human manually measures specific intervals (like P-wave presence or QRS duration), a DNN operates directly on the raw signal. As shown in the input layer of Figure 3, the continuous graphical ECG waveform is discretized into a long vector of numerical voltage values sampled at specific time points (t1, t2, t3, …). Each individual raw voltage reading serves as an input node (x1, x2, x3, … x_N) for the network.
The network receives the exact shape of the signal without any prior human interpretation.
The raw data progresses through multiple hidden layers. Through training on massive datasets, these layers learn to automatically extract informative patterns without being explicitly programmed to look for clinical features.
Finally, the output layer consolidates the highly abstract information from the final hidden layer to generate a probabilistic classification. Instead of a simple binary ‘yes’ or ‘no’, the network assigns a probability score to each possible class. In the example in Figure 3, the network outputs a probability of 0.90 (90%) for SR and 0.10 (10%) for AF, indicating high confidence in the SR diagnosis based on the abstract mathematical patterns it detected in the raw signal.
Deep neural networks
A DNN is characterized by the presence of many HLs between the input and output layers. Each HL is made up of units, or nodes that process the information received and pass it on to the next level. Having multiple layers allows the DNN to capture complex relationships and represent data that follow intricate patterns.
Convolutional neural networks
Convolutional neural networks are a type of NN particularly suited to data with a spatial structure, such as medical images. In a convolutional layer, learnable filters are applied across local regions of the input to generate feature maps. Early layers typically capture relatively simple patterns, such as edges and contours, whereas deeper layers combine these signals into more abstract, task-relevant representations.
Pooling layers can progressively reduce the spatial dimensions of the feature maps, thereby reducing computational complexity while retaining salient information. In the simplified image-classification architecture shown in Figure 4, the final feature maps are flattened and passed through fully connected layers that produce probabilities for predefined image-level classes. Convolutional neural networks can also be configured for image segmentation; in that setting, the network generates a pixel-wise label map rather than a single image-level classification.
Transformer architecture and attention mechanisms
Transformers represent a paradigm shift in NN design. While traditional architectures like recurrent NNs or long short-term memory process time-series data sequentially (one beat after another), transformers use ‘attention mechanisms’ to evaluate relationships between all parts of the input simultaneously.12,13 This overcomes the ‘short-term memory’ limitations of sequential models. The key innovation is self-attention: when analysing an ECG, the model can directly compare the P wave characteristics to QRS patterns across the entire trace, rather than processing sequentially.14,15 This allows capturing long-range dependencies—for instance, subtle rhythm irregularities spanning multiple beats—that traditional architectures might miss.
In cardiology, transformer-based models have shown promising applications in ECG interpretation and multimodal data integration (see the Generative artificial intelligence applications section), where they can support the identification of patterns across different data types, including imaging, ECG, and clinical variables, by learning which features to ‘attend to’ jointly.
Contrastive learning
Contrastive learning is a self-supervised technique where models learn by comparing similar and dissimilar examples, rather than requiring explicit labels for every case. The model learns to map similar cases (e.g. two ECGs from the same patient, or ECGs showing similar arrhythmias) close together in a learned representation space, while pushing dissimilar cases apart.
This approach dramatically reduces annotation burden—instead of labelling thousands of specific abnormalities, the model learns general representations from the structure of the data itself. In cardiovascular imaging, contrastive learning enables models like EchoCLIP (see the Generative artificial intelligence applications section) to learn from vast unlabelled echocardiogram datasets, then adapt to specific diagnostic tasks with minimal labelled examples. This addresses a critical bottleneck in medical AI: the scarcity of expert-annotated data.16
Machine learning applications
Classification in ML involves categorizing patient data into predefined groups, aiding in diagnosis and treatment planning. Techniques such as DTs and NNs analyse features like clinical history, ECG or imaging data to detect cardiovascular risk and stratify patients accordingly. Recent advancements improve accuracy in cardiac risk assessment and disease detection, enhancing personalized medicine.17,18
In contrast, regression in ML estimates continuous outcomes from input data, vital for assessing cardiovascular risk and treatment effects. Techniques like linear regression and advanced methods such as RFs model relationships between variables, including patient demographics and clinical measures. Recent examples show the utility of ML to estimate blood pressure or serum cholesterol levels.19–21
Techniques like classification and regression can both be used to forecast or predict outcomes, like heart failure or arrhythmias. Recent advancements improved predictive accuracy, facilitating early intervention and personalized treatment strategies.18,22–25
Machine learning can also be useful to manage unprocessed data. For example, natural language processing extracts relevant information from unstructured data, such as electronic health records (EHRs), clinical text and physician notes, facilitating risk assessment and decision-making. Recent advancements may improve accuracy and efficiency, supporting personalized treatment approaches and generating research insights.26,27
Moreover, large language models (LLMs) leverage vast datasets to understand and generate human-like text. They assist in automating documentation, summarizing patient records, and enhancing communication. Recent advances in LLMs have shown potential for clinical decision support and patient engagement, with possible benefits for workflow efficiency and care delivery.28,29
Critical model aspects
Model training
Model training is the process through which an ML model learns patterns or structures from data. The exact nature of this process depends on the learning paradigm: in supervised learning, the model learns from labelled examples; in unsupervised learning, it identifies patterns or structures without predefined labels; and in reinforcement learning, it learns through actions and feedback from the environment.
During training, the model progressively recognizes relationships within data, with the goal of generalizing such findings and effectively analysing new data. In cardiology, for example, a supervised model may be taught to analyse ECG recordings to learn signs related to a certain condition, such as recognizing P-wave patterns in patients affected by paroxysmal atrial fibrillation (PAF) during SR.
When models are developed and evaluated using repeated measurements from the same patients, a critical consideration often overlooked is that data splitting should occur at the patient level, not at the individual data point level.
If multiple ECGs, imaging studies, or clinical visits from the same patient appear in both training and test sets, the model may memorize patient-specific patterns rather than learn generalizable disease features. This ‘data leakage’ produces artificially inflated performance metrics that collapse when deployed on truly unseen patients.
For example, if a patient with AF has five ECG recordings—three in training, two in testing—the model may recognize that patient’s unique baseline characteristics rather than learning AF detection generally. Published studies showing exceptional performance without patient-level splitting warrant scrutiny before clinical adoption.30
Accuracy and evaluation
Accuracy defines the diagnostic performance of the model, measured in terms of true results (positive and negative) among examined cases. The evaluation process consists of assessing the model’s performance by using it on new data (data not used to train the model in the first place) and measuring its accuracy. As an example, the aforementioned model may be exposed to new ECGs and assessed in terms of accuracy in correctly predicting which patients are affected by PAF based on P wave analysis during SR.
Explainability
In ML-based AI systems, explainability refers to the ability to interpret a model's internal decision-making process. However, cardiologists must recognize a critical distinction: current explainability methods (SHAP, LIME) identify associations between input features and predictions, not causal relationships.
For example, an explainability tool might highlight that a specific ECG segment heavily influenced an arrhythmia prediction. However, this indicates correlation—the model found this pattern statistically associated with the outcome in training data but does not establish that this ECG feature causally produces the arrhythmia. The model may be capturing proxy markers, confounders, or spurious correlations.31
This limitation has critical clinical implications. Explainability methods indicate what the model pays attention to, rather than what it truly understands; consequently, the identified features may reflect biases inherent in the training dataset rather than underlying pathophysiology. Moreover, explanations based on associations can be misleading when applied to individual patients, particularly those belonging to populations that are underrepresented in the training data.
Therefore, although explainability tools provide valuable insights for model debugging and hypothesis generation, they cannot replace clinical reasoning nor confirm that a model has learned physiologically meaningful relationships. As such, even explainable models require rigorous clinical validation before deployment.
Robustness
Robustness is the ability of a model to deliver acceptable responses even when input data contain unexpected variations, errors, noise or characteristics that were not encountered during the training phase. A robust model can correctly analyse data coming from different sources and pertaining to different populations, with a wide degree of variability.
Preprocessing
Preprocessing, in the initial phase of ML analysis, is the process where raw data are organized, cleaned, and filtered in order to make it feasible for further processing by the model. Effective preprocessing enables reproducible and reliable results, by removing noise and artefacts, standardizing data coming from different sources, and converting it to desired formats.
Bias and variance
Bias and variance describe two statistical sources of prediction error. Bias refers to systematic error introduced when a model is too simple to capture the underlying relationships in the data. High bias may lead to underfitting.
Variance refers to the sensitivity of a model to fluctuations in the training data. High variance may lead to overfitting, because the model captures noise or dataset-specific patterns rather than generalizable information. Balancing bias and variance is therefore essential to allow a model to generalize reliably to new data.
Overfitting and underfitting
Overfitting and underfitting describe the observed behaviour of a model. Overfitting occurs when a model performs well on the training data but poorly on new, unseen data because it has learned noise, artefacts, or dataset-specific patterns. Underfitting describes the opposite condition, in which the model is too simple or has been insufficiently trained to capture relevant patterns.
As a result, it performs poorly on both the training data and new data and may fail to identify features associated with a specific condition, potentially leading to missed diagnoses.
Machine learning methodologies
How does supervised machine learning work?
Supervised ML involves training algorithms on labelled datasets, where each input is paired with a known output. The process begins with data collection, including features such as patient demographics and clinical measurements, such as ECG or imaging features. The algorithm learns to identify patterns and relationships within this data, optimizing its parameters to minimize prediction error.32
During supervised ML development, data are usually divided at the patient level into development and evaluation sets. The training set is used to fit model parameters, such as weights or coefficients. A separate validation, or tuning, set is then used during iterative development to select preprocessing strategies, features, model architecture, hyperparameters, regularization, thresholds, and calibration procedures.
Because the validation/tuning set influences model development, performance on this set should not be interpreted as an unbiased estimate of final model performance. Once model development is complete, a locked internal test set should be used once to estimate performance on held-out data from the same development environment. If the model is further modified after inspecting the internal test results, a new unbiased test set is required.
Finally, external validation evaluates the final locked model in an independent dataset, ideally from different centres, time periods, devices, or populations, and provides a more stringent assessment of generalizability.
In cardiology, supervised ML can support diagnostic and prognostic tasks, including the prediction of heart failure or arrhythmias from ECG, imaging, or clinical data. However, most studies demonstrate model discrimination, calibration, diagnostic yield, or workflow effects rather than improved patient-centred outcomes. A small number of pragmatic randomized studies of AI-enabled ECG have reported improved detection of low ejection fraction in primary care or reduced 90-day mortality in hospitalized patients using an AI-ECG alert system, but these findings are context-specific and should not be generalized to ML tools as a class.
Prospective evaluation of clinical utility, safety, workflow integration, and impact on patient outcomes is therefore required before routine implementation.33,34
What is the difference between classical statistical models and supervised machine learning?
Classical statistical models, such as linear and logistic regression, focus on understanding relationships between variables and making inferences based on underlying assumptions about data distribution. These models often require a priori hypotheses and rely on a limited set of predictors, emphasizing interpretability and statistical significance.35
In contrast, supervised ML uses algorithms that learn patterns from labelled datasets, allowing for complex interactions among numerous predictors without strict assumptions about data distribution. This approach generally has improved accuracy, particularly with large, high-dimensional datasets, common in cardiology.36,37 For instance, ML techniques can identify risk factors for cardiovascular events more effectively than traditional models.17,38
While classical models often prioritize inference and estimation of interpretable associations, supervised ML primarily emphasizes prediction and adaptability. These models may support more individualized risk stratification, but their effect on treatment decisions and patient outcomes requires prospective clinical evaluation.
What is the difference between supervised and unsupervised machine learning?
Supervised and unsupervised MLs are two primary paradigms in AI that differ in the type of data they use and the problems they aim to solve.
As explained before, in supervised learning, the model is trained on labelled datasets, where inputs (features) are paired with known outputs (labels). The objective is for the model to learn the relationship between inputs and outputs so that it can predict outcomes for new patients. For instance, a supervised learning model could analyse datasets containing patient demographic information, blood biomarkers, and imaging features to predict outcomes such as the presence of left ventricular dysfunction.
Techniques like logistic regression (often used for binary classification problems), RFs (one of the most popular ensemble learning methods suitable for both classification and regression tasks), and DNNs (commonly used for complex supervised learning problems involving large datasets) are commonly used in this context.
A concrete example is the application of supervised CNNs to interpret echocardiographic images for detecting structural heart diseases, such as hypertrophic cardiomyopathy, with high diagnostic accuracy.39
In contrast, unsupervised learning uses datasets without labelled outputs. The goal can be to discover hidden patterns or structures in the data without explicit guidance. This method is particularly suited for exploratory analyses.
For example, clustering algorithms such as k-means or hierarchical clustering (discussed later) could group patients with heart failure based on clinical characteristics, biomarkers, and imaging data to identify subgroups with different risks of hospitalization or mortality. Similarly, dimensionality reduction techniques like principal component analysis (PCA) can be employed to condense high-dimensional data, such as multi-omics profiles, into fewer dimensions, facilitating insights into disease mechanisms or risk stratification.
The distinction is clear: supervised learning aims to predict specific outcomes based on labelled data, while unsupervised learning seeks to discover hidden structures within unlabelled datasets. Both approaches complement each other, offering predictive power and exploratory insights, respectively.
How does unsupervised machine learning work?
Unsupervised ML is designed to extract meaningful patterns and relationships from unlabelled data, making it invaluable in cardiology, where datasets often contain complex, multifaceted information without predefined outcomes. Two commonly used approaches are clustering and dimensionality reduction.
Clustering
Clustering methods group patients based on shared characteristics, allowing for the identification of phenotypic subgroups. Examples in cardiology include:
Dimensionality reduction
High-dimensional datasets, such as genomic or imaging data, can pose challenges related to computational efficiency, sparse features, and multicollinearity. Techniques like PCA reduce the complexity of these datasets while retaining critical information, improving computational efficiency and facilitating data visualization.
Examples include:
By revealing hidden structures in data, unsupervised learning provides cardiologists with tools to identify new disease phenotypes, refine risk stratification, and guide hypothesis generation. Although unsupervised learning does not predict outcomes directly, its exploratory nature may help identify clinically relevant patterns or patient subgroups that can generate hypotheses for future risk stratification and precision medicine research in cardiology. However, some unsupervised learning approaches—particularly more complex models—may be perceived as having a ‘black-box’ nature: while they can reveal hidden structures in data, the underlying decision process often remains opaque.
This limitation, with its implications for clinical interpretability, is discussed in more detail in the following paragraphs.
What is the difference between reinforcement learning and the previous models?
Supervised and unsupervised learning methods use fixed databases with or without labelled data, respectively, to train models to conduct specific tasks. However, many learning processes are not as simple as identifying patterns in the data. For example, how to walk takes months of trial and error for humans to learn.
Reinforcement learning is the third basic ML paradigm and is heavily inspired by biology to solve complex tasks within changing environments. Reinforcement learning consists of an agent which is allowed to take different actions within a dynamic environment and is rewarded based on its achievement (e.g. how long it was able to walk). This learning strategy allows for these dynamic algorithms to search a wide range of possible solutions and even explore sub-optimal ones, which may actually maximize the cumulative reward in the long term.
Through the exploration of new strategies and exploitation of learned knowledge, the trained agent will be able to adapt to new circumstances (e.g. walking on rock terrain) and constantly learn from its environment, resulting in a robust agent able to reliably achieve the required task.
Reinforcement learning implementations are limited in the medical field.45,46 Currently, research is focused on exploiting ML to achieve specific tasks using mostly cross-sectional data. Reinforcement learning works best when we can model a dynamic system and train these intelligent agents. However, human physiological systems are so complex that current simulations often are not precise enough.
Furthermore, acquiring vast amounts of longitudinal data is often not feasible, and ethically we may not be able to manipulate the environment to model its consequences (e.g. trying to reinforce the model by running potentially harmful tests in the clinical scenario). Also, even when multiple clinical variables are available, they may still be not sufficient to build robust simulation environments for algorithm development. Once we have the technology to map human physiological processes with sufficient detail in real time, we might be able to build digital twins which mimic human disease processes.
This would allow us to use reinforcement learning to train intelligent agents which can optimize therapeutic strategies.
Machine learning algorithms relevant to cardiology research
Supervised machine learning algorithms: advantages and disadvantages
Supervised ML algorithms have become indispensable tools in cardiology research, providing powerful methodologies for modelling intricate relationships within medical data. This section delves into three prominent supervised ML approaches—regularized regression, ensembles of DTs, and support vector machines (SVMs)—highlighting their strengths and limitations. Additionally, Table 1 summarizes representative scientific articles that have applied these techniques, either individually or in combination, to specific cardiology applications.
Regularized regression
Regularized regression techniques help prevent overfitting by introducing penalty terms to the loss function. These methods are particularly useful in datasets with a high number of predictors relative to observations, a common scenario in cardiology studies involving biomarkers or genomic data. Zou and Hastie highlighted elastic net’s ability to handle correlated features, a frequent challenge in analysing cardiovascular datasets55.
However, the interpretability of regularized models may decline as the degree of penalization increases, potentially obscuring clinically relevant variables.
Ensembles of decision trees
Tree-based ensemble methods like RFs and GBMs have gained prominence for their high accuracy and ability to handle heterogeneous data. Breiman’s RF algorithm leverages bootstrapped datasets and feature randomness to improve generalization, while XGBoost, as introduced by Chen and Guestrin, optimizes computational efficiency for large-scale datasets.56,57 In cardiology, these methods excel in tasks like ECG classification and risk stratification for arrhythmias. However, ensembles are often criticized for their ‘black-box’ nature, which can hinder clinical interpretability, despite recent efforts to integrate SHAP values for feature importance.
Support vector machines
Support vector machines are particularly well-suited for small-to-medium-sized datasets with high dimensionality, a common characteristic in personalized cardiology. Cortes and Vapnik’s foundational work established the theoretical underpinnings of SVMs, which have been used in cardiology for tasks like heart sound classification and arrhythmia detection.58 These models perform exceptionally well in separating classes using kernels to map data into higher-dimensional spaces. However, tuning hyperparameters such as the kernel type and regularization coefficient is critical, as inappropriate settings can lead to suboptimal performance.51
All these algorithms offer significant advantages in cardiology research by enabling the modelling of complex medical data, with immense potential for early diagnosis, personalized treatment, and improved patient outcomes. These algorithms are particularly effective for analysing high-dimensional datasets and identifying patterns that may not be readily apparent through traditional statistical methods. However, their implementation necessitates careful attention to several critical factors, such as balancing model accuracy with interpretability and addressing the computational demands required for processing large datasets.
Additionally, the application of these methods in cardiology presents challenges in ensuring the generalizability of models across diverse patient populations.
Ensuring the robustness of these algorithms also requires high-quality, well-curated datasets to prevent biases that could affect clinical decision-making and patient outcomes. Moreover, regulatory considerations, such as compliance with FDA and EU AI Act guidelines, are essential to ensure the safe and ethical deployment of AI-driven tools in cardiology. Addressing these aforementioned factors is critical for translating ML advancements into reliable, real-world clinical applications.
Unsupervised machine learning algorithms: advantages and disadvantages
As previously discussed, unsupervised learning does not rely on labelled data; instead, it autonomously identifies patterns and structures within data, making it a particularly valuable tool for data-rich but annotation-scarce domains like cardiology. By uncovering patterns without predefined outcomes, unsupervised learning helps reveal underlying structures or relationships among variables within a dataset, with key applications in cardiology: (i) clustering—for example, grouping patients by similar features, such as ECG patterns, to identify high-risk groups or subtypes; (ii) anomaly detection—detecting irregularities in clinical data, which can highlight early warning signs of cardiovascular conditions; and (iii) latent variable modelling—engineering features for low-dimensional data representations of complex data in cardiology, such as multi-modal or genetic data, that can be incorporated into supervised models or used to reduce data complexity.52,53
These techniques are especially relevant in cardiology as cardiovascular diseases are often heterogeneous, with progression influenced by multiple factors. Traditionally, diseases are studied as single entities, potentially overlooking critical differences among patient subgroups. Unsupervised learning could assist by detecting complex correlations in patient characteristics, thus enabling their further clustering into subgroups, allowing for personalized interventions.
For example, patients with similar combinations of clinical, imaging and laboratory characteristics, which could indicate a similar response to treatment, could be clustered together to optimize management.54,55
k-Means clustering
One method is k-means clustering, which partitions data into clusters by minimizing the distance between data points and their respective cluster centres. K-means is effective for straightforward clustering tasks, yet it can struggle with high-dimensional or non-linear data.56,57 Another technique, hierarchical clustering, organizes data into a tree-like structure of clusters based on similarity, often visualized as a dendrogram. This approach is especially useful in fields like genetics for its interpretability, as it enables researchers to observe cluster relationships at multiple levels.52
Finally, latent class analysis is a statistical method used for model-based clustering to identify subgroups or typologies within a population, particularly when the data consist of categorical or dichotomous variables (indicators). Latent class analysis is an unsupervised learning approach that assigns individuals to distinct classes based on similarities in their responses to specific items or variables. Each class is characterized by a set of probabilities that describe the likelihood of certain indicator values, capturing the unique features of that class.
This probabilistic modelling technique helps uncover hidden subtypes or patterns within heterogeneous data, making it a valuable tool for understanding complex populations and drawing statistical inferences.
Topological data analysis
Topological data analysis (TDA) is a newer approach that explores the ‘shape’ of data, capturing common structural patterns of similar entities, which may go unnoticed with traditional methods.58 Although promising, the implementation of TDA in cardiology is still less mature than some other unsupervised methods.
Principal component analysis, singular value decomposition and tensor factorization
For high-dimensional data, several dimensionality reduction techniques like PCA, singular value decomposition (SVD), and tensor factorization help reduce complexity by projecting data onto a lower-dimensional space. PCA and SVD are efficient for capturing linear relationships and reducing dimensions, while tensor factorization can handle and decompose more complex data into lower dimensions. This approach has proven valuable in integrating multiple cardiology data sources, such as EHRs, genetic and imaging data, as well as high-dimensional multi-modal records, but requires careful selection of algorithms and is computationally demanding.
Deep learning
Deep learning methods can support feature extraction and representation learning from complex, high-dimensional data. Among these, autoencoders learn compressed latent representations by reconstructing their input, making them useful for dimensionality reduction. Variational autoencoders extend this approach by learning probabilistic latent representations that can capture complex and non-linear relationships within the data.
However, they require extensive data and computational resources and often present interpretability challenges, which is a barrier to their widespread adoption in clinical practice.
In conclusion, unsupervised learning holds potential, particularly for tasks like patient and disease clustering, early detection through anomaly identification, and precision phenotyping. By subgrouping entities based on intricate combinations of characteristics, unsupervised learning supports personalized care pathways, advancing the goals of precision medicine in cardiology.59 This approach enables deep insights from unstructured, unlabelled cardiovascular data, uncovering patterns and relationships that traditional methods may overlook. Representative applications of unsupervised learning in cardiology are summarized in Table 2.
Critical aspects to consider
Strategies for assessing and controlling artificial intelligence quality and safety
Assessing and controlling the quality and safety of AI in cardiology is paramount for ensuring effective clinical outcomes and patient safety. Several key strategies should be implemented:
Rigorous validation and testing
Machine learning models are highly dependent on the datasets used for their development and therefore require thorough validation on diverse and representative data. A critical component of this process is independent external validation, which is essential for assessing whether a model truly generalizes beyond the population on which it was trained. External validation must include datasets with different characteristics (e.g. demographics, ancestry, disease status) to ensure robust performance across heterogeneous clinical settings.84–86
Evaluation metrics such as accuracy, sensitivity, specificity, and the area under the receiver operating characteristic curve are fundamental for assessing model performance. However, without strong external validation, these metrics may substantially overestimate true clinical utility. When a model does not demonstrate adequate generalizability, applying it to datasets or patient populations different from those used during training would require retraining and additional validation.
Transparency and interpretability
Clinicians must understand how AI systems reach their conclusions. Techniques like SHAP and LIME provide insights into model decision-making processes.63–65 This transparency fosters trust among clinicians and allows for effective integration of AI into clinical workflows.
Continuous monitoring and adaptation
AI models should be regularly monitored and updated to reflect new clinical data and evolving practices.66 This ensures models remain relevant and accurate over time. Continuous learning frameworks can help AI systems adapt to new information, thereby enhancing their predictive power.67,68
Ethical considerations
Ethical deployment of AI in healthcare is critical to avoid perpetuating biases and inequalities. Establishing guidelines for ethical AI use can help mitigate risks and ensure equitable access to AI-driven solutions.69,70
One of the main challenges is bias in AI models, which can arise from imbalanced or non-representative training datasets, leading to disparities in diagnosis and treatment recommendations. Implementing fairness-aware algorithms, adversarial debiasing techniques, and continuous auditing of AI systems can help reduce these risks. Regulatory frameworks, such as the EU AI Act and the FDA’s AI/ML-Based Software as a Medical Device (SaMD) framework, provide essential guidelines for ensuring AI’s transparency, accountability, and patient safety.
Adherence to these regulations is necessary to maintain ethical AI implementation in cardiology and beyond.71
Additionally, enhancing AI explainability through interpretability tools such as SHAP and LIME fosters trust among clinicians and patients. Transparent AI decision-making not only improves adoption in clinical practice but also ensures that ethical considerations, such as informed consent and patient autonomy, are upheld in AI-assisted healthcare.
Multidisciplinary collaboration
Engaging a multidisciplinary team that includes data scientists, clinicians, ethicists, and regulatory experts can enhance the development, assessment, and implementation of AI technologies, ensuring a comprehensive approach to quality and safety.72–75
By adopting these strategies, cardiologists can critically assess AI tools, minimize associated risks, and support safer, more evidence-based implementation in clinical practice.
Key considerations when selecting an artificial intelligence model
For cardiologists, the aim of this section is not to provide technical instructions on how to develop an AI model, but to clarify the main principles that should guide the interpretation of AI studies and the collaboration with data scientists.
Choosing the right AI model has a major impact on the results, both in terms of accuracy and robustness, and depends on factors such as the amount and characteristics of the available data, as well as the specific task that the model is intended to solve. These aspects should therefore be discussed with data scientists at the beginning of the research project or model development process.
While NNs have received considerable attention recently and represent the state of the art for high-dimensional data, such as images, text, or signals, models based on DTs may still perform better for the analysis of tabular data.76,77 Apart from choosing the family of methods to use, it is important to choose the complexity, which is roughly the number of parameters. This can vary from the order of 10 s for a logistic regression, to the order of 109 in the case of the largest NNs. If the problem involves highly non-linear interactions, then more complex methods can be used; however, this has to be pondered with the quantity and quality of data available for training, and the need for interpretability.78,79 A good compromise is available in the form of foundation models, which are NNs trained on large datasets for a variety of tasks involving a given type of data.
Using them as a base component that can be adapted to a specific problem often yields good results; this process is called fine-tuning. Also, for some applications where high interpretability is required, simpler models that are more interpretable can be preferred, as they are also less prone to overfitting.80–82
Human oversight, artificial intelligence design and model transparency
The successful integration of AI in cardiology requires maintaining appropriate human oversight while ensuring transparency in both design and operation. Artificial intelligence systems should be designed to augment, not replace, clinical decision-making, with clinicians maintaining ultimate responsibility for patient care decisions. This requires AI systems with built-in explainability features that allow clinicians to understand the reasoning behind predictions when possible.
If models are inherently unexplainable, model limitations need to be clearly documented, and appropriate safeguards put in place (most notably a ‘human in the loop’). The development process itself must be transparent, including clear documentation of training data characteristics, potential biases, and validation procedures. Such transparency and explainability are critical to build trust.83 A robust governance framework should establish protocols for human oversight, including regular audits of AI performance and clear procedures for handling system errors.
The European Union’s AI Act incorporates many such requirements. Equally important is user-centred design, with intuitive interfaces that facilitate effective human–AI interaction and clear override mechanisms for clinical judgment.
Generative artificial intelligence
Overview
Generative AI (GenAI) is a subset of AI capable of creating new content. Unlike traditional AI, which primarily analyses existing data, GenAI can generate text, images, and even molecular structures. While generative approaches encompass various architectures, including generative adversarial networks and variational autoencoders, recent advances have been driven largely by foundation models: large-scale, pretrained models capable of adapting to a wide range of downstream tasks with minimal additional refinement (called fine-tuning).
These foundation models rely on an NN architecture called transformers, which allows them to handle vast amounts of data and learn complex patterns efficiently. Examples of foundation models include LLMs like GPT and Claude, and vision models such as DALL-E and stable diffusion.
These models can perform tasks they were not explicitly trained for, including learning from just a few examples (few-shot learning) or none at all (zero-shot generalization), potentially transforming various aspects of medical practice and research.104 A key advantage of foundation models is their ability to learn from unlabelled or minimally labelled data through self-supervised learning techniques.84 This significantly reduces the need for extensive manual annotation, which has been a major bottleneck in developing AI models for specialized medical tasks, and contributes to their rapid development and broad applicability.104 Recent progress includes development of specialized medical foundation models and multimodal models, which combine modalities (text, images, ECGs).28
Applications
Although still relatively new, GenAI and foundation models are investigated for various aspects of cardiovascular care, from diagnostic and therapeutic applications to administrative tasks and patient engagement.
Challenges and implementation strategies
These applications demonstrate the potential of AI, including GenAI and foundation models, to transform cardiovascular care across a spectrum of activities. However, it's important to note that while many of these AI applications show promise, only a limited number have been prospectively evaluated in clinical practice. The successful integration of AI into mechanistic research and routine clinical care will require careful validation, consideration of interpretability, and ongoing evaluation of its impact on patient outcomes.
Data quality, privacy, and security are paramount concerns when handling sensitive patient information. GenAI models are also prone to ‘hallucinations’: generating plausible but factually incorrect information. This could lead to inappropriate clinical decisions if not properly monitored.
The ‘black box’ nature of some GenAI models raises critical questions about interpretability and accountability, making it difficult for clinicians to understand how AI-generated insights were derived.
Regulatory and legal frameworks, such as the recent European Union’s AI Act, are evolving to address the use of AI in healthcare, but uncertainty remains regarding legal accountability for AI-influenced decisions. To mitigate these risks, ‘Human-in-the-Loop’ systems and explainable AI techniques are necessary to ensure clinician oversight of AI-generated outputs.
Initial implementation strategies will likely focus on low-risk, high-return applications, such as summarizing patient interactions and reducing healthcare administration burden, and can build trust in using GenAI for more challenging, clinic-facing applications. Active engagement of the cardiology community in the development and implementation of GenAI technologies is crucial to ensure their effective and responsible use in improving cardiovascular care and advancing research.
Limitations and regulations
The EU AI Act is a pivotal regulation for digital medical devices, including AI/ML-based tools in cardiology, such as diagnostic systems, clinical support, and software as medical devices (SaMD). The Act classifies AI systems in medical products as ‘high-risk’, imposing significant compliance obligations, especially regarding risk management, which must align with the European Medical Device Regulation (MDR) frameworks, as these systems are primarily governed by the MDR, which imposes strict compliance requirements to ensure safety and data protection. Given their frequent processing of sensitive health data, they must also comply with GDPR standards.
With extraterritorial scope similar to GDPR, the EU AI Act applies to all AI providers in the EU, irrespective of geographic location, impacting global manufacturers, including those with FDA-approved AI/ML devices. Data governance complements GDPR requirements but adds AI-specific provisions, such as data set validation and bias assessment.
The initial step in assessing AI systems is determining whether they qualify as medical devices under MDR, using the criteria outlined in the guidelines proposed by the MDCG. Following this, risk classification is conducted, critical for establishing appropriate regulatory pathways. It is important to understand that both high risk and non-high risk medical devices according to MDR are classified as high risk according to the Act.
It has to be noted that currently there is a notable lack of evidence regarding the safety and performance of the European regulatory system for AI/ML-based devices. This issue is compounded by the absence of a public registry of approved devices, confidentiality restrictions, and the decentralized CE-marking process. Although the forthcoming Eudamed database promises some transparency, its implementation is delayed.
The USA faces similar transparency issues. Improving transparency in both regions would enhance the evaluation of AI/ML devices, fostering better safety, efficacy, and insights from regulatory experiences.
It is also important to recognize that regulatory frameworks in both the EU and the USA are evolving rapidly; therefore, any recommendations may require updates as the implementation of the EU AI Act progresses and new guidance from the FDA continues to develop.
Beyond the principles outlined here, issues such as structured risk–benefit appraisal, GDPR compliance, and cybersecurity require detailed governance frameworks, which are beyond the scope of this statement but are addressed in dedicated ESC and EU policy documents.
Finally, international regulatory harmonization will be necessary as AI medical devices often require multi-region approval (e.g. MDR in Europe, FDA in the USA). Global manufacturers must navigate differing compliance requirements, underscoring the need for international regulatory collaboration to streamline AI certification and approval pathways.
Conclusion
The significant reduction in data analysis costs, the availability of massive datasets (marking the era of big data), advancements in computing power, and improvements in algorithms and ML techniques have made AI pervasive also in the medical field. This shift has led to a new form of research driven by ‘big data’ and diverse datasets, surpassing traditional statistical analysis. In this scientific statement, we aimed to present an educational overview of AI in cardiology, highlighting the essential knowledge cardiologists need to acquire AI literacy and the understanding required to navigate this emerging domain of cardiovascular research.
This statement highlights the need for clinicians, even non-experts, to become familiar with basic AI concepts, its potential benefits, and its limitations. Looking ahead, we envisage three priorities: (i) strengthening education and training so that cardiologists acquire essential AI literacy; (ii) promoting research that clarifies clinical impact, fairness, and safety; and (iii) fostering collaboration among clinicians, patients, researchers, and regulators to ensure that innovation remains patient-centred.
Supplementary Material