L'intelligenza artificiale nella diagnosi dell'infarto del miocardio: il ruolo del machine learning (apprendimento automatico)

Abstract

Artificial intelligence (AI) comprises computational methods that can perform complex tasks typically associated with human capabilities; Machine learning (ML) is a subset of AI that make models capable of learning from data and generalizing to new cases. The diagnosis of myocardial infarction (AMI) requires a combination of clinical data, electrocardiogram (ECG), and high-sensitivity troponin (hs-cTn). Current diagnostic strategies, particularly ECG in the diagnosis of STEMI (ST-elevation myocardial infarction) and accelerated diagnostic algorithms in the diagnosis of NSTEMI (non-ST-elevation myocardial infarction), still have significant limitations. The aim of this brief review is to identify the potential contribution of ML to improving the diagnosis of AMI. Through the bibliographic search of systematic reviews (SR) and systematic searches on PubMed (machine-learning for ECG diagnosis of STEMI and machine-learning for NSTEMI diagnosis) 2 ML contributions were identified - AI-ECG (AiTiAMI/ROMIAE study) and CNN-LSTM in ambulance - for the diagnosis of STEMI, based on ECG and 3 ML applications for the diagnosis of NSTEMI based on hs-cTn (MI3; CoDE-ACS; ARTEMIS-POC). The reviewed studies have shown how different ML algorithms can improve the ability to identify patients with suspected AMI, increasing the accuracy both in excluding and confirming AMI, even in the presence of atypical or confounding symptoms and complex or atypical ECG patterns. This translates into greater efficiency in the healthcare system with optimization of resources, reduction of unnecessary hospitalizations and faster referral to the hemodynamic unit. Furthermore, some AMI diagnostic algorithms are able to simultaneously offer prognostic assessments by estimating the risk of adverse events, intra- and extra-hospital mortality and the need for revascularization. They can thus support clinical decision-making, especially in complex or ambiguous cases, and potentially reduce interobserver variability. However, methodological risks remain, such as the large amount of training data and their “neutrality”; the risk of overfitting, with a reduced ability to generalize results across different contexts; the presence of selection bias, as most studies come from highly specialized university centers; and the need for prospective multicenter validation and external validation on heterogeneous populations. From a technical point of view, some challenges emerge: the vulnerability of neural networks to small perturbations (the black box problem); models trained on “ideal” data but applied in less standardized real-world clinical settings and the heterogeneity of datasets, which limits generalizability to different populations (ethnicity, healthcare systems, peripheral settings). In terms of clinical implementation, it is focus on practical integration with established ESC pathways, so that AI can serve as a complement rather than a substitute. Finally, on an ethical and regulatory level, significant issues remain: the management of privacy, confidentiality, and security of healthcare data; the definition of clinical liability in the event of AI-supported diagnostic or decision-making errors; the need for adequate legislative and regulatory frameworks; and the risk of reducing physicians’ decision-making autonomy. In summary, ML applications to the diagnosis of AMI appear promising and offer potentially significant advantages in terms of accuracy, timeliness, and operational efficiency. However, without systematically addressing methodological limitations, technical challenges, implementation issues, and ethical and regulatory implications, their clinical impact risks remaining marginal and confined to experimental experiments.

Ähnliche Arbeiten