Personalized medicine and artificial intelligence

The concept of PM is to devise a care plan tailored to each patient’s unique characteristics [24, 25]. Advances in AI have enabled the detailed analysis of patient data, thereby providing a more efficient and individualized approach to medical care [9, 26]. AI algorithms, particularly machine learning (ML) and deep learning (DL), are widely used to improve diagnostic accuracy by learning patterns from past patient records and medical images. Furthermore, DL models such as CNNs have shown excellent results in image-based medical applications, including disease detection and classification [27]. These models help clinicians analyze patient-specific data, which allows them to initiate appropriate treatment early.

Integrating AI and radiogenomics is very useful as it links the genotypic features with the phenotypic ones [28]. Figure 1 depicts the systematic workflows for radiogenomics. The first step is the patient data collection, which includes the radiological images and molecular data [29]. This is followed by radiogenomic analysis, where the AI extracts the most important features and links them with genomic alterations [30]. These correlations point to clinically significant biomarkers that help categorize patients based on risk, disease severity, or anticipated treatment response [31]. This method makes it possible to more precisely customize treatment plans to the distinct clinical features of each patient.

Display full size

Figure 1. Radiogenomics and personalized medicine workflow.

In addition to diagnosis and treatment planning, AI has become an important component of drug discovery and development [32, 33]. It reduces development time and cost while increasing the success rate. While big data technologies and computing resources are becoming more affordable, it would be easier to develop a customized digital model, such as a “digital-twin,” that will mimic the human body and allow doctors to test all potential interventions and treatments before recommending them to actual patients [34–36]. Moreover, by using explainable AI (XAI), the interpretations of the AI-driven outcomes and decisions are enhanced [37, 38]. It helps the doctors to evaluate the reliability of predictions and support informed decision-making rather than replacing clinical judgement.

AI is also being incorporated into robotic platforms, virtual assistants, and monitoring tools to support healthcare delivery [39, 40]. These technologies assist with remote patient monitoring, patient education, and guided therapeutic interventions, including targeted drug discovery. Overall, the integration of AI and PM enhances evidence-based clinical decisions, improves patient outcomes, and has the potential to reduce the burden on the healthcare systems.

CVDs and radiogenomics

After the genomic revolution in the early 1990’s, oncology researchers have focused on investigating the fundamental causes of diseases at the genetic level to enable personalized treatments. CVDs are one of the most difficult diseases to diagnose because of their multifactorial nature, and also, the symptoms appear late during the disease progression [41, 42]. Conventional ways of genetic analysis rely on invasive biopsies or post-operative pathological tissues to perform the procedure; therefore, they cannot be applied to every patient [21, 43]. As we enter the next era of PM and big data, a large number of research studies are being carried out in the field of “Radiogenomics” [44]. It is the high-throughput extraction of image features from medical imaging. It allows data to be collected and used in a clinical decision support system to increase predictive, prognostic, and diagnostic accuracy [45]. Figure 2 depicts the stages in radiogenomics for CVD. The radiogenomics-CVD-based workflow consists of the following four steps: (i) First is multimodal data acquisition and preprocessing. In this step, multimodality imaging technologies such as CT, MRI, and PET are used to capture different physical and chemical properties of tissues. These radiomics features provide a multidimensional view of the CVDs [31, 46]. (ii) Second, feature extraction and multi-omics integration. In this step, high-dimensional feature data is extracted to quantitatively determine the characteristics of ROI [47]. After this, the data are fused with multiple omics data, such as proteomics, clinical, genomics, and transcriptomics, to accurately model clinical outcomes in patients with CVD [48, 49]. (iii) Third, computational modelling and validation. In this step, a DL model is trained and validated to link radiogenomic features with the symptoms of heart disease. They map gene regulatory networks to uncover the molecular relationships that underlie CVD development [50]. To maximize the performance of these algorithms, various cross-validation techniques like k-fold validation are used [51]. (iv) Last is the clinical translation and personalized therapy. Here, the outputs generated in the above-mentioned steps are incorporated into the clinical decision support systems so that they can assist the doctors during real-world decision-making [52]. This also helps in customized therapy planning of the patient, pharmacogenomics, and generating the polygenic risk scores for the early detection of CVDs [53, 54].

Display full size

Figure 2. Four-stage framework for radiogenomics in CVD.

Literature search criteria

The search was made till November 1, 2025, and only studies published within the last 10 years were considered in this review. Figure 3 demonstrates the distribution of studies per year. The count of research papers has varied gradually with time, showing a sudden increase in 2024, indicating that the amount of research is increasing in this field. The key terms used for the search are: “CVD”, “CVD using DL”, “CVD risk assessment using DL”, “AI”, “Radiogenomics”, “AI and radiogenomics”, “Radiomics”, “Genomics”, “PM in CVD”, “AI in the framework of PM”, “Precision medicine”, “Medical imaging”, and “Bias in AI”. This study employed a combination of search criteria by applying Boolean operators like “AND”, “OR”, and “NOT”. This process yielded 860 research articles, of which 80 were deemed suitable according to the selection criteria established for the current study.

Display full size

Figure 3. Year-wise distribution of studies included in the review.

Selection criteria

Our study utilized the published research papers written or translated into English. The inclusion and exclusion criteria used in our investigation are shown in Table 1. We omitted the non-research papers, such as magazines, theses, and dissertations. The final pool of publications focused on AI applications for examining CVDs, incorporating radiogenomics and PM.

Applications of AI in CVD

To put the findings discussed in the review in context, it is essential to examine how recent research has applied AI across a range of CVD use cases. These studies differ significantly in terms of AI model, study design, and clinical goal. Key AI-driven CVD investigations are compiled in Table 5 to enable an organized comparison of their goals, approaches, and published results. Lekadir et al. [56] developed a CNN to determine the proportion of lipid core, fibrous tissue, and calcified tissue in carotid plaque from ultrasound images. The CNN outperformed SVM-based methods with pixel-level accuracy of 0.75 ± 0.16 and correlation with expert area measurements of 0.87–0.93 across tissue types using about 90,000 picture patches from 53 in vivo examples. These findings show that DL can enhance CVD risk stratification and early detection of plaques at risk of rupture. Attia et al. [57] trained a CNN model using paired 12-lead ECG and echocardiogram data to identify asymptomatic left ventricular dysfunction (ALVD). Patients with normal ventricular function at the start who tested positive were four times more likely to develop ventricular dysfunction later. When ECG is combined with AI, it can serve as a powerful screening tool for detecting ALVD in asymptomatic individuals. They further extended their study to identify cardiac dysfunction using a smartwatch’s single-lead ECG [58]. This approach worked even in non-clinical and real-world setting demonstrating the potential of wearable technology to detect asymptomatic heart problems and enable early intervention. Heo et al. [78] used and compared ML techniques to predict the long-term outcomes in ischemic stroke patients. The model that used a deep neural network (DNN) considerably outperformed the traditional Acute Stroke Registry and Analysis of Lausanne (ASTRAL) score (AUC 0.839; P < 0.001), with an AUC of 0.888, whereas ML algorithms, random forest (RF), and logistic regression (LR) models performed similarly to ASTRAL. McGilvray et al. [62] developed an ensemble DL model using longitudinal EHR data to predict which HF patients are at risk of severe decompensation or non-response to medical therapy. The model achieved an AUC of 0.91, which shows high prediction performance. This would spare the rest of the individuals from the associated risk, expense, and unnecessary high-risk procedures. Wang et al. [59] used CMR images to automate screening and diagnosis of 11 types of CVD across several Chinese facilities. They combined cine and late gadolinium enhancement (LGE) MRI with the AI models that have utilized a two-staged DL approach with video-based Swin Transformer models. In fact, this methodology demonstrated robust generalizability across some institutions and outperformed experienced cardiologists in some diagnoses. This outlines the potential of AI to improve the accuracy of CMR interpretation. Furthermore, in a recent study by Saikumar and Rajesh [60], the RCNN integrated with the Adaptive Random Forest (ARF) model outperformed conventional ML techniques in identifying heart disease from CT/MRI images, achieving 99% accuracy, enabling early detection of overlooked cardiac problems. Moreover, recent advances in DL have enabled high-accuracy prediction of CVD events from ECG images. To forecast four heart diseases, researchers in a study by Hasan et al. [61] developed a lightweight CNN model alongside conventional weighted classifiers, including GNB, XGBoost, SVM, DT, RF, and LR. With an incredible 99.29% accuracy, the ensemble model outperformed baseline techniques, while the CNN model extracted 100 significant features per image.

Limitations

While AI offers many benefits for the healthcare sector, it also poses significant challenges. First, when imaging and genomic data are combined, they generate complex, heterogeneous datasets. To trace patterns from such a scarce dataset requires very high-end algorithms, which increase computational cost [69]. Second, an individual’s medical data is highly confidential and must be protected against breaches or cyberattacks. This necessitates robust encryption of patient-specific data. Healthcare institutions must adhere to the legal framework for the ethical use of patient data to protect data integrity from unauthorized use [70]. Third, bias in AI prediction algorithms [71]. It could be at different stages that contribute to health disparities. One of the main reasons is the underrepresentation of a specific population in the training data, which could be due to the inaccessibility of the healthcare facilities [72]. Additionally, AI predictions might yield accurate results on the data on which they were trained, but perform poorly on new, unseen data. It remains unclear to what extent it has understood the general principles [79]. The last is the concept of a “black box,” which results from the decision-making process’s lack of transparency regarding the reasons why AI generated the particular result [73]. For many years, clinical decision support tools have been integrated into EHR systems to assist physicians in identifying patients who are at higher risk, promoting medication safety, and adhering to evidence-based standards [80]. Historically, these systems have been used as advising tools, allowing medical professionals total discretion over whether to accept or reject the advice. The balance of power drastically shifts when AI starts acting on its own rather than just issuing commands. This shift highlights the need for close supervision as autonomy grows since it poses significant challenges to patient safety, accountability, ethical responsibility, and openness.

Conclusion

AI algorithms and radiogenomics have become promising tools for creating individualized therapy for CVD patients. The routine quantitative mapping of cardiac CT scans, which produces a variety of features that can be fed into ML and DL algorithms for CVD diagnosis and risk stratification, has been made possible by the quick development of radiomics. By helping clinicians make the right judgments and relieving them of difficult image processing duties, these innovative technologies have the potential to revolutionize current healthcare. Oncology accounts for the majority of this field’s study, while CVD receives less attention. With more investigation, the lessons learned from oncology-based studies can be applied to CVD.

Also, AI-based CVD imaging is still in its early stages, but it has a lot to offer to both clinicians and patients as it catalyzes the treatment towards more personalized care.