Plaque segmentation attempts to identify plaques precisely and consistently to measure plaque burden and track progression over time. This can be used to diagnose and plan treatment for CVD patients. The segmentation results can be utilized to quantitatively analyze plaque features such as volume, shape, and composition. Regression and progression of the plaque in over 10,000 patients in a few hours retrospectively. Fig. 3 shows the carotid ultrasound image phenotype (CUSIP) parameters derived from carotid artery scans representing the surrogate makers for coronary atherosclerotic disease.

Fig. 3.

Carotid artery scans show plaque in the far wall of the carotid artery. Carotid ultrasound image phenotype (CUSIP) consists of carotid intima-media thickness (IMT, ave., max., min.) and total plaque area (TPA). The shaded yellow region corresponds to the region of plaque burden. (A) Moderate plaque burden in carotid bulb; (B) Moderate plaque burden in common carotid artery; (C) Moderate-High plaque burden in carotid blub artery; (D) High plaque burden in carotid blub artery; and (E) High plaque burden in carotid blub artery. LCB, left carotid bulb; RCB, right carotid bulb (Original image, AtheroPointTM LLC, Roseville, CA, USA).

The segmentation process can be challenging due to the variability in plaque appearance and the presence of other structures in the images. A combination of multiple methods can be used to improve the accuracy and reliability of plaque segmentation as part of the radiomics features. Carotid ultrasound image phenotypes (CUSIP) have been tried before [62, 63, 64]. Therefore, it offers a wide range of applications for regular proactive monitoring of atherosclerotic plaque for CVD risk assessment [65, 66, 67, 68, 69, 70]. The statistical reports in the Pharma Mode shown in Fig. 4 will give the regression or progression of the plaque based on the statins during the 5–10-year follow-up of thousands of patients. The surrogate carotid artery disease and genetic-based biomarkers gathered from blood plasma will be used to track the atherosclerotic disease of the coronary artery disease. In addition, statistical regression plots will be constructed to link CVD risk to other clinical biomarkers.

Fig. 4.

GUI view of AtheroEdge™ 2.0, showing carotid artery plaque measurement. (A) Pharma Mode selection for selection of the image resolution. (B) CUSIP parameters showing cIMT and TPA measurements for low plaque patient. (C) CUSIP parameters (cIMT/TPA) measurements for the high-plaque patient. (D) Pharma Mode report showing the CUSIP parameters and composite CVD risk (marked in color code). This includes (i) renal effect on CVD risk via estimated glomerular filtration rate (eGFR) biomarkers, (ii) arthritis effect on CVD using erythrocyte sedimentation rate (ESR) biomarkers, (iii) diabetic effect on CVD using HbA1c, (iv) changes in CVD due to low-density lipoprotein (LDL) (Original image, AtheroPoint™ LLC, Roseville, CA, USA).

AI-based radiomics techniques have been recently used to extract and analyze segmented features from the carotid artery based on carotid ultrasound imaging modalities [30, 71]. The studies have used different DL algorithms, including convolutional neural networks (CNN), UNet, UNet+, and DenseNet, along with other machine learning (ML) algorithms such as support vector machine (SVM), random forest (RF) and logistic regression (LR) for classification. An example of AtheroEdge™ 3.0 in Pharma mode is seen in Fig. 5A. On clicking the “Pharma Mode” feature of the system, the epidemiologist or physician conducting the Pharma trial will select the spreadsheet with 100,000 patients.

Fig. 5.

GUI view of AtheroEdge™ 3.0, AI-based Cardiovascular Disease Risk Stratification system in Pharma mode. (A) AI-based system in Pharma Mode with 70% (green) patient cohort processed. (B) AtheroEdge™ 3.0: An example of the results of the Pharma Mode trial. The color code shows the CVD risk class. The percentage risk is shown in the second column from the right. The rest of the columns are the CVD risk predictors (zoomed version of Fig. 5B is shown in the Supplementary Material) (Original image, AtheroPoint™ LLC, Roseville, CA, USA).

The patient data file will be loaded into the system and ready to select the trained AI model for CVD risk prediction. On clicking the “Trained Model” button, the user can fetch the trained AI model and the number of risk classes. Finally, on clicking the “Predict” button, the user AtheroEdge™ 3.0 system will compute the risk of all the 100,000 patients in, say, five seconds per patient. Fig. 5B depicts an example report. Table 1 (Ref. [30, 72, 73, 74, 75, 76, 77, 78, 79, 80]) also includes a list of the segment characteristics used in each investigation, such as carotid intima-media thickness (cMIT), carotid wall thickness (cWT), lumen diameter (LD), plaque load (PB), and plaque area (PA). Table 1 also displays the accuracy (ACC), area under the curve (AUC), and study conclusion for each algorithm. Saba et al. [72] used ML-based radiomics to assess intra/inter-observer variability in detecting the common carotid artery (CCA) using SVM and RF classifiers, achieving an ACC of 98.32%. Biswas et al. [30] used DL-based radiomics to detect joint cWT, and PB features with an accuracy of 93% and AUC of 0.89. Vila et al. [73] also used a DenseNet CNN model to detect and estimate the cIMT with an ACC of 96.45% and an AUC of 0.89. Jain et al. [74] used a UNet-based model to detect and segment the CCA plaque area (PA) with an ACC of 88% and an AUC of 0.91. In another study, Jain et al. [75] used UNet+ to detect and segment the PA in the internal carotid artery (ICA) with an ACC of 97% and AUC of 0.99%. Yuan et al. [76] used UNet-based radiomics to segment and analyze cMIT and plaque with an ACC of 97% and a Dice score of 83.3–85.7.

Molinari et al. [77] used a CNN-based radiomics model to analyze cMIT and cWT features, achieving an ACC of 95.6% and an AUC of 0.83. UNet-based tissue characterization of CCA plaque by Gago et al. [78] yielded an ACC of 79.00%. Shin et al. [79] used a CNN-based model to analyze the viscous plaque index with an ACC of 83% and an AUC of 0.87, while Lainé et al. [80] segmented the cWT with an ACC of 86% using a dilated U-Net architecture.

A polygenic risk score, often known as a “polygenic risk score” or “PRS”, is a prediction tool that assesses an individual’s genetic predisposition to various diseases, such as CVD [81]. The PRS is estimated by taking into account the presence and frequency of genetic variants associated with the disease [82, 83].

SNPs associated with CVD can be used as markers to establish an individual’s genetic profile [84]. Individual genotyping data is collected after SNP identification. Large-scale genotyping initiatives or DNA tests can provide genotype data [85]. After that, each person’s PRS is calculated based on their genome’s risk allele count [86]. The CVD association of risk alleles influences their weight [81]. The “effect size” measures how much an SNP affects CVD risk [87]. This statistic typically measures the strength of the relationship between two variables. The odds ratio compares the risk of CVD for those with a particular SNP to those without [88]. The odds ratio usually quantifies the effect [82]. An individual’s PRS is calculated by summing the weights of all risk alleles found in their genotyping data [89]. Higher PRS scores indicate a higher CVD predisposition. Fig. 6 shows using the molecular data to calculate the PRS.

Fig. 6.

Molecular data used for prediction of PRS for CVD (Original image, AtheroPoint™ LLC, Roseville, CA, USA).

The PRS is not a diagnostic tool and should not be used alone to make clinical decisions [90]. The PRS should be used with family history, lifestyle, and other risk factors to make a more informed decision about CVD risk [91, 92]. Table 2 (Ref. [93, 94, 95, 96, 97, 98, 99, 100, 101, 102, 103, 104, 105, 106, 107, 108, 109, 110, 111, 112, 113, 114]) lists In-vitro biomarkers that can be identified in the serum and plasma of asymptomatic and CVD patients. To improve diagnosis and CVD risk stratification, multivariate genomic-based biomarkers and multivariable prediction models have been used [93, 115, 116]. The prediction models assess multivariate atherosclerosis and coronary artery development indicators in circulating molecules. Cellular, biochemical, epigenetic, and/or transcriptional GBBM contributing to CVD are addressed below. ACS biomarkers include C-reactive protein (CRP) [94], a liver protein released in response to inflammation, which increases cardiovascular risk. Another study by Barwari et al. [95] and Barwari et al. [95] found plasma LDL-cholesterol levels to be a biomarker for ACS. LDL-cholesterol is considered “bad” because it causes arterial plaque. Joshi et al. [96] and Cristell et al. [97] noted that PCSK9 regulates cholesterol metabolism and that the identified variant is associated with higher LDL-cholesterol levels and an increased risk of cardiovascular events. Pedicino et al. [98] and Partida et al. [99] identified pro-inflammatory CD4+ cells with poor CD28 expression as a biomarker for ACS. Sato et al. [100] and Montone et al. [101] found that biomarkers of ACS include a drop in the number of regulatory T lymphocytes and their ability to stop cancer cells from growing. The following rows list plaque rupture biomarkers without systemic inflammation. Kelley et al. [102] found inflammasome activation, while Kreiner et al. [103], Toldo et al. [104], and Omerovic et al. [105] found interleukins 1 and 18 and emotional disturbance-induced catecholamine release as CVD biomarkers. Repova et al. [106] found neutrophil activity, while Dang et al. [107] found macrophages or T lymphocytes as ACS biomarkers, eroding plaque. In summary, Table 2 provides a comprehensive overview of different biomarkers associated with plaque progression in patients with CVD. The table highlights the complexity of CVD and the need for personalized treatment approaches based on individual patient characteristics.

Radiogenomics Risk Predictors are valuable biomarkers used for CVD risk stratification, combining radiological imaging data and genetic information to assess an individual’s CVD risk [75, 117]. These predictors analyze quantitative features extracted from imaging modalities like MRI, CTA, OCT, and US, alongside genetic data, serving as specific indicators of pathological changes in the arteries [118]. Referred to as “RBBMs”, they are not mere symptoms but powerful tools for risk evaluation [119]. Numerous studies, detailed in Table 1, have explored and emphasized the significance of these risk predictors.

CAD affects major blood vessels supplying the heart muscle, primarily due to cholesterol deposits causing atherosclerosis, leading to reduced blood flow and potential heart attacks, angina, or strokes [120]. Symptoms of CVD may differ between genders, with men experiencing more chest pain and women often presenting shortness of breath, nausea, and extreme fatigue, indicating underlying heart or blood vessel issues [121]. Abnormal anatomical conditions in CVD, such as cardiomegaly, myocardial infarction, vasculitis [122], atherosclerosis, and thrombosis [123], are observable through medical imaging, physical examinations [124], or laboratory tests, providing critical insights into disease development. Studying these pathological changes is vital for understanding CVD and utilizing radiological biomarkers, to detect stenosis and other risk factors, contributing to a comprehensive comprehension of CVD and its related conditions.

The medical imaging field has noted the progress made in ML and DL [125, 126, 127, 128]. Deep neural networks (DNNs), a subset of DL, work like the human brain [39, 117, 129]. AI has been used in recent studies to figure out the risk of CVD using RBBM [11, 12, 38, 130, 131, 132, 133, 134] and GBBM [33, 135, 136] frameworks. DL is becoming more popular because (i) it automatically extracts features [137], (ii) it can fuse with different ML configurations for classification [117, 138], (iii) it uses UNet and Hybrid UNet-based DL strategies for segmentation [43, 119], and (iv) Lastly, it gives more accurate segmentation and solo or ensemble-based classification because it can go through forward and backward propagation by reducing the loss using different kinds of loss functions [74].

4.2.1 Training and Prediction

The architecture consists of two halves. The left is the training subsystem, while the right is the prediction subsystem.

The DL training classifiers consist of one of the DL classifiers, namely, long short-term memory network (LSTM), recurrent neural network (RNN), gated recurrent units (GRU), bidirectional-LSMT (BiLSTM), bidirectional-RNN (BiRNN), and bidirectional-GRU (BiGRU) (presented in the following subsection).

Along with the DL classifier bank, there are supervised clinical risk labels representing ground truth (GT), such as heart failure (or high CVD risk) and stroke [130, 139]. This GT representing the CAD includes computed tomography (CT) coronary score [140] or quantification of CAD lesions using intravascular ultrasound (IVUS) [141, 142]. Several non-linear training-based approaches have been shown in heart disease risk stratification [12, 132, 134, 143]. Fig. 7 shows AI-based architecture for CVD risk assessment.

Fig. 7.

AI genomics-based architecture for CVD risk assessment (Original image, AtheroPoint™ LLC, Roseville, CA, USA).

In the past, AI models were considered “black boxes”, meaning that it was unknown what specific patterns in the input data would result in the desired output. To put it another way, even if a classification algorithm is nearly perfect, humans must still determine what circumstances led to the categorisation [169]. The lack of transparency in model decisions becomes problematic when a model is trained to make accurate predictions based on features that are irrelevant to the situation at hand, and, as a result, it lacks generalization for that dataset [170]. Fig. 8 (Ref. [171]) shows the eight customizable DL processes: (i) DL training, (ii) quality assurance (QA), (iii) installation, (iv) deployment, (v) prediction, (vi) cross-validation-based testing (A/B test), (vii) monitoring, and (viii) debugging. These processes are made possible by incorporating the feedback loop in XAI. Incorporating XAI enhances its usability, increasing demand for such systems and generating higher revenue [172]. Moreover, it offers greater stability to the AI system, thus extending the lifespan of the product design. XAI strategies expose sophisticated model decision patterns [173]. XAI can provide new information about a dataset and the disease by revealing input variables’ predictive importance, including risk factors, traits, and protein expression levels. However, models like principal component analysis (PCA), LR, and Cox linear regression use model parameters to determine feature relevance [174].

Fig. 8.

Eight aspects of XAI [171].

The average importance of features across samples is determined using feature analysis [175]. Training models to distinguish across classes can be expanded to numerous classes, providing insights such as specific patient groups [176]. XAI clinical power stems from strategies that explain patient decisions. Fig. 9 illustrates the clinical decision system concept employing XAI. If the AI model predicts a patient’s risk, the doctor can inquire about the model’s output for clinical, imaging, or molecular features route databases and gene set enrichment estimates can uncover risk variables and molecular disease pathways. This information can help clinicians choose better treatment plans. Furthermore, by understanding the decision process, the physician will be able to judge the accuracy of the predictions. Explainability has thus been recommended as an ethical requirement for prospective clinical decision systems [177]. This section describes various XAI techniques and summarises medical and clinical applications for patient evaluation [8, 42, 43, 153].

Fig. 9.

Clinical features decisions system using XAI. AI models are trained using clinical, imaging, or multi-omics data into different classes (low, mild, moderate, and high). The trained algorithm predicts (prediction blocks) patient disease or symptom risk. Lastly, the XAI presents decision patterns to help the doctor evaluate the forecast and plan treatment (Original image, AtheroPoint™ LLC, Roseville, CA, USA).

Cloud-based XAI is an approach that allows researchers and healthcare professionals to access powerful ML/DL models and tools for data analysis and prediction without needing specialized hardware or software. In the context of predicting CVD, cloud-based XAI can provide an efficient and cost-effective way to analyze large amounts of data and generate accurate predictions. Cloud-based XAI relies on cloud computing technology, which involves using remote servers to store and process data rather than relying on local resources [43, 72, 178, 179]. Fig. 10 shows a cloud-based XAI system. By leveraging cloud-based infrastructure, researchers can access powerful computing resources and AI models that can process vast amounts of data quickly and efficiently.

Fig. 10.

AtheroCloud™ system depicts a typical daily routine. A yellow color fills the plaque region, representing the total plaque area after the image has been loaded in the Routine mode and the “Auto Trace” button has been clicked (Original image, AtheroPoint™ LLC, Roseville, CA, USA).

To build a cloud-based XAI system for predicting CVD reserchers can follow a few steps: (i) Feature selection: Feature selection is the process of identifying the most important features contributing to predicting CVD. This can be accomplished through the use of statistical approaches such as regression analysis or machine learning/deep learning algorithms that automatically choose the most relevant features. (ii) Model interpretability: The model can be made more interpretable using algorithms that explain individual predictions, such as LIME and SHAP. These algorithms provide feature importance scores and highlight the most important features for a specific prediction. (iii) Data visualization: Data visualization can be utilized to explore data and uncover patterns and relationships among features. Visualization tools such as scatter plots, heatmaps, and box plots can help researchers identify outliers and correlations that may be useful for predicting CVD. (iv) Cloud-based infrastructure: Cloud-based infrastructure can be used to store and process large amounts of data, making it possible to build more complex and accurate models.

Cloud-based AI platforms such as Amazon Web Services (AWS) and Google Cloud Platform (GCP) provide powerful tools for building and deploying ML/DL models [180]. (v) Collaboration: Collaboration is essential for developing accurate and reliable models for CVD prediction. Cloud-based XAI platforms enable researchers to share data and models, collaborate on model development, and work together to identify new insights and approaches [181]. Cloud-based XAI has several advantages over traditional on-premises solutions. First, researchers can access powerful computing resources without needing specialized hardware or software. Second, it provides a scalable and flexible platform that can be easily adapted to meet changing healthcare needs. Finally, it enables researchers and healthcare professionals to collaborate and share data and models, leading to better and more accurate predictions for CVD and other health conditions.

Heatmaps in Explainable AI work by highlighting the regions of input data that significantly influence the model’s output. For image-based tasks, such as object detection or medical image analysis, heatmaps visually depict the importance of different pixels or regions within the input image. The intensity of colors in the heatmap indicates the magnitude of the model’s focus on particular areas [182]. The generation of heatmaps often involves techniques like Gradient-weighted Class Activation Mapping (Grad-CAM) and Guided Backpropagation. Grad-CAM computes the gradients of the target class with respect to the final convolutional layer’s feature maps. These gradients are then global-average-pooled to obtain the heatmap, which highlights the relevant regions in the image. Guided Backpropagation, on the other hand, is used to ensure that only meaningful features are highlighted in the heatmap by preventing the model from backpropagating through unimportant regions [183]. By interpreting heatmaps, researchers, domain experts, and end-users gain valuable insights into why a specific decision was made by the AI model. For instance, in medical diagnosis, heatmaps can help doctors understand which regions of an X-ray or MRI scan contributed most to the AI model’s identification of a certain condition [184].

Skandha et al. [138] proposed the first of its kind method to characterize and classify the carotid plaque using an optimized transfer learning approach and SuriNet (a class of Atheromatic™ 2.0 TL). Eleven Atheromatic™ models were implemented, and the best AUC was 0.961 (p 0.0001) from MobileNet and 0.927 (p 0.0001) from SuriNet. We validated the performance using grayscale median, fractal dimension, higher-order spectra, and spatial heatmaps. TL showed equal and comparable performance to deep learning. The Atheromatic™ 2.0 TL model showed a performance improvement of 12.9% over Atheromatic™ 1.0 ML (AtheroPoint™, Roseville, CA, USA) compared with the previous. Moreover, the Atheromatic™ 2.0 TL model displayed an outstanding 12.9% performance improvement over the previous Atheromatic™ 1.0ML (AtheroPoint™, Roseville, CA, USA), which was based on traditional machine learning techniques. Notably, the system’s validity was established using a widely accepted dataset, reinforcing the robustness and reliability of the findings. This study sets a significant milestone in understanding carotid plaque characteristics, opening new avenues for enhanced diagnosis and treatment in the field of carotid artery disease.

In conclusion, heatmaps are a crucial tool in the toolkit of Explainable AI, offering visual interpretations of AI model decisions and making complex AI systems more transparent and interpretable. As AI continues to impact various aspects of our lives, XAI using heatmaps will play an instrumental role in building responsible and trustworthy AI applications.

The statistical analysis involved the presentation of continuous variables as mean $\pm$ standard deviation. The normal distribution of continuous variables was verified using the Kolmogorov-Smirnov test [185]. The independent samples t-test or Mann-Whitney U test was used to compare continuous data, while the chi-square test or Fisher’s exact test was used for categorical variables, as appropriate [186]. For comparisons involving multiple groups, one-way analysis of variance (ANOVA) or Kruskal-Wallis H test with Bonferroni test was used, as appropriate [187]. Pearson r and Spearman rho coefficient were used for correlation assessment, depending on the nature of the data [188].

Multivariable linear regression analysis assessed the association between left atrium (LA) function parameters and myocardial scar size/location [118]. The regression models included demographic characteristics, traditional cardiovascular risk factors, and Late gadolinium enhancement (LGE) location/size as covariates. To avoid co-linearity, Spearman correlation coefficients were used to test for correlations between continuous variables, and variables with r $>$0.50 were excluded from the same multivariable model [189]. A p-value 0.05 was considered statistically significant. The analysis was performed using IBM SPSS Statistics version 22 (SPSS Inc., Chicago, IL, USA) [190].

AI-based survival analysis is a cutting-edge technique used to predict the risk of CVD [134]. Using huge amounts of medical data, AI algorithms may estimate the chance of an individual having CVD based on parameters such as age, gender, blood pressure, cholesterol levels, and lifestyle patterns [129]. This approach can give physicians valuable insights into patient health and help them create personalized treatment plans for each individual [191].

By utilizing AI-based survival analysis, doctors can also identify high-risk individuals who may require closer monitoring or more aggressive treatment to prevent the development of CVD [192]. AI technology can revolutionize the healthcare industry by improving the accuracy of CVD diagnosis and treatment, leading to better patient outcomes and improved overall public health [69, 193].

This is the first study of its kind (a) that combines radiomics and genomic biomarkers to detect the severity of CVD and stroke risk precisely, and (b) an explainable XAI-based composite risk CVD/Stroke model using DL can predict with precision in a preventive and personalized (aiP${}^3$) framework benefiting pharmaceutical paradigm. Using these two hypotheses, we demonstrated that CVD and stroke risk severity could be determined using RBBM and GBBM biomarkers in the DL framework. Such models can be considered “personalized medicine frameworks”. One significant advancement is the inclusion of carotid imaging and CVD genetic indicators in the DL framework for CVD risk classification. A set of six recommendations for accurate, robust, real-time CVD risk assessment utilizing combined RBBM and GBBM were offered. The study proposed model having significant potential such as (a) radiomics and genomic biomarkers have a strong correlation and can be used to detect the severity of CVD and stroke precisely, and (b) The proposed review suggests a unique, unbiased, and XAI model (AtheroEdge${}^{\text{TM}}$ 4.0) that can predict the composite risk of CVD/Stroke using radiogenomics in the pharmaceutical paradigm. The proposed model required less time than the traditional model to stratify CVD/stroke risk, saving thousands of hours using phrmamode.

AtheroCloud${}^{\text{TM}}$ can be utilized in either the (a) “Routine” mode or the (b) “Pharma” trial mode. Both the batch of carotid ultrasound scans and individual scans are automatically processed in real-time in the Routine mode and Pharma trial mode, respectively, during the patient visit. The system supports large databases, from gigabytes to terabytes, and can operate in batch mode, processing as many as 10,000 images simultaneously.

Table 4 (Ref. [27, 41, 75, 119, 143, 159, 194, 195, 196, 197, 198, 199, 200, 201]) displays the benchmarking studies. It has 17 attributes, namely, K0: Serial Number, K1: Studies, K2: Year, K3: References, K4: OBBM, K5: LBBM, K6: RBBM, K7: GBBM, K8: PBBM, K9: EBBM, K10: Preventive, K11: Prediction, K12: Personalized, K13: AI Type, K14: FDA discussion, K15: Clinical Setting, K16: Risk of bias; K17: AI explainability. Our observations showed that there were only eight studies [27, 41, 159, 194, 195, 196, 197] that discussed the role of GBBM in CVD using DL, while the remaining studies predicted CVD using RBBM [75, 119, 143, 198, 199, 200, 202, 203]. However, all the studies mentioned the role of OBBM and LBBM in CVD risk. Five studies adopted DL technologies [27, 41, 75, 196, 198, 199], while eight studies used ML [11, 131, 197, 202], and two studies used HDL [143, 201]. Only four studies discussed the risk of bias (RoB) [75, 143, 202]. Except for the proposed study, no other study addressed clinical settings, AI explainability, and fusion of GBBM and RBBM.

Evaluating bias in AI models has gained much greater significance in recent years [204, 205]. Earlier computer-aided diagnosis techniques showed a lack of bias in evaluations [206]. To reduce bias, a large sample size, appropriate clinical testing, the incorporation of comorbidities, the use of big data configurations, the usage of unseen data analysis, and the scientific validation of training model design are all strategies that can be utilized [140, 207]. Essential phases in patient risk stratification include determining the AI RoB [191, 207, 208] and suitably modifying diagnostics and treatment.

Edge devices are becoming increasingly important as cloud-based systems and the internet improve. Edge devices are extremely important when using trained AI models for future predictions or disease risk stratifications in mobile frameworks. There is a requirement to deploy compressed models since huge data models cannot be deployed on edge devices. Image-based DL models such as fully convolutional networks (FCN) or segmentation networks (SegNet) [40] can be pruned using evolutionary algorithms such as Particle Swarm Optimization (PSO), Genetic Algorithms (GA), Wolf Optimization (WO), and Differential Evolution (DE) [209]. The future of radiomics-based CVD risk stratification fused genetic-based paradigms can be compressed and deployed on edge devices for rural areas, especially in third-world nations.

The following are recommendations for a suggested Radiogenomics model that can be used to stratify CVD/stroke risk. The study proposes two hypotheses: (a) radiomics and genomic biomarkers have a strong correlation and can be used to detect the severity of CVD and stroke precisely, and (b) The proposed review suggests a unique, unbiased, and XAI model (AtheroEdge${}^{\text{TM}}$ 4.0) that can predict the composite risk of CVD/Stroke using radiogenomics in the pharmaceutical paradigm. The following recommendations are: (i) requires a clinical evaluation and scientific validation for reliable detection and CVD risk stratification, and (ii) requires hyper-parameter optimization in CVD/Stroke risk stratification. (iii) balancing the risk classes (control, low-risk, and high-risk) is the most effective way to minimize DL bias; (iv) with proper pruning and compression, DL systems can be adapted to edge devices; (v) A DL system that relies on surrogate carotid imaging can be cost-effective without compromising precision in CVD risk stratification.

This pilot review’s ability to risk stratify CVD and stroke patients by integrating RBBM and GBBM was a major strength. The biomarkers generated from radiological, biochemical, and morphological complexity that demonstrated a link to CVD supported the first hypothesis. An XAI-based composite risk CVD/Stroke model using DL can predict, with precision, in a preventive and personalized (aiP${}^3$) framework benefiting the pharmaceutical paradigm. A DL approach was presented to evaluate CVD and stroke risk by integrating RBBM and GBBM. While the system is pretty straightforward, it requires optimization to eliminate the possibility of bias and generalization to account for comorbidities. Better comprehensive feature space can be tried for superior DL-based classification [210]. Ensemble-based solutions embedding PCA for best feature selection followed by recurrent neural networks are possible extensions for superior CVD/stroke risk solutions [211].

In clinical trials, researchers often analyze the effectiveness of a new treatment or intervention compared to a control group. One of the essential statistical measures used in this context is the “Odds Ratio” (OR). The odds ratio helps to assess the association between an exposure (such as receiving a treatment) and an outcome (such as a positive response to the treatment) [212].

The odds ratio is calculated by taking the ratio of two odds: the odds of an event occurring in the treatment group and the odds of the same event occurring in the control group. Mathematically, it can be expressed as:

The odds ratio is calculated as follows:

(3)$$\mathrm{OR}=(\mathrm{a}/\mathrm{b})/(\mathrm{c}/\mathrm{d})$$

Where:

$\bullet$ a is the number of people in the exposed group who have the outcome

$\bullet$ b is the number of people in the exposed group who do not have the outcome

$\bullet$ c is the number of people in the unexposed group who have the outcome

$\bullet$ d is the number of people in the unexposed group who do not have the outcome

The odds ratio can be used to determine if there is a significant difference in the odds of the outcome between the treatment and control groups. If the odds ratio is equal to 1, it suggests that there is no association between the treatment and the outcome, meaning the treatment has no effect. If the odds ratio is greater than 1, it indicates that the treatment is associated with a higher likelihood of the outcome. Conversely, if the odds ratio is less than 1, it implies that the treatment is associated with a lower likelihood of the outcome [213].

Researchers typically calculate the odds ratio along with a confidence interval to assess the statistical significance of the results. If the confidence interval includes the value 1, the results are considered not statistically significant, while values above or below 1 indicate a statistically significant association [214]. The odds ratio is a useful tool in clinical trials as it provides valuable information about the relative impact of a treatment compared to a control group. However, it’s essential to interpret the odds ratio in the context of the study design and the specific clinical question being addressed [214]. In conclusion, the odds ratio is a critical statistical measure used in clinical trials to analyze the association between treatment and outcomes, providing valuable insights into the effectiveness of interventions.