This scoping review aimed at presenting the state-of-the-art of ML applications on exposomics data, specifically focusing on CVDs. To this end, a substantial amount of current literature on the selected topic has been identified and analyzed with respect to diverse aspects i.e. the use case of each application and their distinguishing features, ML techniques employed and the categories of exposomics data exploited. Building on the insights from the previous sections, we will summarize findings and elaborate on the strengths, the challenges and the opportunities that ML-enabled exploitation of exposomics data holds for CVD-related variables.

As reflected by the timeline of identified publications, the field is constantly expanding, particularly from 2021 onwards, with the expansion in exploitation of ML on exposomics targeting CVDs being predominantly driven by research in the US and Asia. However, the past 4 years the EU has shown an increasing interest in the Human Exposome Project initiating multimillion funding and thus recognizing potential to enhance public health [85].

Overall, ML approaches have provided novel insights into the exposomic factors affecting CVD by overcoming the inherent limitations of traditional epidemiological models, which assume linear and independent relationships between factors and outcomes. Unlike conventional regression-based models (such as linear regression and LASSO), which can only produce accurate interpretations under the assumption of linearity, ML techniques can capture nonlinear, synergistic, and threshold effects among complex exposures while also improving accuracy (e.g. in [44], [63]). This is particularly important in exposomics where high dimensional data is involved, and multiple factors interact in intricate ways that cannot be easily disentangled using standard statistical methods. By leveraging algorithms that are more robust to multicollinearity (e.g., random forests) and capable of identifying hidden patterns in high-dimensional data, ML enables the detection of previously unrecognized interactions [55].

As presented in detail in the context of RQ1, two main categories of studies have been identified based on the application context: the first one focusing on disease management, prevention and understanding, and the second one addressing healthcare resource planning and management. Overall, the direction of research efforts falls within the scope of AI-driven health investigations aimed at effective risk stratification, public health surveillance and health policy planning and the ML assisted exploratory association analysis between novel risk factors and CVD outcomes, all of which are described to contribute to health-related sustainable development goals [86].

Regarding ML techniques, it is noteworthy that ML tasks have been primarily framed in a supervised context, leaving ample space for the exploration of under-researched unsupervised techniques. Overall, a variety of ML algorithms spanning from linear to non-linear and ensemble algorithms have been exploited. In many cases different categories of algorithms are employed and compared against each other. Ensemble algorithms seem to be the most widely adopted approach and especially Random Forest, Stacking Ensemble and XGBoost which frequently rank as the most efficient solutions in comparative experiments. In the context of unsupervised learning and data clustering, more methods could be explored imposing minimal assumptions on data, trying to identify patterns, simplify large datasets and enhance predictive modeling. A valuable addition to the toolbox of the well-established linear clustering methods could be the case of non-linear clustering algorithms and specifically the so-called manifold learning, which aims to cluster data by identifying the intrinsic manifold upon which the data lies [87].

As presented in the context of RQ3, environmental exposure seems to be the most widely investigated aspect of the feature space followed by lifestyle-related exposure, with both of them often handled individually. In contrast, socioeconomic factors are typically studied in combination with other types of variables. Most environmental exposure studies are exploratory in nature, seeking to confirm associations of variables with health outcomes, instead of developing complex prediction models; they often aim to expand the understanding of the exposure-outcome relationships through identifying new or inadequately studied risk factors that may play an important role on disease prevalence and progression [11], [42], [65], [66]. Such efforts showcase the need for public health policies related to the environment [62].

A critical consideration in applying ML to exposomics research is the validity of inference. Many widely used ML algorithms, such as regression-based (e.g., LASSO) and tree-based (e.g., Random Forest and XGBoost) methods, are primarily designed for prediction rather than for modelling exposure-outcome associations (e.g., Bayesian Additive Regression Trees and Bayesian Kernel Machine Regression). Thus, they do not necessarily provide valid statistical inferences (e.g., they are often unable to quantify uncertainty around estimates) and/ or indicate true exposure-outcome associations. These algorithms are typically optimized with respect to prediction-oriented measures (e.g., AUC for classification, MAE for regression etc.) and then used to report associations - often relying on various feature importance metrics to identify key determinants [35], [39], [40], [41], [42], [44], [53], [55], [59], [60], [61], [62], [63], [65], [66], [68], [69], [71], [73], [74], [78], [81], [84], [88]. It is important to note that this approach raises concerns about post-selection bias, as feature selection is data-driven, and traditional statistical measures (e.g., p-values, confidence intervals) do not account for the fact that variables were selected based on observed data. For example, when conducting multiple statistical tests across many variables (e.g., in the context of linear regression tools), the likelihood of false positives increases, necessitating multiple testing adjustments (e.g., Bonferonni correction, False Discovery Rate etc.). Without proper correction, significance can be overstated resulting in potentially misleading conclusions. This is especially pertinent in exposomics, which frequently involve high-dimensional data with numerous, highly correlated exposures, and where identifying true predictors amidst correlated covariates is a significant challenge [89]. For example, in [62] the authors report the relative importance of each feature and compare the results with General Linear Model (GLM) where they consider the features significant based on p-value <0.05 without correcting for multiple comparisons.

Closely related to the above considerations is the issue of causal reasoning. When correlated variables are fed into ML models predicting CVD outcomes, their relative importance will be estimated higher when their impact on a predicted score is higher without necessarily implying causality. To enhance the reliability and generalizability of ML-driven exposomics research, a sound methodology should pay special attention to rigorous validation frameworks, including carefully chosen post-selection inference methods, such as cross-validation, permutation-based significance testing, stability selection, and, ultimately, external validation. However, the causality of the identified associations still needs to be validated independently within a causal inference framework [88] and/or through comparison with conventional association analysis [68]. When these aspects cannot be meticulously addressed, or are only partially addressed, authors should clearly acknowledge the limitations of their study. Expanding the use of causal inference methods (e.g., instrumental variable analysis, propensity score matching) within the ExWAS context, can help establish robust links between exposures and CVD outcomes, mitigating the impact of non-causal correlations and promoting the understanding of the underlying causal pathways.

Several additional limitations have been also reported across the identified literature. Firstly, the validity of the exploited ground truth cannot be ascertained, for example when using diagnostic codes (e.g., the international classification of diseases [65], [71]) corresponding to hospitalizations or death certificates [90]. Secondly, in the case of self-reported data categories such as lifestyle data [11], [39], [41], [42], [44], [45], [63], [78], [88], subjectivity and recall bias cannot be avoided. In addition, the bulk of conducted research only involved incidents that occurred in a single region and/ or the cohort was ethnically homogeneous [40], [44], [65], [71], [84]. Another important limitation stems from the lack of temporal alignment between the collection of medical data concerning CVD incidents and the collection of exposome-related data with many of the identified studies being observational in nature. Most of the time, the data sources were not comprehensive and/ or standardized [36]. Several works report limited availability of data sources [35], [44], [88], and finally, it is difficult to directly compare results from different studies since they are heterogeneous in terms of target, problem framing, study design, sample size and outcome assessment as reflected by the occasional lack of an external validation process but also by the diverse metric scores reported.

Most of these limitations are data-related, primarily arising from issues such as data heterogeneity, lack of standardization, inconsistent terminology, and incomplete, biased, or limited data sources. To address these challenges, it is essential to integrate standardized protocols for exposome data collection and harmonization across studies, reducing bias and enhancing reproducibility. Establishing a common language for exposomics is crucial for eliminating inconsistencies in terminology and classification. This would ultimately enable more effective comparison of findings across studies. It has been proposed by Schmitt et al. 2023 [34] that exposomics research would benefit significantly from a comprehensive, standardized, and interoperable data ecosystem framework, facilitating the integration of diverse longitudinal datasets across multidisciplinary [91] and international collaborations—similar to the Human Genome Project [92]. Such an ecosystem would link studies across continents, merging data from both existing and future longitudinal research. By establishing a federation of global data resources, this initiative would enhance the visibility, accessibility, and usability of exposomics data. Achieving this vision requires collaboration among multiple stakeholders, working together to form a global consortium that unifies frameworks and provides guidelines for an Exposome Study Design and a comprehensive exposome database [91]. In this context, it should be noted that the Environmental Exposure Assessment Research Infrastructure (EIRENE [93]) is the first European infrastructure dedicated to the human exposome.

The adoption of longitudinal study designs with synchronized environmental and health data collection can enhance the ability to track exposures over time and infer causality more reliably (e.g., the UK biobank utilized in [40], [84]). Additionally, integrating diverse data sources, including computerized and standardized data from wearables, can help mitigate recall bias, which is often introduced by self-reporting [11], [39], [40], [41], [42], [44], [45], [63], [65], [78], [84], [88]. Lastly, Explainable AI (XAI) techniques are valuable for enhancing model interpretability, enabling qualitative comparisons and ensuring that associations identified between exposomic factors and CVD-related outcomes are meaningful in real-world applications. It is noteworthy that many studies identified in this review implement XAI techniques in meaningful ways, attempting to interpret their models and explore potential associations (e.g., by using the model-agnostic SHapley Additive exPlanations) [11], [40], [42], [60], [63], [68], [74], [78], [81], [88].

All in all, the growing intersection of ML and CVD exposomics offers unprecedented opportunities to refine exposure assessment, improve early disease detection, and guide public health interventions. Expanding the feature space to include non-traditional CVD risk factors—spanning environmental, social, and lifestyle domains—has shown promise in enhancing the performance of conventional models targeting CVD-related variables. Such findings may serve as a foundation for developing high-precision tools for personalized healthcare management. On a broader scale, this approach holds potential for prioritizing healthcare resource allocation, optimizing workflows, and implementing targeted prevention and intervention strategies to reduce CVD-related healthcare costs.

However, to translate these advancements into actionable insights, stronger methodological rigor and standardization efforts are needed. Even though ML algorithms effectively enhance predictive modeling, most of them are not inherently designed for causal inference, raising concerns about the validity of reported exposure-outcome associations. Addressing limitations requires rigorous validation frameworks, careful post-selection inference methods, and integration of causal inference techniques into ML pipelines. Additionally, inconsistencies in data sources, lack of standardization, and the heterogeneous nature of studies hinder comparability and reproducibility. Establishing standardized protocols and interoperable data-sharing frameworks, akin to initiatives in other healthcare domains, holds the potential to significantly advance the field.