Human health risk assessment is a well-established methodology that identifies and quantifies the potential risk of adverse health effects in humans who may be exposed to chemical contaminants in environmental media, such as air, water, soil, and food (Fryer et al., 2006). Human health risk assessments can extend to other areas such as infectious pathogens and transmissible diseases (Rees et al., 2019).
Traditionally, human health risk assessments rely on manual processes. Skilled experts evaluate available surveillance, clinical, or exposure data often from diverse sources. The process results in constructing risk profiles, risk scenarios, qualitative heat maps and quantitative mathematical models. The analyses are used to inform risk management, used as an evidence base for decision-making, and for policy recommendations. These assessments, while valuable, are time-consuming and labour-intensive. Risk analysis done manually can be limited by the volume of data and evidence that can practically be processed by humans. Humans may bias risk assessments by focusing on certain evidence, excluding other evidence or make assumptions to simplify analysis.
Machine learning (ML) is emerging as a transformative force, offering the potential to automate and augment the steps of the risk assessment process. ML is a subfield of artificial intelligence that harnesses algorithms trained on data sets to create models, enabling machines to perform tasks previously carried out by humans. These tasks encompass categorizing images, analyzing data, and predicting outcomes.
One of the strengths of ML is its capacity to automate risk profiling and the ability to deal with multiple large data sets improving scalability. Ultimately, ML can enhance the efficiency and accuracy of risk assessments from its ability to handle different types of data (Alaa and van der Schaar, 2018). An example of data limitation in standard epidemiological approaches is seen using the Cox proportional hazards model. This is a commonly used statistical regression model in medical research for investigating the association between the survival time of patients and one or more predictor variables. It encounters limitations when attempting to amalgamate diverse data sources and modalities- ranging from demographic, surveillance, incidence, social, longitudinal, imaging, and multi-omics data. However, ML combined with neural networks can incorporate and use multi-faceted data inputs.
AI-based models efficiently harness available observational data to deduce the impact of various types of exposures (eg. medical treatments on individuals, infectious disease outbreaks or chemical exposures) taking into account their unique characteristics (eg. pre-existing comorbidities, synergistic risks or confounders). This approach aids in determining the most suitable risk management strategies for maximizing risk reduction. By extracting insights from diverse data sources and learning how to apply this knowledge, ML can contribute to a more comprehensive understanding of risk and how it relates to population health (van der Schaar et al., 2021).
In risk assessments, quantifying uncertainties is important. The Monte Carlo method is a well-established approach for factoring in uncertainty related to parameter values within risk assessment modeling. However, these models often serve as screening tools in scenarios where information is typically sparse and imprecise. This raises questions about the ability to assign true probabilities to parameter estimates. Therefore, scenarios are often modeled with varying degrees of uncertainty to account for gaps in knowledge. Similarly, Probabilistic Risk Assessment (PRA) approaches are used for quantifying risks and uncertainties in various fields, including engineering, nuclear safety, and environmental risk assessment. PRA involves the systematic analysis of possible failure modes, their probabilities, and consequences to determine overall risk. PRA considers both aleatory (inherent randomness) and epistemic (lack of knowledge) uncertainties (Guyonnet et al., 1999).
ML techniques offer a promising solution to model uncertainty and reduce bias by employing transfer learning methods. These methods implicitly address uncertainty in predictions for individuals from different populations than those used for training. Moreover, transfer learning is valuable not only for transferring models across populations but also for updating ML models within a single population, adapting to evolving cohorts over time. As diseases and populations change, scalable Bayesian optimization techniques (easily integrated with ML) have been developed to update models with new data efficiently while building on past optimizations (Zhang et al., 2019).
Currently machine learning methodologies can be applied to risk assessments but are still in their infancy. It is hoped that over time more robust ML methods will be developed which will automate some if not all risk assessment steps. ML holds much promise for the future and until further developed as standardized tools the role of the human risk assessor remains indispensable for now.
References
Alaa, A., & Schaar, M. (2018). Autoprognosis: Automated clinical prognostic modeling via Bayesian optimization with structured kernel learning. In International conference on machine learning (pp. 139-148). PMLR.
Fryer, M., Collins, C. D., Ferrier, H., Colvile, R. N., & Nieuwenhuijsen, M. J. (2006). Human exposure modelling for chemical risk assessment: a review of current approaches and research and policy implications. Environmental Science & Policy, 9(3): 261-274.
Guyonnet, D., Côme, B., Perrochet, P., & Parriaux, A. (1999). Comparing two methods for addressing uncertainty in risk assessments. Journal of environmental engineering, 125(7), 660-666.
Rees, E. E., Ng, V., Gachon, P., Mawudeku, A., McKenney, D., Pedlar, J., Yemshanov, D., Parmely, J., & Knox, J. (2019). Risk assessment strategies for early detection and prediction of infectious disease outbreaks associated with climate change. Canada communicable disease report = Releve des maladies transmissibles au Canada, 45(5), 119–126. https://doi.org/10.14745/ccdr.v45i05a02.
van der Schaar, M., Alaa, A. M., Floto, A., Gimson, A., Scholtes, S., Wood, A., McKinney, E., Jarrett, D., Lio, P., & Ercole, A. (2021). How artificial intelligence and machine learning can help healthcare systems respond to COVID-19. Machine learning, 110(1), 1–14. https://doi.org/10.1007/s10994-020-05928-x.
