Counterfactual-based Recommendations for Prevention of High-Prevalence Chronic Diseases

Chronic diseases, among the leading causes of death and disability worldwide, are characterized by persistent symptoms that require long-term care. Effective prevention through the management of modifiable risk factors (e.g., lifestyle habits) is key to mitigate the associated health and socioeconomic burden. Artificial Intelligence (AI) predictive models trained on large sets of clinical data can support prevention efforts by identifying individuals at risk and highlighting patterns associated with disease onset and progression. Explainable AI (XAI) techniques have shown considerable potential in supporting the development of trustworthy predictive models by elucidating the internal mechanisms of their decision-making process. However, the integration of XAI-based medical decision support systems into clinical practice remains limited, as existing approaches often fail to generate explanations that are genuinely interpretable by humans, actionable, and clinically feasible. Counterfactual explanations, a class of purely data-driven, local, post-hoc XAI methods, offer a promising way to address these limitations by generating “what-if” scenarios that illustrate how changes in input features could alter model predictions. Nevertheless, this technique solely relies on associative patterns in data, which may reflect spurious relationships, resulting in recommendations that are impractical or infeasible in real-world settings. Incorporating causal reasoning in medical decision making can help uncover underlying disease mechanisms and produce more reliable and actionable insights. In this regards, causal learning techniques can be used to estimate the efficacy of hypothetical lifestyle interventions (e.g., improvements in physical activity and diet) by explicitly representing cause–effect relationships, thereby supporting the design of personalized preventive strategies. This Thesis project investigates advanced methodological frameworks using data-driven and domain expert-driven knowledge for developing predictive models aimed at chronic disease prevention. Counterfactual analysis, combining both purely data-driven XAI methods and causal learning approaches, is applied to two chronic disease prevention case studies (type 2 diabetes and cardiovascular disease) leveraging large-scale routinely collected data from primary care electronic health records and disease progression models. Constraints on feature mutability and adherence to a causal structure are imposed to promote actionability and clinical feasibility of the outcomes, while also maintaining compliance with quantitative performance metrics. The results demonstrate that the proposed methodologies enhance model transparency while producing clinically relevant insights, illustrating how targeted modifications of modifiable risk factors can improve patient outcomes, support personalized care, and reveal variability in benefits across patient groups.