Harder to mend than to break?—counterfactual explainable artificial intelligence for lifestyle medicine and heart disease prediction

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

Background: Heart disease is the most prevalent cause of mortality around the world, with estimates up to 80 % being caused by modifiable risk factors. Counterfactual (CF) techniques on machine learning (ML) models based on mixed behavioral and clinical data enable generation of synthetic transitions from healthy to unhealthy (downward) or unhealthy to healthy (upward), in order to simulate pathways between the states. We used a public package to examine characteristics of the created new instances. Methods: Our methodology employs a 3×3×2 analytical framework, integrating three ML models [naive Bayes (NB), random forest (RF), and support vector machine (SVM)] with three distinct CF generation algorithms [multi-objective CF (MOC), Nested and Interpolated Counterfactual Explanations (NICE), and What-If analysis] across two health outcome scenarios (heart disease to healthy and healthy to heart disease). Results: Our findings reveal that the sensitivity of ML algorithms to the generation of CFs is relatively low, indicating robustness in the ML models’ performance irrespective of the CF methodology employed. Conversely, the choice of CF algorithm significantly influences the quality and characteristics of the generated CFs, particularly regarding the number of features altered. This result underscores the impact of CF methodologies on the interpretability and applicability of CFs in clinical settings, where understanding the minimal changes required to alter a health outcome is essential for targeted interventions. Our analysis of transitions from positive to negative health outcomes (Pos2Neg) and vice versa (Neg2Pos) unveils profound differences, highlighting the asymmetric nature of health state transitions. Conclusions: This research contributes to the burgeoning field of explainable artificial intelligence (XAI) in healthcare, providing valuable insights into how different ML and CF approaches interact and impact the utility of predictive models in clinical decision-making. Our work emphasizes the careful selection of CF algorithms to enhance the actionability and relevance of CFs in health outcome predictions, paving the way for more nuanced and effective patient care strategies.