Enhancing healthcare explainability through multiobjective counterfactual explanations.

AI-driven decision support systems possess the capability to improve health outcomes. However, advanced AI-driven systems face challenges related to interpretability and slow real-world adoption due to the use of abstract, black-box models. As a result, Explainable AI (XAI) approaches are evolving to provide explanations for predictions. Nonetheless, the predominant XAI approaches have several significant drawbacks. Firstly, these techniques require an understanding of model internals; secondly, they lack the potential to provide insights into feature interaction; and thirdly, they cannot generate alternative solutions based on What-if scenarios. In this context, counterfactual explanations (CEs) can improve the explainability of black-box models by generating what-if scenarios for input data points. However, producing clinically plausible CEs is complex due to the presence of the "Rashomon Effect", multiple conflicting objectives, and the trade-offs among them. This requires a robust Mechanism that ensures the interpretability and plausibility of CFs under balanced, optimized constraints. To address the above issues, this study proposes a novel multiobjective optimized counterfactual explanation (MoCE) methodology for generating clinically significant CEs. Our proposed methodology consists of two phases. Firstly, a Random Forest (RF) classifier was trained on the chronic kidney disease (CKD) dataset, using a single instance from the dataset as a patient profile to obtain the predicted outcome. Secondly, for baselines, we generated five CEs using diverse counterfactual explanations (DiCE) using a CKD instance. Next, we applied the NSGA-II algorithm-based MoCE methodology to generate multiple CEs against the same CKD instance. We compared both approaches, DiCE and the MoCE, qualitatively and quantitatively, for generating clinically plausible CEs while efficiently handling the confidence, validity, and sparsity trade-offs. Clinical relevance and significance of the CEs were assessed in an initial clinical evaluation conducted by a nephrologist expert. Both approaches achieve 100% validity. DiCE achieved a mean sparsity = 0.175 by changing approximately 2 features with medium confidence levels (0.59). Conversely, the proposed method generated CEs with enhanced confidence levels (0.9) by altering 7 attributes with sparsity (0.8-0.9), yielding more diverse and plausible solutions. The proposed MoCE methodology generated 100 Pareto-optimal CEs candidates, indicating a rich trade-off surface among validity, proximity, prediction confidence, and sparsity. The comparison highlights that DiCE generates CEs by altering a minimum number of features, resulting in highly sparse but less diverse and less clinically relevant CEs. In contrast, the proposed MoCE approach generated a diverse set of CEs by simultaneously optimizing multiple objectives, including distance, sparsity, and prediction confidence. This demonstrates a balanced handling of trade-offs, yielding more plausible CEs, particularly in high stake domains where feature interactions are crucial. This also suggests that the proposed MoCE provides a balanced and clinically meaningful alternative to existing methods.