Counterfactual Explanations for Gallstone Prediction: A Comparative Evaluation of DiCE and PertCF
International Conference on Computer Sciences, Zagreb, Hırvatistan, 17 - 18 Eylül 2026, ss.1499-1507, (Tam Metin Bildiri)
- Yayın Türü: Bildiri / Tam Metin Bildiri
- Basıldığı Şehir: Zagreb
- Basıldığı Ülke: Hırvatistan
- Sayfa Sayıları: ss.1499-1507
- Açık Arşiv Koleksiyonu: AVESİS Açık Erişim Koleksiyonu
- Gazi Üniversitesi Adresli: Evet
Özet
Abstract Gallstone disease is a common gastrointestinal disorder for which non-invasive pre-imaging risk prediction using routine clinical and bioimpedance data may serve as a low-cost screening aid. Beyond predictive performance, counterfactual explanations can show clinicians how changes in input features could alter a model's decision, without implying that such changes represent actionable clinical recommendations. However, the consistency and trade-offs of different counterfactual generation methods remain insufficiently characterized in gallstone prediction. This study compares two counterfactual explanation methods, DiCE and PertCF, using 31 test-set patients predicted as gallstone-positive by a tree-based ensemble trained on clinical and bioimpedance data from 319 patients. The methods were evaluated using matched, paired statistical tests under a common backward-elimination post-processing protocol. DiCE and PertCF did not differ significantly in sparsity (median 2.0 features for both, Holm-adjusted p=0.313) or proximity (Holm-adjusted p=0.544). In contrast, PertCF was substantially faster (median 0.625 s vs. 3.80 s; Holm-adjusted p < 0.001). PertCF also exhibited a numerically higher rate of one-sided changes within strongly correlated feature pairs than DiCE (54.8% vs. 38.7%), although this difference was not statistically significant (exact McNemar p=0.125). The ensemble classification backbone showed no significant advantage over a simpler two-model average under nested cross-validation (p=0.206). SHAP analysis and univariate comparisons provided convergent evidence highlighting C-Reactive Protein and Vitamin D. Overall, neither counterfactual method was uniformly superior: PertCF demonstrated a clear computational advantage, while differences in the preservation of strong empirical feature dependencies remained statistically inconclusive. Keywords: counterfactual explanations, DiCE, PertCF, explainable AI, gallstone disease, clinical decision support.