Çulcu S., Ekinci F., Akarkamçı H., Demir Ö., Koçkar N., Tamam S., ...Daha Fazla
Digital Health, cilt.12, sa.1, ss.1-40, 2026 (SCI-Expanded, SSCI, Scopus)
-
Yayın Türü:
Makale / Tam Makale
-
Cilt numarası:
12
Sayı:
1
-
Basım Tarihi:
2026
-
Doi Numarası:
10.1177/20552076261501475
-
Dergi Adı:
Digital Health
-
Derginin Tarandığı İndeksler:
Health Research Premium Collection (ProQuest), Scopus, Science Citation Index Expanded (SCI-EXPANDED), Social Sciences Citation Index (SSCI), Directory of Open Access Journals
-
Sayfa Sayıları:
ss.1-40
-
Ankara Üniversitesi Adresli:
Evet
Özet
Objective
This study evaluated a leakage-aware machine learning framework integrating comparative performance assessment and multi-layer explainability for exploratory classification of recorded survival status in gastric cancer patients.
Methods
The cohort included 152 patients, comprising 90 non-survivors and 62 survivors. Twelve classifiers were assessed using an isolated hold-out test set and stratified five-fold cross-validation within the training subset. Preprocessing was fitted exclusively on the corresponding training partitions. Evaluation incorporated discrimination, class-wise performance, calibration curves, and statistical comparisons using Friedman and pairwise Wilcoxon tests with Bonferroni correction. Explainability combined KernelSHAP, feature ablation, counterfactual explanations, and anchor-based rules.
Results
Linear Discriminant Analysis (LDA) achieved the highest hold-out ROC–AUC (0.803), with an F1-score of 0.818 and balanced accuracy of 0.692. LDA also attained the highest mean cross-validation ROC–AUC (0.664 ± 0.103), followed by Passive-Aggressive (0.631 ± 0.122) and k-nearest neighbors (0.626 ± 0.136). The Friedman test indicated overall differences (χ
2
= 20.628, p = 0.037), but no pairwise comparison remained significant after correction. KernelSHAP and ablation identified tumour size as the dominant predictor. Anchor rules achieved maximum precision of 0.93 and maximum coverage of 0.34 across different rules, while calibration curves revealed probability miscalibration.
Conclusions
The framework provides a comparative and interpretable analysis of recorded survival outcomes. Although LDA led the observed rankings, statistical superiority was not established. Limited sample size, resampling variability, and an undefined mortality horizon constrain prognostic interpretation, requiring further validation before clinical application.