Klasifikasi Risiko No-Show Pasien Berdasarkan Karakteristik Penjadwalan dan Sosiodemografi Menggunakan Kombinasi Instance Hardness Threshold dan Symbolic Regression

Abstract

Patient non-attendance at scheduled medical appointments (no-show) places significant burdens on healthcare systems, including resource inefficiency, disrupted care continuity, and substantial financial losses. Medical appointment datasets are inherently imbalanced, making accurate identification of the no-show minority class challenging with conventional classification methods. Moreover, most prior studies have primarily focused on improving predictive accuracy through complex black-box models, while model interpretability for clinical decision-making has received comparatively little attention, leaving such models unable to produce explanations that healthcare personnel can directly act upon in operational settings. This study proposes a two-stage classification pipeline integrating Instance Hardness Threshold (IHT) and genetic programming-based Symbolic Regression (SR), applied to the no-show Appointments dataset from Kaggle (N=110,527 records, Vitória, Brazil) with 14 sociodemographic and scheduling features. IHT selectively eliminates ambiguous majority-class instances, while SR produces explicit mathematical formulas that can be read and translated directly into intervention priority rules without requiring additional interpretability tools. IHT reduced the class imbalance ratio from 3.95:1 to 1.08:1 by removing 44,790 instances. SR identified waiting_days and same_day_appt as the most critical feature pair, generating an explicit symbolic formula that achieved 92% recall, an F1-Score of 0.43, and a ROC-AUC of 0.67 on the test set, competitive against the best-performing Gradient Boosting baseline (F1=0.43, ROC-AUC=0.73) with markedly superior interpretability. The SR formula can be directly operationalized as a reminder dispatch priority rule, patients with longer waiting times who did not book same-day appointments are prioritized for intervention, while same-day appointment patients exhibit near-zero no-show risk. Future work is recommended to explore multi-objective SR, contextual feature integration, and validation on Indonesian clinical datasets.