Machine learning – based financial fraud detection: Evidence from healthcare and related firms in Turkiye

Financial fraud and accounting irregularities distort financial transparency, weaken corporate governance, and reduce the efficiency of resource allocation.Machine learning techniques have emerged as tools for detecting complex fraud-related patterns that traditional approaches may fail to capture.This study examines the effectiveness of machine learning in identifying financial fraud risk among healthcare and healthcare-related firms in Türkiye.Using firm-level data covering 2021-2025 (279 firm-quarter observations, 15 firms), six algorithms -Logistic Regression, Naive Bayes, k-Nearest Neighbors, Support Vector Machine, Decision Tree, and Random Forest -are evaluated through firm-level five-fold cross-validation with hyperparameter tuning and explicit class-imbalance handling.Fraud-risk indicators are derived from the Beneish framework, while predictor variables are kept independent from the fraud-risk label to avoid target leakage.The findings reveal a clear trade-off: Random Forest achieves the highest accuracy and specificity (72.4% and 86.7%) but the lowest sensitivity (21.3%), whereas Decision Tree achieves the highest sensitivity (60.7%) and F-score (0.354) despite lower accuracy (51.6%).No single algorithm dominates across all criteria.The pruned decision tree indicates that changes in gross profit margin and quarterly net profit margin are the signals most consistently associated with elevated fraud risk.The results highlight both the potential and limitations of datadriven approaches for financial monitoring, showing that model selection should depend on the relative cost of false negatives versus false positives.Rather than replacing auditor judgment, the evaluated models are best understood as decision-support tools that help prioritize firms for closer review within a transparent, reproducible framework.