Thesis (M.Sc)-Cairo University, 2024.

Bibliography: pages 138-143.

Healthcare fraud is a persistent and expanding issue in modern healthcare
systems, resulting in substantial financial losses and negatively impacting the
equitable distribution of medical resources. Documented practices include false
billing, cost inflation, and claims for services not rendered. The transition to
digital health records and claim systems has improved efficiency but also created
new pathways for complex fraud that elude traditional auditing methods.
To address this challenge, a scalable and accurate fraud detection framework
based on Machine Learning (ML) techniques was proposed. The study involved
the evaluation of nineteen supervised classification models, including linear
classifiers, neural networks, and ensemble methods such as XGBoost and Extra
Trees. Structured claim data was used as input, with suspicious patterns targeted
for identification.
Given the high-class imbalance typically found in fraud detection datasets,
several resampling techniques such as Synthetic Minority Over-sampling
Technique (SMOTE), Adaptive Synthetic Sampling (ADASYN), Random Over
Sampling (ROS) and Random Under-Sampling (RUS) were applied.
Additionally, statistical features were engineered to extract behavioral insights
from financial variables, resulting in enriched datasets. A total of ten
experimental configurations were implemented, with the best-performing
models achieving up to 99.99% accuracy when combining feature engineering
and resampling.
The experiments were conducted using a publicly available dataset titled
Healthcare Provider Fraud Detection Analysis, developed in 2022 by the
University of California and published on Kaggle. As a practical reference, the
Egyptian Company for Metro (ECM) was cited to demonstrate the system’s
potential applicability in large institutional healthcare environments.
The study concludes that ML offers a powerful tool for detecting healthcare
fraud, especially when combined with effective preprocessing and domain-
aware feature construction. The proposed framework provides a solid foundation
for developing future real-time fraud detection systems that can enhance
decision-making and reduce operational losses in healthcare delivery.

تتناول هذه الدراسة تطوير إطار اكتشاف الاحتيال في الرعاية الصحية باستخدام تقنيات التعلم الآلي (Machine Learning) مع التركيز على المعالجة المسبقة هندسة الميزات. تم اختبار 19 نموذجًا تصنيفيًا مثل XGBoost وExtra Trees باستخدام بيانات المطالبات، مع تطبيق تقنيات SMOTE وADASYN وROS وRUS لموازنة البيانات. حققت النماذج أداءً عاليًا بدقة وصلت إلى 99.99% عند دمج تحسين الميزات وإعادة التوزيع. تقدم الدراسة إطارًا عمليًا لمنع الاحتيال في مؤسسات ذات بيانات مطالبات كبيرة ومعقدة، مع استخدام الشركة المصرية لإدارة وتشغيل المترو (ECM) كنموذج تطبيقي لدراسة إمكانية تطبيق الإطار في بيئات مؤسسية ذات طابع تشغيلي

Issues also as CD.

Text in English and abstract in Arabic & English.