Bernhard Schrupp

Bernhard Schrupp

Master's Thesis

Simulation of Inpatient Claims Data and Fraud Detection from a German Health Insurance Perspective


René Raab (M. Sc.), Kai Klede (M. Sc.), Prof. Dr. Björn Eskofier


06 / 2023 – 11 / 2023


In 2021, 11,328 cases of healthcare fraud were reported in Germany [1]. The AOK Bundesverband quantifies the secured insurance claims (proven fraudulent behavior) in hospital treatment at 10% of a total of 37 million euros. This contrasts with 35.1% of total AOK expenditures for hospital care, suggesting a high rate of undetected false claims. However, the literature on healthcare fraud detection focuses on US data (Medicare), Taiwanese data (NHI), or other private data, and detection approaches based on individual claims are virtually nonexistent [2].

Detecting healthcare fraud is difficult because health insurers are prohibited from sharing data on individual treatments. Audits of inpatient hospital billing are primarily performed by the ‘Medizinischer Dienst’ (MD) at the request of health insurers under § 275c SGB V [3]. However, the health insurance funds must meet hospital-specific quotas for the allocation of audits to the MD.

Hospital claims are calculated by cumulating gender, age, diagnoses, operations, and length of stay. These parameters are grouped by certified tools, called DRG Groupers, which generate a Diagnosis Related Group (DRG) and its base ratio (in points). This base ratio is multiplied by a state-specific value per base ratio point. The result is the rate paid by the insurance company. Small changes in diagnoses or operations, like changing the importance of a diagnosis, can have a large impact on the base ratio. DRG upcoding occurs when these changes cannot be justified by medical reality [4].

The goal of this thesis is to detect fraud in hospital claims based on billing information available at health insurance companies, to detect cases of not or only partially provided services (typical cases of DRG upcoding), and to detect billing of treatments that are not necessary for healing (e.g., prolongation of ventilation time) [5, 6]. Therefore, we develop an inpatient care billing data simulator and validate its statistical properties using internal data from AOK Bayern. We conduct expert interviews to identify and inject common fraud schemes into the simulation. Finally, we develop and evaluate machine learning models to detect healthcare billing fraud.

To achieve this goal in a privacy-preserving manner, we use a Monte Carlo approach to generate synthetic billing data from publicly available hospital treatment data. These public data consist of information provided by the German Federal Statistical Office on the number of primary and secondary diagnoses [7, 8] as well as operations and procedures [9]. Using these data, probabilities for hospital treatments are calculated and randomly assigned to individual treatment cases in an agent-based model [10]. The resulting claims are calculated by using a certified DRG Grouper.

Discussions are ongoing with AOK Bayern as a potential data provider on how to validate the data. The current plan is to validate the generated data by comparing the distribution of values with claims data provided by AOK Bayern to ensure comparability.

The so-generated and validated claims data is then analyzed for irregularities and models are  trained using various supervised and unsupervised Machine Learning approaches drawn from the existing literature on hospital claims fraud [11, 12], such as:

  • Deep Neural Networks
  • Support Vector Machines
  • k-Nearest-Neighbor
  • Gradient Boosting
  • Logistic Regression

Although the area of inpatient healthcare fraud in the German market has not been well researched, several problems can be expected. First, it is not clear how many cases and what types of inpatient billing fraud exist. The estimated number of unknown cases is likely to be higher than the known number because the cases detected to date have been detected by trained humans. Second, the number of cases of potential fraud is likely to be much smaller than the number of cases of correct billing and, if not properly addressed, will lead to incorrect predictions.

The first challenge will be addressed by interviewing experts in the detection of the aforementioned types of fraud. The second is addressed, if necessary, with different versions of over- and under-sampling. In comparable research – mainly done on credit card fraud – SMOTE [13, 14] shows the most promising results.


[1] Bundeskriminalamt, Ed., “Wirtschaftskriminalität: Bundeslagebild 2021,” Wiesbaden, 2022. [Online]. Available: https://​​/​SharedDocs/​Downloads/​DE/​Publikationen/​JahresberichteUndLagebilder/​Wirtschaftskriminalitaet/​wirtschaftskriminalitaetBundeslagebild2021.html

[2] Li, K.-Y. Huang, J. Jin, and J. Shi, “A survey on statistical methods for health care fraud detection,” Health Care Management Science, vol. 11, no. 3, pp. 275–287, 2008, doi: 10.1007/s10729-007-9045-4.

[3] Sozialgesetzbuch V: SGB V.

[4] Salomon, “Ökonomie und Ethik im Klinikalltag – Der Arzt im Spannungsfeld zwischen Patientenwohl und Wirtschaftlichkeit,” (in ger), Anasthesiologie, Intensivmedizin, Notfallmedizin, Schmerztherapie : AINS, vol. 45, no. 2, pp. 128–131, 2010, doi: 10.1055/s-0030-1248148.

[5]    D. Schirmer, “Betrugsbekämpfung in der deutschen gesetzlichen Krankenversicherung: Beschreibung eines Phänomens und dessen ökonomischer Wirkung,” Masterarbeit MHBA, Friedrich-Alexander-Universität Erlangen-Nürnberg, Nürnberg, 2020.

[6]    European Healthcare Fraud & Corruption Network, The EHFCN Waste Typology Matrix©. [Online]. Available: https://​​/​what-​is-​fraud/​ehfcn-​waste-​typology-​matrix/​

[7]    Statistisches Bundesamt, “23131-0003: Krankenhauspatienten: Deutschland, Jahre, Geschlecht, Altersgruppen, Wohnort des Patienten, Hauptdiagnose ICD-10 (1-3-Steller Hierarchie),”

[8]    Statistisches Bundesamt, “23141-0003: Nebendiagnosen der vollstationären Patienten: Deutschland, Jahre, Geschlecht, Altersgruppen, Wohnort des Patienten, Nebendiagnosen ICD-10 (1-3-Steller Hierarchie),” [Online]. Available: https://​​/​genesis//​online​?​operation=​table&​code=​23141-0003

[9]    Statistisches Bundesamt, “23141-0111: Operationen und Prozeduren an vollstationären Patienten: Bundesländer, Jahre, Geschlecht, Altersgruppen, Operationen und Prozeduren (1-4-Steller Hierarchie),” [Online]. Available: https://​​/​genesis//​online​?​operation=​table&​code=​23141-0111

[10]  E. Bonabeau, “Agent-based modeling: methods and techniques for simulating human systems,” Proceedings of the National Academy of Sciences of the United States of America, 99 Suppl 3, Suppl 3, pp. 7280–7287, 2002, doi: 10.1073/pnas.082080899.

[11]  T. Ekin, L. Frigau, and C. Conversano, “Health care fraud classifiers in practice,” Appl Stoch Models Bus & Ind, vol. 37, no. 6, pp. 1182–1199, 2021, doi: 10.1002/asmb.2633.

[12]  N. Kumaraswamy, M. K. Markey, T. Ekin, J. C. Barner, and K. Rascati, “Healthcare Fraud Data Mining Methods: A Look Back and Look Ahead,” Perspectives in Health Information Management, vol. 19, no. 1, 1i, 2022.

[13]  A. Fernández, S. García, F. Herrera, and N. V. Chawla, “SMOTE for Learning from Imbalanced Data: Progress and Challenges, Marking the 15-year Anniversary,” Journal of Artificial Intelligence Research, vol. 61, pp. 863–890, 2018.

[14]  V. N. Dornadula and S. Geetha, “Credit Card Fraud Detection using Machine Learning Algorithms,” Procedia Computer Science, vol. 165, pp. 631–641, 2019, doi: 10.1016/j.procs.2020.01.057.