“Process” – art by blprnt_van on Flickr.
Fraudulent, wasteful, and abusive behaviors account for an estimated one-third of the $2.2 trillion spent on healthcare in the US each year. Unnecessary medical interventions contribute to healthcare cost inflation and can be harmful to patients. A fractional reduction in the number of wasteful procedures can yield far greater savings than direct cuts in care and coverage. Industry acknowledgement of widespread waste in healthcare has led several medical professional societies to publish “Top 5” lists of the most over-utilized and unnecessary procedures as part of the Choosing Wisely campaign.
Unfortunately, medical fraud, waste, and abuse is difficult to detect. Fraud/Anomaly Detection is hard in any context since the modeler needs to deal with imbalanced class distribution (rare event detection) However, fraud detection is a tougher challenge in healthcare than in financial services, where there is usually a victim who can flag a fraudulent transaction. In the case of healthcare, the patient does not have the expertise to judge the necessity of an operation and hence the modeler cannot have access to labeled data and cannot utilize supervised learning techniques. Waste is even more difficult to detect, as it requires a subjective judgment to be made about the medical necessity of a treatment. Despite attempts to reduce variation in medical necessity definition and decision making, there is still little agreement on medical necessity among professionals.
Rather than utilizing existing “pay-and-chase” approaches to detecting fraud and abuse, a Specialty Benefits Management Company (SMB Company) worked with Pivotal Data Labs to catch waste and fraud before it occurs.. Adjudicating billions of dollars in claims annually on behalf of many large insurance providers, the SMB Company aimed to triage claims in real-time, during the authorization process.
The SBM Company boasts expert knowledge of what questions to ask about patient history and symptoms being exhibited, that allow it to determine whether a procedure is medically necessary or not. Over years the company has collected a very unique data set that is much richer than claims data alone. The challenge lied in identifying fraud, abuse and waste before authorization and payment, while employing safeguards to protect medical resources and patients’ well-being.
Pivotal Data Labs proposed a predictive approach, modeling outcomes across a range of contexts by using historical information about patients and the providers responsible for ordering a procedure. The model we developed accounts for patient demographics, past procedures and treatments, diagnoses, and the ordering physician’s utilization patterns. We then asked: “How different does the current patient look, compared to all other patients under similar conditions and circumstances?” For example, whether a patient has diabetes may be relevant to one service request, but not another. We needed to be able to create custom feature sets for the particular service requested. Cases inconsistent with the model’s expectation would then be flagged for further investigation.
When considering tens of procedures and hundreds of pathways for these procedures, building custom models with custom feautures can quickly become a daunting task. However, our methodology of using regularized regression for feature selection for each model and building models for pathways in parallel using in database analytics allowed us to create hundreds of custom models with custom features within minutes.
We implemented fraud, waste, and abuse detection algorithms across the entire cardiology line of business, which represents 60 different medical procedures. In order to measure the potential impact we have on our customer’s business we focused on a cardiac imaging diagnostic called Myocardial Perfusion Imaging, a procedure listed as the second most over-utilized procedure by the American Society of Nuclear Cardiology 1 and the highest volume procedure for SMB Company. Following just this procedure alone, the SMB Company was able to detect thousands of anomalies and to prevent millions of dollars in medical waste and abuse.
Besides the monetary impact, there are substantial benefits to patient health and safety by preventing unnecessary medical procedures, which can be invasive and pose substantial risks. With the success of the initial model, we then applied the same approach to all 60 procedures in the cardiology line of business with hundreds of models trained on historical data and stored in the database. The predictive approach allows SMB Company to prioritize cases for a physician to review and triage, and identify waste and abuse before authorization, resulting in reduced physician workload and improved patient care.
1 H. Brody, “From an Ethics of Rationing to an Ethics of Waste Avoidance,” pp. 3–5, 2012
About the Author
BiographyMore Content by Noah Zimmerman