Finding Anomalies in Medicare

Abstract

The general behaviour of a medical practitioner can be assessed by metrics related to the services rendered. The values of these variables are inter-related and also related to the demographics of the practice. If there is a functional form that relates specified metrics to the others and the demographics, machine learning techniques can be used to determine the function. Both boosted regression trees and feedforward neural networks trained with backpropogation are used to learn a functional form. Where the predicted value of the function does not match the actual one, the behaviour is anomalous. This may be because there is fraud. Another technique for determining atypical behaviour is to reconstruct a value from the significant principal component scores, and compare this to the actual ones. Self organising maps can also be used to find similar, and dissimilar patterns. Both the targets themselves and the Karhunen-Loeve transform are used to train self organising maps. The coordinates on these maps are used to compare the doctors identified by the various techniques.