Manutenção automática de um sistema de auditoria inteligente em uma seguradora de saúde

Abstract

The use of machine learning to solve real problems provides an increase in the speed of data analysis and overlaps the human bias with greater assertiveness in activities such as prediction, classification and optimization. The model generated by a machine learning algorithm requires periodic adjustments to maintain effectiveness, even if new input data changes over time. The main motivation for the development of this work is the fact that the machine learning models become outdated over time. The objective of this work is to automate the processing of requests at Unimed-BH that cannot be audited automatically and require human intervention. This happens when: (i) there are no clear business rules for these situations or (ii) the request is complex or expensive. Currently, these types of requests are directed to a group of expert auditors who perform manual analysis of the information presented. The company has about 1.2 million customers and audits around 500,000 requests for medical exams and procedures per month. The main contribution of this work includes the creation of a maintenance mechanism for the Unimed-BH electronic audit system capable of monitoring one or more risk metrics and calibrating model parameters automatically. The application of the maintenance mechanism dynamically monitors in real time the effectiveness of the analysis of requests originally sent for human audit, aiming to maintain minimum effectiveness parameters that minimize the impacts of automation and act according to the current scenario. If the model approves a request that should be denied, a financial loss is generated that involves a risk metric that regulates the automation rate. The relationship between the automation rate and financial risk uses the Theory of Derivative Integrated Proportional Controllers so that the proposed system is more permissive in times of little financial loss and allows to increase the automation rate. Similarly, it is desirable that the system becomes more prohibitive in cases where the performance of the model deteriorates, which allows to dynamically control the performance of the model.