97 Innovative approaches to proactively identify members with special medical needs

Abstract

Objectives Bupa’s purpose is longer, healthier, happier lives. We do this by providing a broad range of healthcare services, support and advice to people throughout their lives. Bupa is commited to becoming the most customer centered health and wellbeing organisation in the world. Meeting a patient’s individual care needs is right at the heart of this comittment. The objective is to develop a predictive model which accurately identifies patients whose claiming behaviour is likely to escalate in the near future. This allows for timely referral to specialist support nurses, medical directors/forums for discussion and input and case coordination to help the most vulnerable patients at their time of need, potentially avoiding unnecessary treatment at the same time. Method The modelling dataset is a sample of a half a million patients with Bupa PMI cover who claimed or were due to claim on their policy in last 15 months and more than 650 indicators. The indictors included member demographics and claims-based variables with severity (claimed amount), frequency (number of care episodes), and timing (months since last treatment) aspects. Taking into consideration the business needs, we wanted to create a model that generates both accurate predictions and meaningful ‘insights’, which could be converted into triggers for patients’ case management. We considered several traditional statistical methods (logistic regression) and more innovative machine-learning techniques (mainly tree based models). The latter can capture very complex relationships and therefore be more accurate but often lack insights. Results Compared to the traditional method we ran, tree based algorithms, in particular xgboost, provided the highest accuracy, with 2 out of 3 patients correctly classified. Despite the general belief that machine-learning models are considered ‘black boxes’, we were able to generate 3 levels of insights: • A list of the most important factors at a population level (age, previous cancer claim, etc.). • Insights at individual indicator level. For example, we found that once over 55, a patients’ likelihood of their care escalating increases dramatically. • The contribution each indicator has on patient level to their individual probability. Conclusions This talk demonstrates the value and potential applications of predictive modelling in the UK private medical settings. Such an application enables us to create triggers for case management, pathways tailored for an individual patient, and potentially avoiding unnecessary treatment.