Machine learning algorithms prove their value in distributing healthcare premiums
Algorithms have become indispensable in our lives: from targeted offers from your supermarket to support for doctors in making medical diagnoses. Gupta Strategists and i2i recently explored a new application of algorithms within the healthcare system on behalf of the Ministry of Health, Welfare and Sport. What did we learn? Machine learning algorithms are also highly valuable in predicting healthcare costs.
Predicting cost of care is crucial for Dutch healthcare system
Risk adjustment is an indispensable component of the Dutch healthcare system. In fact, our risk adjustment scheme is regarded worldwide as the best system for enabling solidarity in healthcare. In short, it ensures that health insurers are financially compensated for the healthcare needs of their policyholders. The contribution insurers receive is based on the predicted healthcare costs per person. Thus, an insurer with mainly elderly people in poor health receives a higher contribution than an insurer with many healthy young people. Without risk adjustment, (chronically) ill people would have to pay very high premiums or become uninsurable.
Demonstrably better predictions with machine learning
Our research ‘Exploring the use of machine learning in health insurance risk adjustment’ shows that the application of machine learning algorithms significantly improves the prediction of healthcare costs. We conducted this research together with experts from i2i and on behalf of the Ministry of Health, Welfare and Sport. In the study, we show that with machine learning the predicted healthcare costs are closer to the actual costs than with the current risk adjustment system. To be precise: predictions are more than 3 percentage points better, measured in R2. We were able to reduce the spread between insurers from 310 euros per insured person to 235 euros.
Advantages and disadvantages of different algorithms
In the study, we compared different machine learning algorithms. The neural network and the GBM algorithm performed best. The predictive power of both algorithms was significantly better than that of the classic linear regression. That’s because with machine learning we are better able to capture the interactions between different characteristics of insured persons, like age, chronical disease and socioeconomic status. However, these algorithms also have disadvantages with regards to required computing power and interpretability of results. Moreover, they require adjustments to the current way of working. In the research we describe the working mechanism of different algorithms, the results in adjusting effect and the advantages and disadvantages. The source code is also available for download.
Would you like to know more about the application of machine learning in healthcare?
Please contact Roxanne Busschers or Daan Livestro.
Download the full report "Exploring the use of machine learning in health insurance risk adjustment" (in Dutch) here » Download the full report "Exploring the use of machine learning in health insurance risk adjustment" »