Lydia AI: three ways for a quicker acceptation of machine learning to speed up underwriting
The promise of machine learning to make underwriting digital, faster and more accurate is well understood. However, the lack of a standardized validation framework to help machine learning scientists, underwriters and actuaries collaborate efficiently has slowed adoption. Machine learning scientists train models that are often left idle in sandbox environments or small scale adoption.
Lydia AI, an AI insurtech startup helping insurers use alternative data to make life and health insurance products easier to buy and more inclusive, recently worked with Taiwan Life to launch a series of machine learning models into production. The machine learning models predicted different health indicators that united into a health score that could potentially qualify candidates for accelerated underwriting, instead of lengthy medical examinations. Their work was recognized by Celent with the Model Insurer of the Year award for data, analytics and AI.
The Lydia AI team has gathered 3 techniques that they used to launch machine learning models into production faster:
1. Understand and adapt validation method to key stakeholders
There are 3 key stakeholders to validate a machine learning workflow: the actuary, the underwriter and the machine learning scientist. These stakeholders each have a different way of working with data and different expectations of how machine learning models should perform. Hence, the validation framework needs to be adapted to the specific needs of each stakeholders. For example, comparing underwriters with actuaries: underwriters expect ML models to achieve a high level of agreement with underwriters ) whereas actuaries expect ML models to on average conform with actuarial experience.
Clear communication between stakeholders and mutual understanding on their differences helps to facilitate the validation process. The team formulated a validation framework incorporating the needs of all stakeholders and get the model into production as seen in Figure 1.
2. Incorporate actuarial standards into machine learning models
Machine learning scientists tend to tune models to optimize for predictive performance, based on the data the models were trained and tested on. However, models with high predictive performance measured by the machine learning standard metrics do not necessarily make sense to actuaries. Machine learning scientists need to translate model performance into incidence rate plots that show the likelihood of an event (i.e. mortality, chronic illness, cancer) for each health score.
Actuaries expect incidence rate patterns that can be compared against industry benchmarks: the incidence rates should rise as health scores rise (less healthy), and people with the lowest health scores should have lower incidence rates than the benchmark. In order to achieve this, ML scientists must create customized actuarial objective functions to emphasize the expected pattern. That means, machine learning models will often not be selected only based on the highest standard model performance metrics .
3. Use underwriting simulation to compare machine versus human decisions
The confusion matrix is a familiar tool for machine learning scientists to compare their model’s predictive results versus the correct answers. The Lydia AI team adapted the tool to compare the machine learning’s predictive results with an underwriter’s decisions on the same cases.
The comparison provided underwriters with confidence that the health score can be trusted as a strong reference marker. The Lydia AI team recommends using disagreements between machine learning models and underwriters to spark discussions on how to improve models and the underwriting process as demonstrated in the diagram below. Most notably, with the models making medically explainable decisions, the Lydia AI team is able to convince human writers to revise their previous decisions to deny coverage for some people, thus expanding the population that can benefit from insurance protection.
About Lydia AI
Lydia AI is a health AI insurtech startup, on a mission to insure the next billion people, by making insurance personalized, easier to buy and more inclusive. Insurance companies tap into the company’s Lydia AI risk prediction engine to make accurate dynamic health risk predictions based on alternative data. These actuarially validated health scores are used to make personalized product recommendations, accelerate underwriting and create new insurance products.
Established in 2015, Lydia AI, formerly known as Knowtions Research, is backed by Alibaba Entrepreneurs Fund, Information Venture Partners and 500 Global. Lydia AI has offices in Toronto and Taiwan. In 2022 their work was recognized by Celent with the Model Insurer of the Year award for data, analytics and AI.
For more information about Lydia AI, click here! For more information about Lydia AI, listen to Lydia AI machine learning scientists share their experience at the Toronto Machine learning Summit here or read the Celent analyst paper here.