Interpretable AI

A key subgenre of explainable AI is ‘interpretable AI.’

Interpretable AI is the ability for an underwriter or a regulator to be able to audit an individual application at the case level to determine why the model made a particular decision for a particular consumer. This means that we're able to look at the thousands of features that were potentially used as part of the model decisioning for example, medical disclosures, smoker status and other information that may have been received from an APS (Attending Physician Statement) or EHRs (electronic health records) and explain to a consumer why a particular decision was made, either favourably or unfavourably.

This set of combined tools - bias detection, model drift detection, and interpretability - are all part of the overall ecosystem that Munich Re has included when we designed our predictive models.

It is really about the culture of building responsible AI into all your processes and procedures from the beginning and ensuring you meet your regulatory requirements.

The ultimate philosophical question when we discuss predictive models is what is an actual predictive model? Is it the data or is it the algorithm? And ultimately, it is a mixture of both. 

One does not work without the other. The historical data contains the existing philosophy of your underwriting decisions, and the algorithm optimises it in a way that can be used for decisioning or recommendations. When we talk about data, it is key that we have data in an open fashion.

As we discussed in the original article in this series, one of the main challenges today for insurers is siloed and unstructured data. It comes in many different formats and is difficult to interpret by either a data scientist, or an underwriter. It is key then when we talk about data, we talk about data as an asset. And once we determine that data is an asset, it should be treated as such. 

Once data is obtained, it should be labelled and structured in a way that can be understood by data scientists and other consumers, making it a valuable asset. In recognition of its value, it should come with its own SLA (service level agreements), like any other product. And it is key that it comes with data dictionaries and metadata analysis that allows you to understand what the fields are as they are working.

For example, across multiple different siloed systems, gender, or date of birth, may be in different formats that can be difficult to link, post-analysis. Another key aspect of an open data strategy is ensuring that data is labelled and accessible in easily usable trans-consumable ways, for example via APIs (Application Programming Interfaces).

And finally, when data is to be treated as an asset it should be clean, validated and understood. This is key to success. Using unvalidated data can result in biases or other unexpected outcomes from predictive models.

And so, we have come full circle. To reiterate the challenges insurers are facing, which we’ve detailed in this series of articles, insurers need to ask themselves:

Do we have the correct tools in place? And do we have the talent to be able to achieve our goals?

Is our data labelled, structured and clean with a strongly understood provenance? Can it be used and consumed easily by data scientists? Does it contain any implicit or explicit biases that we need to resolve so as not to negatively affect outcomes?

Are we compliant within the existing regulatory environment? Is our data model explainable, auditable, and interpretable? 

Ultimately, the decisions that are made by AI model must be fair to the customer at the end of the process. Our AI challenges can also be our AI opportunities and forward-thinking insurers have a generational opportunity to gain a competitive advantage by using predictive and descriptive analytics to their fullest potential.