AI & machine learning: League’s principles

The importance of AI principles

Artificial intelligence (AI) is a profoundly effective technology. However, it’s incumbent upon those who develop this technology to ensure an ethical framework governs its construction and application. At League, our primary implementation of AI is through machine learning, which powers the data-driven, personalized experience available to every user of our CX platform.

Personalization is a critical component of our platform. As we continue to find new ways to leverage machine learning and improve our offerings, we must also keep ourselves in check to ensure we are implementing personalization in a thoughtful and responsible manner. For example, we must consider the fact that a model based on a singular prototype may not apply to an entire population. We also need to determine when it’s appropriate to communicate with users about our nudges and recommendations (i.e. why am I seeing this?).

Why have AI principles

AI is an innovative way to transfer the heavy lifting of robust data analysis to technology. For example, our Health Journey™ experience is personalized using models that take into account a user’s personal health history when recommending a next-best health action. League understands that a tool this impactful should be operated deliberately and within a self-mandated framework. 

Regulatory bodies in the U.S, Canada and EU are placing an increased emphasis on the ethical usage of AI, especially as it relates to human health. With this in mind, we have established self-imposed principles to ensure our AI usage follows an intentional and directed framework that keeps our diverse user population in mind. This helps mitigate the potential risk of future guidelines and ensures we uphold the mission and values of League.

As our offerings evolve, our principles guide everything we do. Monitoring the impact machine learning can have on our user base is a critical component of our product development. At League, we follow these main principles when it comes to how we use machine learning: transparency, privacy by design and bias evaluation and mitigation.

Our AI principles

Transparency 

Explainability and transparency are top priorities in machine learning model development and user-facing features. Models built using machine learning are complex and operate within the set parameters identified by its creators, but they can also make connections of their own accord. These unknown parameters may be useful in certain scenarios but may result in a model that is not representative of an entire population and all its sub-groups. 

We prioritize transparency when it comes to the questions and suggestions we pose to users. This leads to our data science team taking approaches with a preference towards explainability and making deliberate choices to inform users when an explanation may be warranted. 

Privacy by design

We remain highly committed to keeping information secure and ensure any machine learning usage aligns with our privacy policy. Machine learning models inherently deal with copious amounts of data and, in the healthcare sector, this data requires the highest level of security to protect patient information. Our team is committed to adhering to the patient privacy policy and ensuring any machine learning usage protects this sensitive information. Sometimes, this may mean exploring methods like federated learning, where models can be trained without commingling data from multiple cohorts of users, or even training models specific to users. 

Evaluate and mitigate bias

We endeavor to actively check for performance biases in our machine learning models that may perform differently for certain user populations. A critical piece of mitigating bias in machine learning models is ensuring that the data that these models are trained on represent a sufficiently diverse population. Designing for product inclusivity helps ensure that our technology performs effectively on a number of population groups, thus enabling us to utilize diverse data sets to train our machine learning models and further mitigate potential bias.