All in on AI, Transparent & Explainable

In terms of explainability, machine learning models can be divided into two general classes: black-box and white-box.

One of the primary challenges faced by AI developers is this trade-off between model accuracy and explainability. On one hand, the predictive accuracy of AI models should be a priority, so that they can identify complex, nonlinear relationships between variables, and provide valuable insights that drive informed decision-making.

However, the more sophisticated an AI system is, the harder it becomes to explain how it operates, which can negatively affect the integrity of its outputs.

For example, AI models are vulnerable to biases stemming from unrepresentative data, leading to outputs that can perpetuate inequitable outcomes. Furthermore, AI models can experience “model drift,” a phenomenon in which model performance degrades over time because real-world data differs from the data the model was trained on. A lack of explainability can hinder human operators from monitoring model outputs, lead to poorly informed decision-making, and undermine trust in AI systems.

Explainability allows developers to tackle these issues, allowing us to shine a light into the “black box” of AI. Depending on the use case, many strategies to improve the explainability of AI models have been explored, including the usage of comprehensible text, approximations, or visualizations. At Sanofi, all AI Systems comply with our documentation standards during the development phase.

Adverse events are a major concern in many clinical trials. They are a common contributor to clinical trial failures, and their causes are often complicated and difficult to untangle.

At Sanofi, we are using AI to predict trial participants that are at high risk or low risk of an adverse event. Our developers are prioritizing transparency and explainability by using glass-box models, which are models with interpretability built in. They pair the model with a model card which captures information such as how the model was trained, the model’s characteristics, the model’s performance and the model’s inferred outputs.

At Sanofi, we understand that transparency and explainability are vital in ensuring that our AI systems are trustworthy, accountable, and aligned with emerging AI regulations. Ensuring that the outputs and decisions of our AI systems are understandable allows us to navigate the future of AI responsibly.