PretopoMD: The Key to Making AI Transparent by 2027

The AI Act and Its Requirements for AI

The AI Act, a European legislation set to come into effect between 2026 and 2027, imposes transparency and explainability requirements on artificial intelligence systems deemed high-risk. This regulation aims to ensure that algorithms, particularly those used in sensitive sectors such as healthcare or justice, do not operate as "black boxes." Companies that fail to comply with these rules could face severe penalties, reaching up to €35 million or 7% of their global annual revenue.

Currently, many high-performing algorithms, especially those based on deep learning and neural networks, produce effective results but remain incomprehensible in their internal logic. This opacity poses major problems, particularly when critical decisions need to be explained, such as in the case of a medical diagnosis or a judicial decision.

Pretopology: An Approach for Explainability

To address these challenges, pretopology offers an innovative method that allows for the creation of readable "neighborhoods" around each data point. This approach uses disjunctive normal form rules to process mixed data in an explainable manner. PretopoMD, an open-source implementation of this method, is currently being proposed for pilot studies in areas such as healthcare, human resources, and maintenance.

For example, in the medical field, an early cancer detection algorithm could analyze various factors such as hemoglobin levels, family history, smoking status, and a patient's level of physical activity. However, if a doctor requests explanations regarding a patient's risk classification, the current system may fail to provide a clear answer, thus illustrating the "black box" problem.

The Limitations of Current Methods

Current clustering methods, such as hierarchical clustering, are often used to organize similar observations into a hierarchy of nested groups. This allows experts to identify typical profiles without having to open the black box of a predictive model. However, these methods encounter difficulties when it comes to mixing numerical and categorical data.

For instance, measuring the distance between two numbers is relatively straightforward, but assessing the proximity between two categories, such as "smoker" or "non-smoker," is more complex. Existing solutions, such as k-means, HDBSCAN, and DIANA, all have limitations. Transforming categories into artificial numbers can introduce biases into similarity calculations, making the results less reliable.

PretopoMD: A Promising Solution

To overcome these challenges, PretopoMD proposes defining "neighborhoods" using disjunctive normal forms. For example, a patient can be included in the neighborhood of a group if certain conditions are met, without requiring artificial distance measures. Once the neighborhoods are defined, iterative calculation allows for determining the set of patients that "adhere" to each group. This process results in a hierarchy of successive groupings, without imposing artificial distances between categories and numbers.

This approach could revolutionize the way AI systems process and explain data, particularly in critical fields such as healthcare and human resources. PretopoMD thus offers a potential solution for making algorithms more transparent and compliant with upcoming regulatory requirements.