AI Academy

The project develops a system to classify whether patients have taken their medication. The system analyzes data from an IMU device (Inertial Measurement Unit) that contains a gyroscope and an accelerometer. By interpreting sensor data, the system can automatically detect when patients take their medication. The main purpose is to ensure that psychiatric patients follow their treatment plans and take their medication as prescribed.

In the long term, the project aims to enable deeper studies of the relationship between medication and patient well-being through so-called unsupervised learning. This method can provide valuable insights for researchers into treatment effects and behavioral patterns.

AI Academy participant Arian Behtoui is working on this project, which is a collaboration with the Department of Biomedical and Clinical Sciences (BKV) and the company Seber Medical.

The project focuses on further developing the existing software DyKnow, which is used to reason about information streams in real time. Information streams are continuous flows of data coming from various sources, such as sensors, games, or networks. DyKnow analyzes these flows to draw conclusions about what is happening in the system.

The goal of the project is to give DyKnow a new capability – not only to analyze data but also to perform actions. Actions mean that the system can act based on its analysis, for example by sending a command, changing a setting, or influencing a process in the environment. These actions are controlled by logical formulas expressed in ProbSTL (Probabilistic Signal Temporal Logic), a language that makes it possible to formulate requirements for information streams even when those requirements are uncertain or stochastic. Stochastic simply means that something involves an element of randomness or probability, allowing the system to handle uncertainty in data.

AI Academy participant Leo Jarhede is working on the project. To demonstrate the concept, he uses data from the real-time strategy game StarCraft. DyKnow receives information streams from the game, such as the number of units or the amount of minerals, and converts them into the type of data needed to evaluate ProbSTL formulas.

“If a formula shows that all requirements for building a certain unit are met, an action is sent to the game to actually build the unit,” says Leo Jarhede.

This work demonstrates how stream reasoning can move from passive analysis to active control, which could have major implications for future AI systems that need to act in complex, dynamic environments.

This project is led by the Swedish National Courts Administration, in collaboration with Linköping University (LiU), the Swedish Defense Research Agency, and the Swedish Police Authority. AI Academy participants Nils Alenäs, Aleksi Maxim Andreev and Victor Lagerbring are working on the project. Its goal is to automate the protection and anonymization of sensitive information in legal documents.

Today, anonymization, sometimes called “masking”, is done manually by lawyers and administrative staff. This process takes a lot of time and resources. Automating it would free up valuable time for legal professionals and make the process faster, more accurate, and consistent across all courts. If successful, this solution could also be used by large companies, government agencies, and even the military and changing how sensitive data is handled everywhere.

The project uses AI technology based on Swedish BERT, a language model trained to understand Swedish text. It combines this with NER (Named Entity Recognition), which is a technique for finding specific types of information in text, such as names of people (PERSON), organizations (ORG), locations (LOC), and codes. Once these details are detected, they will be anonymized using methods like pseudonymization (replacing real names with fake ones) and format-preserving masking (keeping the same structure so the document looks natural). If the system is unsure, a human will review the case to ensure accuracy.

Because legal documents contain highly confidential information, the system will only be deployed within each individual court. It will not share data between courts, and the AI will only be trained on public and synthetic data, never on real court cases, so that privacy is fully protected.

To ensure high quality, the system will be tested using metrics like precision, recall, and F1-score. These are standard measures in AI that check how accurately the system finds and anonymizes sensitive information. Combined with user feedback, these evaluations will help guarantee that the final solution is both accurate and reliable, reducing workload and strengthening legal certainty.