for Research

Reducing patient burden and research cost: the AI4Nof1 project leverages cutting edge developments in (causal, neurosymbolic) reinforcement learning, psychometrics and digital epidemiology to build adaptive personalised treatment regimes for chronic conditions, simultaneously identifying phenotypes and causal pathways while minimising the time and number of measurements needed for patients to find a treatment that is right for them.

curATime is a collaborative, multi-centre effort applying AI methods to high-dimensional biomedical data to promote an understanding of biological processes associated with cardiovascular illness. This exciting project involves highly granular data collected as part of the Gutenberg Health Study — a prospective cohort study of a representative sample (N = 15000) of the population of Mainz — including genotyping, DNA methylation, transcriptomics, proteomics and extensive time-varying clinical information.

The project explores methods of selecting and extracting features from whole genome data for survival problems (individual health), the effect of different levels of data aggregation on spatiotemporal models (societal health) and the challenges behind building a transparent, explainable and useful data science pipeline for such complex temporal data. A particular focus is paid to automation of the data cleaning process, as well as to effective combination of scientific and machine learning knowledge on problems where data are messy, sparse or only available in aggregate form.

The aim of this project is to develop a new form of AI: physics-informed deep anomaly detection. This involves utilising state-of-the-art neural anomaly detection algorithms, which have recently drastically reduced error rates in common benchmarks. Neural network architectures are being developed that specifically incorporate information from the data domain to further increase data efficiency. In addition, physics-informed generative AI methods are being developed to expand and improve the existing training data. The application of the new physics-informed anomaly detection methods is being evaluated specifically in additive manufacturing, particularly with regard to efficient process execution.

Managing a large building requires knowing exactly what equipment is inside it, but this vital data is usually trapped in thousands of pages of messy blueprints, tables, and text documents. Traditionally, compiling this information into a usable register takes weeks of tedious, manual paperwork and costly site inspections. This project changes that by using AI to automatically read, extract, and organize this hidden data from your existing documents. By turning a chaotic mountain of paperwork into a structured digital catalog, this project aims to save massive amounts of time, cuts maintenance costs, and seamlessly connect with modern digital building software to optimize energy efficiency.

The TrustifAI project aims to contribute a set of concrete solutions to improve trustworthiness of AI applications in health and wellbeing at various stages of development lifecycle. A quality platform for development of trustworthy AI applications will enable users to create efficient and effective data science analytics pipelines through a human-in-the-loop approach with the goal of increasing trustworthiness.