Dr. Kata Vuk

From Data to Discovery in the Healthcare Information Age: Interpretable Machine Learning with Piecewise Constant Models

The project ‘From Data to Discovery in the Healthcare Information Age: Interpretable Machine Learning with Piecewise Constant Models’ aims to improve the application of machine learning models in the healthcare sector. The focus is on the development of models that not only enable precise predictions, but are also easy to understand. This is particularly important in the healthcare sector, as medical decisions must be comprehensible and transparent.

The project centres on piecewise constant models, in particular decision trees and change-point models. Piecewise constant models are mathematical methods that visualise complex relationships in a simple form. Decision trees and change-point models help to analyse data, for example to make decisions or identify changes in data patterns. Decision trees are a fundamental component of many machine learning methods and are particularly suitable for tabular data, such as tabular patient data. However, they only offer high predictive power if several decision trees are combined in the form of so-called ensembles, such as random forests or boosting methods. At the same time, this leads to a loss of interpretability in complex combinations. The project is developing approaches to adapt these models in such a way that they both make individualised predictions and remain comprehensible. On the other hand, change-point models, which are used for serial data such as time-dependent vital parameters or genetic information, are optimised to ensure interpretability even at high dimensions. This is crucial for identifying specific changes in a patient’s state of health or in genetic sequences.

In the long run, the project aims to use such interpretable models for personalised medicine in order to develop customised treatment strategies.