PHOTONAI

The machine learning framework photonai is motivated by the rising need to apply multivariate pattern recognition methods to scientific problems without in-depth technical training.

In order to apply machine learning in practice, researchers need to select suitable algorithms from numerous different toolboxes, learn toolbox-specific syntaxes, decide on a validation scheme, manage the data flow and integrate all these components within a specifically implemented training and testing procedure. This highly repetitive process requires technical knowledge and programming experience, while also being time-consuming and error-prone. Interestingly, however, the basic workflow to construct, optimize and evaluate machine learning models follows a static scheme. It can be framed as the (systematic) search for the best combination of data processing steps, learning algorithms, and hyperparameter values under the premise of unbiased performance estimation. Therefore, it perfectly qualifies for automation within a software layer of abstraction.

Research Project

We develop photonai as a software covering the full predictive analytics workflow - i.e. model development, optimization, evaluation, and deployment, motivated by the lack of suitable software for medical machine learning analyses.

Built on top of existing state-of-the-art machine learning libraries, it automates the supervised training, hyperparameter optimization and testing loop with nested cross-validated and ensures valid performance estimates. Thereby, it not only condenses machine learning analysis to a set of important design decisions, but also mitigates the problem of reporting inflated performance values and models with a high risk of bias, as predominantly present in current medical machine learning publications (Navarro et al., 2021).

Most importantly, it is easy to customize and allows the construction of complex multi-part pipelines for specialized medical use cases, such as stacking classifiers for each region of interest (ROI) in the brain, trained distinctively. Moreover, it offers convenient constructs for typical machine learning development use cases, such as choosing the right learning algorithm for the learning problem. On top of that, we have a responsive result visualization platform and several plugins for handling advanced medical data modalities (neuroimaging, graphs, time series).