Nowadays, an increasing number of machine learning methods are being used to develop medical applications. Many implementations use state-of-art methods, such as deep learning for tasks of classification, clustering and regression. But also new concerns arise: privacy preserving, model interpretability and fairness can be as important as accuracy metrics when evaluating a model. To have a set of tools and services available to developers may enable them to develop new trustworthiness-aware applications more easily.
This integrates different third-party libraries, platforms, practices and tools in a solution toolset able to process large volumes of data, to provide reports about the results of algorithms regarding trustworthiness and to provide higher level of software abstractions (workflows in Lemonade). The toolset includes tools to big data processing (Apache Spark framework), data anonymization (privacy preserving in ARX library), fairness (Aequitas framework) and interpretability (LIME and SHAP libraries) reports and deep learning processing (Keras and Tensorflow libraries in GPGPUs). The toolset is available in a higher level abstraction as Lemonade workflows or directly as development libraries.
Lemonade Framework was developed in the context of EUBra-BIGSEA project. Now, we are integrating features related to trustworthiness and deep learning in the framework. We have plans to allow developers to define neural networks directly in Lemonade, using Keras as abstraction. Also, different views of fairness (e.g. if a machine learning algorithm is “racist”) and interpretability (e.g. display which parts of images are relevant for an algorithm in order to a physician evaluate a diagnosis) are going to be implemented as reports in the interface.