back to list

Project: Plausible Counterfactual Scenario Generation for Explainable AI in Psychosis Prognosis Prediction


Introduction: Artificial intelligence (AI) has shown great promise in different domains including the clinical domain. However, the applications of the developed AI model in clinical practices remained limited mainly due to the lack of model explainability. Clinicians, in general, want to know why an AI model came up with certain decisions or predictions. While the most effective AI models, such as deep learning models, are considered backbox models, it is very difficult to interpret or explain their predictions. Nevertheless, counterfactual inference provides a means for opening this black box. In counterfactual inference, the predictions of the model are calculated for counterfactual samples (e.g., by manipulating a certain feature in an actual sample), and the difference between model predictions for the actual and counterfactual samples is used to explain the effect of each feature on the final prediction. While simple and effective, the application of counterfactual inference in clinical domains is challenging because generating clinically plausible counterfactual scenarios for tabular data (that are common in the clinical domain) is not straightforward.

Aim: Within the U-HEAL project (an EWUU-funded project, see for more details), we aim at finding new solutions, e.g., using variational autoencoders (VAE)  or generative adversarial networks (GAN), to generate plausible counterfactual scenarios for specific applications in psychosis prognosis prediction.

Objectives:  The student is expected to meet the following three objectives:

  1. Literature review on tabular data generation using deep learning

  2. Employing an existing tabular data generation method for counterfactual scenario generation in psychosis prognosis prediction 

  3. Developing new methods to impose plausibility constraints in the data generation process

Requirements: The student should be familiar with basic concepts in machine learning and deep learning. Experience with deep generative models such as VAE and GAN is a plus. The student should also be familiar with Python packages for deep learning such as Tensorflow, Keras, or PyTorch.

Funding opportunities:  In case the master thesis ends up in a conference or journal publication (this is optional), all the costs (conference registration and travel) will be covered by the U-HEAL project. 

For more information about this project, please contact Dr. Maryam Tavakol ( or Dr. Seyed Mostafa Kia (

Maryam Tavakol
Secondary supervisor
Dr. Seyed Mostafa Kia