Johann FaouziPostdoctoral researcher in machine learning applied to Parkinson's disease

Home Publications Teaching Workshops Talks Open Source Software

I’m currently a postdoctoral researcher at the Paris Brain Institute in both the ARAMIS team and the Corti-Corvol team. I’m working on machine learning for precision medicine in Parkinson’s disease and I’m supervised by Olivier Colliot and Jean-Christophe Corvol. I’m interested in machine learning with medical applications and Python programming. I love contributing to open source software when I have the time to do so.

Postdoctoral research

I just started, so I do not have much to say yet, but I’m working on machine learning for precision medicine in Parkinson’s disease, with a focus on cognitive decline and falls. I’m under the supervision of Olivier Colliot and Jean-Christophe Corvol.

More information soon.

PhD thesis

During my PhD thesis, I worked on machine learning to predict impulse control disorders (ICDs) in Parkinson’s disease (PD). I was under the supervision of Olivier Colliot and Jean-Christophe Corvol.

Parkinson’s disease is the second most common neurodegenerative disease in the world, with no cure available for the moment. The therapeutic strategy is based on the dopamine replacement therapy. Although available since the 1960s’, it is only relatively recently that behavioral disorders associated with these drugs have been described. Gathered under the term of “behavioral addiction”, they include impulse control disorders, dopamine dysregulation syndrome, and punding.

Impulse control disorders are a class of psychiatric disorders characterized by impulsivity. These disorders are common during the course of Parkinson’s disease, decrease the quality of life of subjects, and increase caregiver burden. Being able to predict which individuals are at higher risk of developing these disorders and when is of high importance.

The main objective was to build predictive models of ICDs in Parkinson’s disease, in order to improve the quality of life of the patients. We were particularly interested in investigating the predictive performance of algorithms trained using known and suggested risk factors, as well as identifying new risk factors, especially genetic factors.

As the literature on predicting ICDs in PD was almost non-existent, we studied the predictability of impulse control disorders in Parkinson’s disease. With the objective of predicting their presence or absence at the next clinical visit (binary classification task with longitudinal data), we showed that the predictive performance was acceptable. We also showed that removing the genetic variants as input of the algorithms did not significantly change the predictive performance. As importantly, we detailed our rigorous methodology (nested cross-validation, replication in an independent cohort) that is missing in the few studies currently published on this topic. An article summing up this work is currently under review.

Since the genetic variants suggested in the literature did not seem to be important contributors to the models, we studied the genetics of impulse control disorders using genetic risk scores. A genetic risk score is a number indicating the risk of developing a given phenotype based on genetic data and can be obtained from a genome-wide association study. Because there is no genetic risk score for impulse control disorders to date and the sample sizes in the available datasets were relatively small, we computed genetic risk scores for other phenotypes and studied their associations with impulse control disorders. We could not report any association after correction for multiple comparisons. An article summing up this work is currently under review.

Finally, in a more methodological study, we investigated the combination of static and dynamic data in recurrent neural networks. We performed a literature review, identified four approaches and proposed a new one. We evaluated the five approaches in the use case of predicting impulse control disorders in Parkinson’s disease. We used clinical data as dynamic data and genetic data as static data. All five approaches yielded very similar results. These negative results were not really surprising due to the lack of positive results with these genetic data in the other two studies.

Open source software

I strive for open science and open source software in particular. I find creating and contributing to open source software to be a great learning experience.

More information on my contributions in the Open Source Software section.