PhD thesis: data driven decision support systems for Alzheimer’s disease
Predicting diagnosis evolution
I focused on predicting the evolution from mild cognitive impairment to Alzheimer’s disease status. In a first project I conducted a systematic quantitative review of 172 articles. I supervised a group of 19 readers who extracted information from these articles, and conducted a statistical analysis to identify current trends and the methodological elements that most impact the performance of the prediction. I also identified methodological issues and proposed guidelines to ensure the performance evaluation reflects what can be achieved in clinical practice. In a second project, I proposed a method for predicting the diagnosis evolution while maximizing interpretability. This method is composed of two parts: (1) Prediction of the evolution of the features at the given date; (2) Prediction of the corresponding diagnosis using the predicted features. I participated in the Tadpole challenge; this proposition ranked 6th in the global leaderboard and 1st in the university leaderboard.
Pre-screening tool for clinical trials
I worked on lowering the recruitment cost for Alzheimer’s disease clinical trials targeting amyloidosis (abnormal protein build up), by proposing a pre-screening tool for selecting subjects having amyloidosis. I benchmarked a variety of algorithms: Random Forest, SVM, logistic regression, adaptive boosting, and adaptive logistic regression. I studied the impact of the input features (cognitive or imaging variables, cross-sectional or longitudinal), feature selection and data set size. Testing this method on 3 different data sets showed it could lead to a 20% cost decrease!
Longitudinal study of prescription patterns in electronic health records
In a last project I worked on GERS-Data, a French medical data base of electronic health records. I analyzed the health records of about 35,000 patients, spanning 15 years before Alzheimer’s disease diagnosis and 10 years after. I compared the prediction patterns of patents with Alzheimer’s disease with those of 2 control groups: patients suffering from a mild cognitive impairment and patients with no neurological or psychiatric disorder. I also studied the impact of the diagnosis for patients suffering from Alzheimer’s disease. This project was especially interesting for me as it highlighted the differences between research and clinical data, as well as the biases one can encounter in real world data.
Post-doc project: non-linear statistical model of the spine
Modeling the spine and its possible shapes is an important step for many applications in spine imaging. It is involved in the regularization of 3D reconstruction from 2D images and in vertebra segmentation, and could be useful for predicting the evolution of adolescent idiopathic scoliosis for example. Current statistical models of the spine rely on principal component analysis and are limited by this linear representation. I am currently working on learning a non-linear representation, with the hope that a more advanced modeling could positively impact a wide range of applications.