Double Séminaire:
Kaniav Kamary : Bayesian survival time prediction of French Childhood Cancer Survivors’ Study using re-parameterized mixture distribution
In the last few decades, survival rates after childhood cancer have considerably in- creased due to recent treatment technique advances in developed countries. However, even though this progress leads to a growing number of long-term survivors, treatments can damage healthy tis- sues. Consequently, childhood cancer survivors (CCS) carry a significant risk of cancer treatments related to late effects. One of the leading late effects of childhood cancer radiotherapy treatment is cardiac pathology that can lead to mortality and Valvular Heart Disorder (VHD). Early di- agnosis can prove lifesaving, so it is essential to identify early the patients who are at high risk of experiencing at least one cardiac disease to improve therapeutic and follow-up protocols. This study focuses on French Childhood Cancer Survivors Study (FCCSS) patients who had received radiotherapy and chemotherapy. The main objectives of this study are the statistical modeling of the survival time of (FCCSS) patients using mixture distribution and establishing a Bayesian inference to identify the model parameters, predict the survival time, and then determine the risk of cardiac disease. Regarding the model’s statistical inference, we have developed an adaptive Metropolis-within-Gibbs algorithm to estimate the model parameters.
Tabea Rebafka : Conformal novelty detection for metabolic networks
Our goal is the detection of novelties in a set of metabolic networks, which describe the metabolic reactions of hundreds of bacteria. This amounts to perform graph classification in a semi-supervised setting. We introduce a statistically sound approach that identifies novelties in a dataset, where a graph is considered to be a novelty when its topology is significantly different from those in the reference class. As our procedure is a conformal prediction approach, the false discovery rate is known to be controlled at a user specified nominal level without making any distributional assumptions on the data. The method can be seen as a wrapper around traditional machine learning models, so that it takes full advantage of existing graph classification methods. The performance of our procedure is assessed on several standard benchmarks and on our dataset of metabolic networks. We show that our approach efficiently controls the false discovery rate and detects more novelties than existing alternative methods.