In collaboration with Joan LaRovere and Sebastien Boyer

When discharged from hospital after an inpatient stay there is an 8 to 18% chance of readmission to the hospital for the same condition within 30-days. This is both clinically bad for the patient and expensive to the medical insurer. Readmission within 30 days of discharge represents a $41.3 billion dollar market and accounts for 11% of total hospital costs. Clinicians strive to avoid readmission for their patients and do their best to be certain that they are clinically ready for discharge but obviously this system is imperfect. Building a predictive model of 30-day readmissions to assist hospitals and clinicians could aid this process.

In this project we implement a topic modeling algorithm that we combine with several supervized learning algorithms to provide a predictor with an AUC of 0.7. We also perform a clustering analysis that provides an easy-to-interpret way to understand our results.

Research report available here.

In collaboration with Sebastien Geerart, Vincent Bogart, Cyril Colin, and Antoine Caillaud

Automated electric defibrillator are a key part of first aid in the case of cardiac arrest. The defibrillator measures the electric activity of the heart, and applies a shock if it detects fibrillation. There is scientific evidence that the shock may be harmful to the patient. In particular, when the patient has been in cardiac arrest for a long time, it may be better to first perform CPR for a few minutes.

In this project, we try to predict whether an electric shock will be succesfull in restoring normal cardiac activity. We used recordings of a 500Hz pre-shock electrocardiogram signal, to compute a series of features, including FFT and AMSA. We then trained models on these features to predict whether an electric shock was going to be efficient. Efficiency here was defined by the absence of cardiac fibrillation for at least 60s. This was labeled by a physician. We managed to obtain reasonable accuracies in terms of prediction (order of 90%) but this work would need to be improved in order to actually be applied in defibrillator algorithms.

Research report available here (in French).

In collaboration with Sebastien Boyer

Amateur sports clubs often participate in local and regional championships. We assume that the first phase of the championship consists of a round robin where every team plays against all other teams. The winner advances to the second phase. This round robin phase requires all participants to travel. When this travel is frequent and over long distances, this not only places a burden on families, but also generates significant CO₂ emissions. In this project, we wish to find a way to group clubs that minimizes the sum of all travel distances.

We model this problem as a graph, where vertices represent clubs, and edges represent the travel distance between two clubs. We provide heuristic algorithms that use Semi Definite Programming, as well as Branch and Bound techniques to find a partition that minimizes the sum of edges between all the vertices of the same partition.

Research report available here (in French).