Inserm U1219 Bordeaux Population Health · AHeaD team · University of Bordeaux

Research themes

One question runs through all of it: how do you produce, from imperfect health data, a measurement reliable enough to act on.

Modelling patient flow and processes in emergency care

The core of my PhD. An emergency department is a queueing network coupled to finite resources, clinical decisions and a saturated hospital downstream. I build models there whose purpose is to let a reorganisation be tried before it is applied to real patients: where to place a nurse, how to order a queue, whether to open a bed.

The work falls into four stages. An exploratory analysis first characterises the department in its own right — who arrives and when, which pathways, which durations, how occupancy moves minute by minute — and determines what a model needs to represent. A deterministic compartmental model gives an aggregate reading of it, built by successive additions of mechanisms, each evaluated before the next is introduced. Then comes a discrete-event simulation, following patients one by one: routing is learned by gradient-boosted models specific to each area, durations are estimated from visits observed under low congestion, and delays emerge from queues, capacities and blocking rather than being supplied as parameters. Once validated, it serves as a laboratory for comparing organisational levers, each scenario replaying the same patients at the same arrival times. Out of this work came EDSim, a patient-flow modelling tool for emergency departments, which separates the simulation engine from the configuration of a given department.

Repairing degraded hospital records

No flow model survives false data. The traces a care pathway leaves behind are produced in the course of clinical work, not for research: misordered events, impossible durations, duplicates, broken trajectories. Reconstructing those pathways is therefore a piece of research in itself, not a preparatory step to be rushed through.

I built a three-stage pipeline: deterministic expert rules for the repetitive errors, a probabilistic transition graph for interpretation and for spotting structurally abnormal pathways, then an encoder-decoder Transformer that learns to produce the corrected trajectory from the corrupted one — area sequence and timestamps together. The whole is evaluated against a gold standard annotated by clinical experts. The same concern for what is feasible inside the hospital drives parallel work on record deidentification with language models compact enough to run on the hospital’s own servers.

Related publications

Algorithmic fairness, bias and serious games

A model trained on a department’s history also learns its biases. With the team, I study the detection of human bias in triage — gender, age, socioeconomic status — using language models, and the use of serious games to make those biases perceptible to clinicians in training.

The question is not only technical. Optimising patient flow means deciding who waits: an average gain can conceal a loss for one particular group. That is why the scenarios I evaluate are broken down by subgroup rather than judged on an average, and why protecting the most acute classes is treated as a constraint rather than a side effect.

Related publications

Computer vision and road safety

Automated assessment of pedestrian crossing safety across France, pairing aerial and street-level imagery to score nine risk factors. An applied project outside the hospital, sharing the same logic: producing at scale a measurement that human expertise cannot deliver across a whole country.