Research¶
Research is one part of the engineering record. The work below preserves the scientific questions, data constraints, and implementation contributions without turning the site into an academic CV.
Epileptic Seizure Detection From Video and Audio¶
The doctoral work investigated infantile spasm detection from ordinary video and limited audio. It produced separate action-recognition and classical optical-flow systems, supporting child detection and state clustering, and the infrastructure required to process long clinical archives.
Recorded results include 85% ROC AUC for X3D-M and 0.96 ROC AUC for DualTVL1 optical flow with a Gaussian Mixture Model on infantile spasm detection. The collection program obtained CNIL authorization and gathered data from more than 100 patients across six epilepsy centers.
Multimodal Cartography of Human Lymphopoiesis¶
High-dimensional cytometry was being analyzed manually, limiting the population structure that researchers could resolve. Automated pipelines combined multimodal data, UMAP, and PCA to map human lymphoid progenitor populations and support figure generation and interpretation with biology researchers.
The recorded workflow reduced analyst time by approximately 60% and contributed to two co-authored publications in Cell Reports and iScience.
Neonatal Thermography¶
This work explored newborn thermogenesis where no established neonatal thermal-video method was available. Flood fill segmented thermal zones; barycenters summarized them; heat-distribution graphs and a theoretical diffusion model described the first ten minutes after birth.
The pipeline ran in real time. The study recorded a correlation of 1 between thermal image and Apgar score within the 36–37°C range and resulted in a 2025 Frontiers in Pediatrics publication.
Explainability for Clinical Computer Vision¶
Clinical collaborators needed a way to inspect model decisions. SHAP was proposed and implemented for explainability, and the method contributed to the explainability section of work on spinal muscular atrophy hypotonia detection published in JAMA Pediatrics.
Clinical Data Infrastructure¶
Research data could not be separated from collection and governance.
Collection¶
An air-gapped Node.js backend, iPad form, and Raspberry Pi nodes collected recordings across five patient rooms. The deployed layer handled clinical collection; trained detection models remained research prototypes.
Secure sharing¶
A separate backend used the Signal protocol to receive child videos. The implementation was supported by a comparative security and open-source review of Signal, WhatsApp, and Telegram and was presented to the AP-HP DPO. Approval established the legal right to use the technology when hosted in Europe.
Dataset construction¶
Collection required collaboration with medical teams across multiple centers. An automated segmentation-based pipeline identified child occlusion and video quality for labeling, producing 1,928 labeled child images, recorded as the largest public child-detection dataset at the time.
Research Practice¶
Two interns were supervised with different operating rhythms: stepwise tasks for a junior profile and milestone-based ownership for a more experienced profile, supported by daily unblock meetings. The recorded outcomes were a Python package with C++ bindings and 2x speed improvement, reusable preparation pipelines reducing team preparation time by 5x, and an I3D detector with an 80% success rate.
Publications¶
The complete bibliography available in the source material is preserved on the publications page.