Research¶
My research has focused primarily on computer vision and machine learning for medical applications, from automated analysis of clinical videos to multimodal biological data and neonatal thermography.
The projects below describe the scientific questions, methods, and contributions developed in collaboration with clinical and research teams.
Infantile Epileptic Spasm Detection¶
My doctoral research investigated whether infantile epileptic spasms could be detected using ordinary video recordings.
I developed and evaluated two complementary approaches:
- Action recognition using pretrained spatiotemporal networks, including X3D-M and I3D.
- Classical computer vision using DualTVL1 optical flow and Gaussian Mixture Models (GMM).
The X3D-M approach used visual embeddings to classify short video sequences, achieving a ROC-AUC of 0.813 ± 0.058 on the evaluated dataset.
The optical-flow approach transformed motion information into temporal signals and used a GMM for classification, achieving a ROC-AUC of 0.964 on a separate clinical dataset.
These results were obtained on different patient cohorts and are not directly comparable.
The research also involved child detection, exploratory audio analysis, and the development of video-processing tools for large clinical archives.
The detection methods remained research prototypes and were not validated for clinical decision-making.
Multimodal Cartography of Human Lymphopoiesis¶
During my research at EPHE-PSL, I worked on the analysis of high-dimensional cytometry data to investigate human lymphoid progenitor populations.
I developed automated analysis pipelines combining multimodal datasets, dimensionality reduction, and visualization techniques, including PCA and UMAP.
The work supported the identification and characterization of cell populations, in collaboration with biology researchers.
The automated workflows reduced analyst processing time by approximately 60% compared with the previous manual approach.
I contributed to two publications:
- Distinct Subsets of Multi-Lymphoid Progenitors Support Ontogeny-Related Changes in Human Lymphopoiesis — Cell Reports, 2023.
- Multimodal Cartography of Human Lymphopoiesis Reveals B and T/NK/ILC Lineages Are Subjected to Differential Regulation — iScience, 2023.
Neonatal Thermography¶
This research investigated whether infrared thermography could provide useful information about newborn thermal adaptation during the first minutes after birth.
I contributed to the development of an image-processing pipeline using flood-fill segmentation, thermal-region analysis, and barycenter-based representations of heat distribution.
The work also involved modeling how thermal patterns evolved during the first ten minutes of life.
The resulting prospective study included 223 newborns and examined thermal distributions alongside Apgar scores recorded during the same period.
The published results described differences in early surface temperature patterns between newborns with lower and higher Apgar scores, with progressive warming and convergence over time.
The study established the feasibility of characterizing early thermal adaptation using infrared imaging. It did not establish a clinically validated prediction method.
Publication:
Newborn clinical condition assessment using infrared thermography: correlation with the Apgar score in a prospective cohort study — Frontiers in Pediatrics, 2025.
Explainability for Clinical Computer Vision¶
I contributed to research on explainability for computer vision models used in clinical assessment.
The objective was to make model predictions more interpretable to medical collaborators.
I proposed and implemented an approach based on SHAP (SHapley Additive exPlanations) to analyze model predictions.
This contribution was incorporated into research on automated hypotonia detection in children with spinal muscular atrophy.
Publication:
Clinical Data Infrastructure¶
Medical machine learning research required more than model development. Collecting, processing, and sharing clinical recordings introduced constraints involving privacy, infrastructure, and data quality.
I contributed to several supporting systems.
Clinical Video Collection¶
I developed and deployed a data-collection system using a Node.js backend, an iPad interface, and Raspberry Pi recording devices.
The infrastructure operated on an isolated hospital network without external internet connectivity.
It supported video collection across five patient rooms and was used to collect recordings from approximately 50 patients.
This was a deployed clinical data-collection system. The machine learning models developed during the research remained experimental.
Secure Video Sharing¶
I developed a separate backend using the Signal protocol for secure video transfer.
The work included a comparative review of Signal, WhatsApp, and Telegram, focusing on security properties and open-source implementations.
The proposed system was presented to the AP-HP Data Protection Officer as part of the review of its intended use.
Dataset Construction¶
I participated in the development of clinical video datasets through collaboration with medical teams across six epilepsy centers.
The broader collection program received CNIL authorization and included recordings from more than 100 patients.
I also developed an automated image-processing pipeline to identify child visibility, occlusions, and image quality.
This work contributed to a specialized child-detection dataset containing approximately 1,930 labeled images, used to evaluate object detection models in clinical settings.
The dataset and child-detection methods were described in:
Child detection research — IEEE Access, 2025
Publications¶
My complete bibliography, including journal articles, conference publications, and doctoral research, is available on the Publications page.