Context
Prematurity, defined as birth before 37 weeks of gestation, affects 6% of live births in France and 10% worldwide and represents a major risk factor for neurodevelopmental complications, such as cerebral palsy, cognitive delay, language disorders, and behavioral disorders (1). Standard clinical and radiological neurological evaluations of preterm children during the neonatal period to assess long-term developmental outcomes lack sensibility and specificity (2). This prognosis uncertainty excludes children from early intervention and targeted follow-up. Also, it induces parental anxiety, which impacts negatively on parental attachment and child development (3). Neonatal evaluation of spontaneous motor activity – also called general movement assessment – can predict the absence of occurrence of cerebral palsy with a sensibility of 80 to 100% (4) and later cognitive development (5). However, this evaluation cannot be performed in clinical practice because it requires a prolonged video recording followed by a visual analysis after manually selecting sequences of interest (5).
PhD thesis subject
The objective is to validate an early and innovative prognosis tool for the development of premature children, based on the automation of the analysis of spontaneous motor activity. To this aim, we have already managed the best world database, which includes 129 premature children followed up to the age of two years and contains 282 annotated videos for the analysis of spontaneous motor activity with their associated 3D poses (6).
Videos can be considered as being weakly annotated by doctors; in this case the movements of a video are considered or not as complex, fluid and variable. From this information, short extracts of videos, or « video patterns », and the associated 3D poses, we will learn a representation based on dictionaries of sequences characteristic of each class (normal or abnormal movements, i.e. not enough complex, not enough variable or not enough fluid), by combining physical models of movements with methods of analysis and recognition of actions allowing to decompose them (7) and to learn classes not previously observed (8), to take into account movements not observed in the database produced. To this end, we could work on the encoding and representation of the temporal sequences obtained, using graph auto-encoders (10), in order to define prototypes associated with normal movements and abnormal movements linked to a lack of complexity, variability, or fluidity. A distance from the prototypes would then provide a new quantification tool. Ultimately, we will recognize sequences of continuous movements, as for sign language (9), to develop a neurodevelopmental prognosis tool. To this end, we will use the Bayley Scale of Infant and Toddler Development – 4th Edition at 2 years of age to associate the neurodevelopment and neonatal spontaneous motor activity. This work was approved by the Ethical Committee (IDRCB 2020-A03335-34; IRBN012023/CHUSTE). Written parental consent was obtained from each participant.
1. Pierrat V, Marchand-Martin L, Marret S, Arnaud C, Benhammou V, Cambonie G et al. Neurodevelopmental outcomes at age 5 among children born preterm: EPIPAGE-2 cohort study BMJ 2021; 373 :n741.
2. Hintz S, et al. Preterm Neuroimaging and School-Age Cognitive Outcomes. Pediatrics 2018, 142.
3. Huhtala M., Korja R., Lehtonen L., Haataja L., Lapinleimu H., Rautava P. Parental Psychological Well-Being and Behavioral Outcome of Very Low Birth Weight Infants at 3 Years. Pediatrics. 2012, 129: e937–e944.
4. Kadam AS, et al. General Movement Assessment in Babies Born Preterm: Motor Optimality Score-Revised (MOS-R), Trajectory, and Neurodevelopmental Outcomes at 1 Year. J Pediatr X. 2022, 17;8:100084.
5. Hadders M. Neural substrate and clinical significance of general movements: an update. Developmental medicine and child neurology. 2018, 60(1):39-46.
6. Soualmi A, Ducottet C., Patural H., Giraud A., Alata O. A 3D pose estimation framework for preterm infants hospitalised in the Neonatal Unit. Multimedia Tools & Applications. 2023, 83: 24383-24400.
7. Erdal Aksoy E, Orhan A, Woergoetter F. Semantic Decomposition and Recognition of Long and Complex Manipulation Action Sequences https://arxiv.org/abs/1610.05693 (IJCV preprint).
8. Gowda S-N, Moltisanti D, Sevilla-Lara L. Continuous Learning Improves Zero-Shot Action Recognition. ACCV 2024, https://arxiv.org/abs/2410.10497.
9. Aloysius N, Geetha M, Nedungadi P. Continuous Sign Language Recognition with Adapted Conformer via Unsupervised Pretraining. Arxiv. 2024, https://arxiv.org/abs/2405.12018. 10. Tan M, Yang C, Li P. Graph Auto-Encoder Via Neighborhood Wasserstein Reconstruction https://arxiv.org/abs/2202.09025 (ICLR 2022).
Work environment
This Ph.D. will take place at U1059 INSERM – SAINBIOSE laboratory, Université Jean Monnet, Saint-Étienne, France. It will be under the supervision of Dr Antoine Giraud, Associate Professor of Pediatrics at the U1059 INSERM – SAINBIOSE laboratory and the Neonatal Intensive Care Unit of the Saint Etienne University Hospital, expert of infants’ spontaneous movement assessment; Pr Olivier Alata, Professor at the Hubert Curien Laboratory, UMR CNRS 5516 (https://laboratoirehubertcurien.univ-st-etienne.fr), expert of computer vision; Dr Radia Spiga, expert in public health and analytical epidemiology.
Candidate
We are looking for a motivated student holding a Master degree or equivalent (on the 1st of October 2025) in the field of computer science, with a good background in applied mathematics (probability, data analysis, estimation, and optimization, …) and software development (algorithmic, proficiency in Python, C/C++ or Matlab/Octave/Scilab …). Knowledge in computer vision and machine/Deep learning and a particular motivation towards medical applications would be appreciated.
Salary
Net salary: 1650€ per month for 3 years. Additional paid teaching activities can be envisaged on demand.
Application process
The application should include the following documents:
– Letter of intent
– Grades and ranking during Master 1 and Master 2
– Scientific CV
– List of publications (if applicable)
– Names of Referees (at least 2)
