Automated analysis of facial dynamics in newborns: a path to earlier diagnosis of Rett syndrome

Background: Rett syndrome (RTT) is a rare neurological disorder primarily affecting females and is caused by de novo mutations in the gene encoding the epigenetic regulator methyl-CpG-binding protein 2 (MECP2). Affected individuals typically experience severe cognitive, social, motor, and physiological impairments emerging during toddlerhood [1]. Because early developmental milestones often appear normal, diagnosis is commonly delayed until around two years of age, after critical periods of brain development have already passed, and interventions are likely to be less effective. Evidence from retrospective home videos suggests that atypical facial and body dynamics are already present in RTT patients during postnatal life [2], however, there is a significant shortage of clinicians trained to detect these subtle early signs. We recently demonstrated that automated tools for behavioral analysis can discern between mouse pups carrying disease-causing MeCP2 mutations and wild-type littermates based on their atypical patterns of spontaneous movements [3], providing proof that such tools may be effectively applied to anticipate diagnosis in the context of RTT. Objectives: In this project we aim at translating the methodology developed in our previous preclinical study to the clinical setting and develop a computer vision-based tool able to automatically recognize newborns affected by RTT. A focus on facial dynamics will ensure maximal exploitation of real-world videos recorded by caregivers. Design: We are collecting retrospective amateur videos of newborns aged 0 to 6 months who were later diagnosed with RTT, along with videos of their typically developing (TD) siblings or unrelated control infants. A system previously developed for the automated assessment of facial muscle movements and expressive behavior in infants, based on the Facial Action Coding System for Infants and Young Children (Baby FACS), will be used to extract facial action features from the video data. These features will serve as input for training machine learning algorithms in a binary classification task aimed at distinguishing RTT from TD infants. The performance of each proposed classifier will be evaluated on an independent set of video recordings, with a focus on both prediction accuracy and computational efficiency. The model that achieves the best balance between these two criteria will be selected for further analysis. Explainability techniques will then be applied to this model to gain insight into the facial expression anomalies that may characterize the presymptomatic stages of RTT. Expected results: We plan to develop a framework capable of identifying newborns with suspected RTT before the onset of overt symptoms, by analyzing atypical facial dynamics captured in amateur videos. This tool could eventually be translated into a user-friendly application for the automated detection of suspected medical conditions using smartphone recordings. Such a system would support earlier referral to specialized centers, thereby accelerating diagnosis. With the recent expansion of therapeutic options, marked by the FDA approval of Trofinetide and the ongoing advancement of gene therapies in clinical trials, enabling timely intervention could significantly improve clinical outcomes and increase the likelihood of meaningful recovery for affected individuals. Funded by the European Union -Next Generation EU-NRRP M6C2-Investment 2.1 Enhancement and strengthening of biomedical research in the NHS (code PNRR-MR1-2022-12376808). [1] Hagberg B., 2002. Clinical manifestations and stages of Rett syndrome. Mental Retardation and Developmental Disabilities Research Reviews 8(2), 61-5. [2] Cosentino L., Vigli D., Franchi F., Laviola G., De Filippis B., 2019. Rett syndrome before regression: A time window of overlooked opportunities for diagnosis and intervention. Neuroscience and Biobehavioral Reviews 107, 115-135. [3] Cosentino, L., Urbinati C., De Filippis B., 2024. Automated recognition of mouse pups modelling Rett Syndrome before symptom onset: exploiting spontaneous movements for diagnosis acceleration. Neuroscience Applied 3, 104337. Topics: Neurodevelopmental conditions, artificial intelligence and computational models