Automated early prediction of cerebral palsy: interpretable pose-based assessment for the identification of abnormal infant movements

McCay, Kevin D. (2022) Automated early prediction of cerebral palsy: interpretable pose-based assessment for the identification of abnormal infant movements. Doctoral thesis, Northumbria University.

Preview

Text (Doctoral Thesis) mccay.kevin_phd_11028440.pdf - Submitted Version Download (11MB) | Preview

Abstract

Cerebral Palsy (CP) is currently the most common chronic motor disability occurring in infants, affecting an estimated 1 in every 400 babies born in the UK each year. Techniques which can lead to an early diagnosis of CP have therefore been an active area of research, with some very promising results using tools such as the General Movements Assessment (GMA). By using video recordings of infant motor activity, assessors are able to classify an infant’s neurodevelopmental status based upon specific characteristics of the observed infant movement. However, these assessments are heavily dependent upon the availability of highly skilled assessors. As such, we explore the feasibility of the automated prediction of CP using machine learning techniques to analyse infant motion.

We examine the viability of several new pose-based features for the analysis and classification of infant body movement from video footage. We extensively evaluate the effectiveness of the extracted features using several proposed classification frameworks, and also reimplement the leading methods from the literature for direct comparison using shared datasets to establish a new state-of-the-art. We introduce the RVI-38 video dataset, which we use to further inform the design, and establish the robustness of our proposed complementary pose-based motion features. Finally, given the importance of explainable AI for clinical applications, we propose a new classification framework which also incorporates a visualisation module to further aid with interpretability. Our proposed pose-based framework segments extracted features to detect movement abnormalities spatiotemporally, allowing us to identify and highlight body-parts exhibiting abnormal movement characteristics, subsequently providing intuitive feedback to clinicians.

We suggest that our novel pose-based methods offer significant benefits over other approaches in both the analysis of infant motion and explainability of the associated data. Our engineered features, which are directly mapped to the assessment criteria in the clinical guidelines, demonstrate state-of-the-art performance across multiple datasets; and our feature extraction methods and associated visualisations significantly improve upon model interpretability.

Actions (login required)

View Item

Downloads