Premium
Automated pose estimation captures key aspects of General Movements at eight to 17 weeks from conventional videos
Author(s) -
Marchi Viviana,
Hakala Anna,
Knight Andrew,
D'Acunto Federica,
Scattoni Maria Luisa,
Guzzetta Andrea,
Vanhatalo Sampsa
Publication year - 2019
Publication title -
acta paediatrica
Language(s) - English
Resource type - Journals
SCImago Journal Rank - 0.772
H-Index - 115
eISSN - 1651-2227
pISSN - 0803-5253
DOI - 10.1111/apa.14781
Subject(s) - movement assessment , pose , artificial intelligence , skeleton (computer programming) , kinematics , medicine , movement (music) , estimation , cerebral palsy , computer science , computer vision , physical medicine and rehabilitation , pattern recognition (psychology) , motor skill , philosophy , physics , management , classical mechanics , psychiatry , economics , anatomy , aesthetics
Aim General movement assessment requires substantial expertise for accurate visual interpretation. Our aim was to evaluate an automated pose estimation method, using conventional video records, to see if it could capture infant movements using objective biomarkers. Methods We selected archived videos from 21 infants aged eight to 17 weeks who had taken part in studies at the IRCCS Fondazione Stella Maris (Italy), from 2011 to 2017. Of these, 14 presented with typical low‐risk movements, while seven presented with atypical movements and were later diagnosed with cerebral palsy. Skeleton videos were produced using a computational pose estimation model adapted for infants and these were blindly assessed to see whether they contained the information needed for classification by human experts. Movements of skeletal key points were analysed using kinematic metrics to provide a biomarker to distinguish between groups. Results The visual assessments of the skeleton videos were very accurate, with Cohen's K of 0.90 when compared with the classification of conventional videos. Quantitative analysis showed that arm movements were more variable in infants with typical movements. Conclusion It was possible to extract automated estimation of movement patterns from conventional video records and convert them to skeleton footage. This could allow quantitative analysis of existing footage.