The current study demonstrated that the complexity of term-equivalent writhing movements in preterm infants quantitatively analyzed using a deep learning algorithm is associated with early neurological development. The complexity indices of both joint angles and joint angular velocities were different between the very preterm infants with HINE<60 and ≥60, and showed positive correlations with the HINE scores in most joints of the upper and lower extremities. The similarity indices among each joint angle or among each joint angular velocity did not differ between the infants with HINE<60 and ≥60 in most of the joints at the upper and lower extremities.
In this study, reduced complexity of writhing movements was significant or demonstrated a trend towards significance for all of the upper and lower extremities in the preterm infants with HINE<60 compared to those with HINE≥60. The complexity indices of both joint angles and joint angular velocities showed a significant positive linear relationship with the global HINE scores even though the strength of relationships was weak. These results are overall consistent with those of previous reports that describe decreased complexity as one of the main findings in infants at a high risk of cerebral palsy20,21. Hypothetically, spontaneous movements are endogenously generated by the central pattern generator network in the spinal cord and brainstem4. Its activity may result in simple body movements but can be modulated by the activity of the supraspinal structures (e.g. subplate and cortical plate) during brain development; this may induce movement complexity22. The deviant patterns of the spontaneous movements are characterized by a lack of variability; this can result from injuries or dysfunction of the subplate or cortical plate and/or its connective fibres21.
Complexity, which comprises spatial and temporal variability, was investigated to objectively measure spontaneous movements in preterm infants since reduced complexity is strongly associated with cerebral palsy4. SE was adopted as a key index to measure complexity quantitatively. Generally, reduced complexity indicates fetal or neonatal compromise and has been regarded as one of the important characteristics to distinguish normal and abnormal spontaneous movements during the writhing period18. A previous study demonstrated that SE of time-series kinematic data acquired from infants with wearable sensors attached to the lower extremities is associated with a risk of developmental delay23. In that study, SE of the lower-limb movements was significantly decreased in the infants at risk of developmental delay, which is consistent with the results of the current study. Conversely, similarity in terms of the Pearson correlation coefficients did not correspond with the results of complexity. It may be assumed that the similarity of the inter- or intra-limb movements reflects the characteristics of cramped-synchronized movements, but the similarity index, in terms of Pearson’s correlation coefficients, may not be sufficient to exclude normal writhing movements. It is necessary to plan future studies that reveal the quantitative features of each type of spontaneous movements according to GMA.
Automatic computer-based analyses of spontaneous infantile movements have been explored for clinical use in numerous previous studies24,9. One of the main approaches applied is a two-dimensional video-based approach. This has been strengthened by the rapid progress of computer vision technologies, such as deep learning algorithms, and the popularization of smartphone cameras25. In the current study, AlphaPose, an open-source pose-estimation model that demonstrates state-of-the-art performance, was used to extract the positional coordinates of each joint at the upper and lower extremities16. The acquisition of positional coordinates was conducted automatically using AlphaPose, facilitating the efficient performance of kinematic analyses of spontaneous infantile movements. The pose-estimation model provides automatic acquisition of spatiotemporal data on the trunk, arms, and legs of infants without labour-intensive and time-consuming manual marking. This study determined the kinematic parameters as joint angles and joint angular velocities, rather than as the raw data of the positional coordinates. Joint angles were considered beneficial since they can be less affected by the camera’s height, tilt/roll angles, and directions compared to positional coordinates and are straightforward for clinicians to analyze26–28.
This study has several limitations. First, in this study, a neurological outcome in infants was the global score of HINE at 4 months of corrected age. Even though a low global score of HINE at an early stage is suggestive of cerebral palsy, future studies are necessary to identify the long-term longitudinal association between complexity indices of spontaneous infantile movements and poor neurological outcomes. Second, the sample size of this study was relatively small and a further study is warranted with a large population of preterm infants. Third, kinematic analyses were performed on the two-dimensional video images, not on the three-dimensional image analyses of spontaneous movements in preterm infants. To obtain easy access to video images of the infants after hospital discharge, this study adopted the image acquisition instruments of conventional or smartphone cameras.
In conclusion, quantitative assessments of spontaneous movements in preterm infants using a deep learning algorithm are feasible. Complexity indices in terms of SE of joint angles and joint angular velocities at both the upper and lower extremities are associated with early neurological development. This study indicates that complexity indices of writhing movements can be a potential candidate to detect cerebral palsy in high-risk infants. Further studies are warranted with larger cohorts of preterm infants and a focus on developing automatic computer-based models using complexity indices for predicting cerebral palsy in clinical practice.