Machine learning enables update to pediatric neurorehabilitation

摘要

Emerging new technologies are reshaping healthcare. At the cutting edge of this transformation stands artificial intelligence, which has presented the potential to significantly enhance healthcare outcomes. As a pivotal branch of artificial intelligence, machine learning (ML) involves the development of intelligent algorithms with self-improvement through experience. In recent years, ML has been shown to be instrumental in tackling complex challenges in numerous medical domains, that is, disease diagnosis,1 medical device development,2 and biological networks.3 As we will discuss below, there are opportunities for moving the field forward by integrating pediatric neurorehabilitation with novel ML approaches. In pediatric neurorehabilitation, specialized training and interdisciplinary collaboration are used to care for children with diseases, injuries, or disorders of the nervous system, to optimize functional improvement. Conditions such as autism spectrum disorders (ASDs), cerebral palsy (CP), and intellectual disability are among those that can potentially benefit from this specialized intervention. As a spiral management process, the neurorehabilitation progress starts from a personalized treatment program and is subsequently refined and updated in response to therapy-mediated advancements.4 Thus, for better-identifying problems and understanding individual needs, the accessibility of continuous and accurate assessment is essential for the rehabilitation process.5 Efficiency and accuracy in assessments hold utmost importance in clinical practice. Notably, in pediatric neurological disorders, early and accurate screening allows for treatment during a critical window of neuroplasticity so as to improve patient outcomes.6, 7 However, some existing assessment and/or screening tools are cumbersome and time-consuming. Also, assessments based on experienced clinical professionals are not amenable to widespread use.8 Recent studies indicate that some CP and/or ASD patients have limited access to accurate screening early on, such that the early window of neuroplasticity is underutilized.6, 7 ML methods may have the potential to improve these situations. In practice, most assessment scales with high reliability and validity are typically labor-intensive. ML methods could streamline them into more concise versions with similar accuracy as evidenced by relevant research in the context of the 66-item version of the Gross Motor Function Measure9 and the Autism Diagnostic Observation Schedule (ADOS).10 Moreover, ML algorithms have the ability to develop low-cost, swift, and easily applicable assessment and/or screening tools through the analysis of clinical data. For instance, Abbas et al.6 introduced a multimodular ASD assessment involving three ML modules within a mobile application. Evaluation by a blinded clinical study demonstrates that its accuracy is similar to that of gold-standard instruments such as the ADOS and autism diagnostic interview-revised (ADI-R), and its implementation does not demand hours from certified clinicians. Groos et al.8 presented a fully automated assessment based on deep learning (an ML method) for early CP prediction in high-risk infants using a single video recorded during the spontaneous movement period from 9 to 18 weeks corrected age. External validation of the well-trained deep learning model is assessed in the paper, and the results suggest a potential avenue for using this model to provide objective early CP detection in clinical settings. Zhang et al.11 demonstrated the effectiveness of supervised ML algorithms in the classification of sagittal gait patterns in children with diplegic CP. This work illustrated that integrating ML algorithms into a three-dimensional gait analysis system has the potential to interpret gait data and automatically generate high-quality analysis results. After assessments are concluded, clinical professionals will formulate a rehabilitation program based on the assessmen