Automation to approximate the Gestalt: applying machine learning to the general movement assessment

Abstract

Cerebral palsy (CP) is a common physical disability in children and is associated with physical, emotional, psychological and financial stressors.¹ Many individuals with CP lead long and fulfilling lives. Nonetheless, they often need long-term specialized medical care and individualized intensive rehabilitation services.² Early identification of CP informs the family and facilitates early intervention and timely referral to specialized rehabilitation services, which lead to improved long-term motor functioning due to neuroplasticity in the early years of life. Recent recommendations highlight that CP can be accurately identified before 5 months corrected age using standardized evaluations, including Prechtl’s General Movements Assessment (GMA).³ The GMA is a non-invasive tool for predicting CP involving brief assessments of infants’ spontaneous movements. The GMA was first described a half-century ago by Heinz Prechtl⁴ and there has been increasing enthusiasm for and uptake of the GMA by neonatal clinicians. In a 2013 meta-analysis of four studies with a total of 326 patients with risk factors for CP, the GMA demonstrated a pooled sensitivity of 98% and specificity of 91% for CP.⁵ During the Writhing Period (birth-6 weeks corrected age), cramped synchronized movements are a distinct abnormal movement with abrupt abdominal flexion and synchronous displacement of large proximal joints. In a 2018 meta-analysis of five studies, cramped synchronized movements were 70% sensitive and 97% specific for CP.⁶ One advantage of assessing for cramped synchronized movements is that they can usually be diagnosed or ruled out prior to discharge from the neonatal intensive care unit.

In clinical medicine, artificial intelligence (AI) and machine learning have improved the quality and delivery of care and show promise to improve them even more in the coming years.⁷ The use of AI and machine learning has already been incorporated in the interpretation and classification of medical images.⁸ AI algorithms have made major strides in diagnosing cancers in radiology, pathology, and gastroenterology by identifying aspects of images that deviate from the norm. Beyond image classification, AI can learn from a variety of inputs including medical signal data to predict outcomes. For example, machine learning applied to electroencephalogram signals from adults with brain injuries unresponsive to spoken commands allowed the detection of brain activation in response to commands, which is a predictor of eventual recovery.⁹ The ongoing integration of AI into clinical medicine will not only optimize prognostic accuracy, but will also allow less experienced clinicians or institutions to take advantage of examinations that require a high degree of experience and training, such as the GMA.¹⁰ As it stands, acquiring specific high-quality training is a prerequisite for the certified GMA assessor alongside regular practices.¹¹ Although published studies of the GMA and later CP report high interrater reliability and accuracy,⁶ individual assessor accuracy in the real world surely varies and may wane over time without deliberate and time-consuming recalibration exercises. As such, AI may allow the GMA to maintain its high accuracy as this assessment is scaled up internationally.