Automated detection of abnormal general movements from pressure and positional information in hospitalized infants.

Background: Prechtl's general movements assessment (GMA) allows visual recognition of movement patterns that, when abnormal (cramped synchronized, or CS), have very high sensitivity in predicting later neuromotor disorders; however, training requirements and subjective perceptions from some clinicians may hinder universal adoption of the GMA in the newborn period.

Methods: To address this, we used a three-phased approach to design a preliminary and clinically-oriented approach to automated CS GMA detection. 335 hospitalized infants were dually recorded on video and a pressure-sensor mat that collected time, spatial, and pressure data. Video recordings were scored by advanced GMA readers. We then conducted a series of unsupervised machine learning and supervised classification modeling with features extracted from clinician- and mat-driven datasets. Finally, the resulting algorithm was converted to a software interface.

Results: A classification model combining normalization, clustering, and decision tree modeling resulted in the highest sensitivity for CS movements (100%). Results were delivered via the software interface within 20 min of data recording.

Conclusion: The combination of clinical research, machine learning, and repurposing of existing sensor mat technology produced a feasible preliminary approach to automatically detect abnormal GMA in infants while still in the NICU. Further refinements of software and algorithms are needed.

Impact Statement: Machine learning can differentiate cramped synchronized general movement patterns in the neonatal intensive care unit with good sensitivity and specificity. Increasing access to the GMA through automated detection methods may allow for earlier identification of a greater number of children at high risk for movement delay. Large studies leveraging new artificial intelligence approaches could increase the impact of such detection.