Improving Adherence To Neonatal Resuscitation Using Machine Learning at Quality Improvement Approach(MALA)

High quality neonatal resuscitation is a key to save newborn lives, prevent brain injury and optimize child development, yet the quality of care remains far below standards. In this project, we investigate the use of video filming of neonatal resuscitation (source of data) to develop a machine learning application which automatically detects neonatal resuscitation activity. This artificial intelligence system will assist to standardize resuscitation in neonates requiring assisted ventilation on the resuscitation table. The MAchine Learning Application (MALA) installed in a tablet mounted on the resuscitation table detects the baby’s crying (sound), breathing (chest movement) and health worker’s resuscitation action (stimulation, suctioning and bag-and-mask ventilation), and provides real-time feedback (reminder) on steps of resuscitation.The real-time feedback will be in the form of audio and visual signals from the tablet during resuscitation. Following the completion of resuscitation, MALA provides a summary feedback on the resuscitation steps followed as per the resuscitation guideline.

Globally, every year among the 140 million neonates, 10-15 million do not cry or breathe at birth. These babies require resuscitation to transit from intra-uterine to the extra-uterine environment. Neonates who do not receive timely and adequate resuscitation either die or develop brain injuries and long term disability. Every year, it is estimated that one million neonates die due to intrapartum-related complication, also known as “birth asphyxia”. Two million neonates have hypoxic ischemic encephalopathy and 1.2 million have developmental delays. Most of the deaths take place in low and middle income countries. Of the non-crying neonates, two third of them die before 28 days of birth due to poor quality of care. Improving quality of neonatal resuscitation care remains a global health challenge for further reducing the burden of neonatal deaths and neuro-development delays.

Significance and scientific novelty

Machine learning approach- Machine learning approach is a well established artificial intelligence technology in different science frontiers. It is a computational model or deep learning model which uses a multiple processing algorithm in visual objection detection and many other domains. The deep learning approach has made breakthrough in processing image and video. Deep learning discovers intricate structure in large data sets by using back propagation algorithm to indicate how a machine should change its internal parameters that are used to compute the representation in each layer from that of the previous layer.Machine learning frameworks have been applied to non-emergency medical conditions, to assist with clinical decision making in screening diabetic retinopathy, retinal disease  and autoimmune disease. Machine learning technologies have also been used to support clinical decision-making in an acute medical condition such as using medical dispatchers for identifying out-of-hospital cardiac arrest on the emergency phone calls.

For the first time in the field of medicine, video objection detection using machine learning approach (MALA) will be used in for improving neonatal resuscitation care. MALA will detect crying (sound) and breathing (chest movement) of neonate, resuscitation action (stimulation, suctioning, bag-and-mask ventilation) and quality of ventilation care. The key to a good clinical outcome depends upon the quality of neonatal resuscitation care. Despite different quality improvement interventions such as reminder and feedback system to improve quality of neonatal resuscitation, real-time measurement of improved resuscitation care has not feasible. The improvement in resuscitation care is either based on change in APGAR score or heart rate immediately after birth in almost all the clinical settings. Both these indicators (apgar score or heart rate) are output of resuscitation and also depend upon a number of physiological parameters such as cardio-vascular and neurological status. Recognizing the challenge in real-time objective measurement of neonatal resuscitation and ethical concern of independent human observation, the machine learning application using activity recognition will change the process of care. MALA will be a real-time resuscitation guide for the improved resuscitation care on the resuscitation table and will provide summary feedback. MALA is like a real-time geo-positioning tracking device (GPS) during driving; it will guide health workers on resuscitation care and provide feedback to understand what improvement is required as per the neonatal resuscitation guideline. MALA will not video record the events in the resuscitation table, rather detect the sound and movement, based on which using the AI system will generate output.

Timeline of implementation of overall project- An estimated 2.5-3 years will be required to develop and pilot such machine learning application. Study I will be done during the 2nd-3rd quarter 2022. Study II will be conducted in 4th quarter 2022 and 1st– 2nd quarter of 2023. Study III will be conducted in 3rd– 4th quarter 2022 and first-second quarter of 2023. The analysis and report writing will be done during the 3rd-4th quarter 2023.