Automatic oxygen controller design for premature infants with clinical evaluation

[ACCESS RESTRICTED TO THE UNIVERSITY OF MISSOURI--COLUMBIA AT REQUEST OF AUTHOR.] For premature infants, the peripheral oxygen saturation (SpO2) level has significant effects on their health. Manual adjustment of fraction of inspired oxygen (FiO2) to keep SpO2 within the prescribed range is the common approach in neonatal intensive care unit (NICU). However, manual control of FiO2 is a heavy workload task for nurses, and due to the variability of SpO2, it is difficult to maintain it within the target range. Motivated by these limitations, the research of automatic control of FiO2 is the topic of interest. In this study, a new automatic oxygen controller was designed, then a pilot clinical study was conducted to evaluate the control system's performance. The new automatic oxygen controller was designed based on the previous controller design concept, model estimation plus adaptive control. Firstly, the disturbance observer was modified, the expressions of the matrices in the extended state observer were derived, especially the observer gain vector. Based on the estimation system, two adaptive proportional-integral (PI) controllers were designed. The stable region of controller gains was calculated according to the stability criterion based on a simplified system and used as the guidance of the first-generation adaptive controller design. The second-generation adaptive controller was designed based on simulations of actual system, the centroids of the stability and performance regions were the target of setting controller gains. In order to reduce the influence of time delay caused by gas transport to the control system, Smith predictor and derivative feedback were compared through simulations. Derivative feedback was selected to be added into the control system. Additionally, the integral windup problem was observed through the first clinical test on human subject. An anti-windup mechanism was designed that can effectively handle the two main causes of the integral windup issue, one is manual operation switch to automatic control, the other is infants having difficulties receiving oxygen. Finally, non-clinical tests were conducted to evaluate the performance of the new control algorithm design before applied to the real clinical tests. A pilot clinical study was conducted to compare the blood oxygen saturation targeting performance of the automatic oxygen control system with the standard of care, manual control of inspired oxygen, for preterm and low birthweight infants. A crossover study was designed with several endpoints including the comparison of the percentage of time that the SpO2 is in a target range for automatic oxygen control and manual oxygen control. Other metrics were also compared to assess performance including numbers of bradycardia events. The pilot study included six patients that fit the inclusion criteria. The results showed that there were improvements in the measured outcomes considered including statistically significant improvements in the number of bradycardia events for the period where automatic oxygen control was used. In order to notify healthcare providers of the upcoming alarm events, and potentially increase patients' health condition by providing extra time for healthcare providers to take action, a NICU alarm events forecasting model was built using long short-term memory recurrent neural network model. This forecasting model was trained on the clinical data collected from the pilot study. The overall accuracy of forecasting SpO2 alarm status could reach 75 [percent], in which the true positive rate for desaturation alarm is 77 [percent], and for oversaturation alarm is 84 [percent]. The current alarm event forecasting research is a proof of concept, the preliminary results demonstrated the potential of using deep learning methods. With further study, the alarm events forecasting model could be integreted into the automatic oxygen control system for better control performance