**1. Introduction**

The drive and control of breaths is an intricate system involving the circulatory, pulmonary, and central nervous systems. Chemoreceptors throughout the body monitor for hypoxia and hypercarbia, altering respiratory rate (RR) to maintain organ perfusion and optimal pH. Evidence has shown RR to be a better predictor of cardio-pulmonary deterioration than blood pressure and pulse rate [1]. In the neonatal population, ventilation and appropriate pulmonary support is critical. Effective ventilation is the key objective in neonatal resuscitation, and RR is included in all early warning systems for neonatal sepsis and necrotizing enterocolitis [2,3].

Traditionally, the gold standard for measuring RR is counting breaths for a minute while auscultating the patient or palpating for chest rise. While this is the most accurate, it is time consuming and not practical for an intensive care unit where continuous vital signs monitoring is required. Since the late 1800s, multiple innovations have been developed to automate measurements of respiration [4]. Early inventions measured breaths by airflow into a device, spirometry, or by attaching a circumferential strap around the chest, respiratory inductance plethysmography (RIP). Spirometry, however, is designed to measure lung volumes, and RIP is currently used in neonates to detect airway obstruction and for sleep apnea monitoring as it accurately records respiratory waveforms over time [5]. While RIP does measure RR, it requires bulky chest straps and interpretation from a pulmonologist or somnologist.

Similar to RIP, impedance pneumography (IMP) measures changes in electrical signal secondary to the movements of the chest and diaphragm and has been widely adapted to multiple clinical settings. IMP uses mathematical algorithms to convert other electronic signals such as pulse oximeters and electrocardiograms (ECGs) to a respiratory wave form and RR, while RIP is derived directly from the attached straps. IMP can be integrated into standard continuous monitoring systems used by most hospitals [6]. While any manipulation of data or signal can introduce error, studies have shown strong correlation between RR gathered from RIP and IMP [6,7].

Nevertheless, there are several limitations of IMP. Information has to be gathered from electronic leads attached directly to the skin, and the signal needs to be amplified to be measured, making it susceptible to artifacts or noise. Artifacts and signal noise can originate from inadequate attachment of the electronic leads as well as any movement of the patient not related to breathing. In the neonatal population, these limitations are particularly apparent and produce several unique adverse outcomes. The frequency of false alarms secondary to the frequent movement of newborns is associated with provider alarm fatigue, infant hearing loss, and a disruptive environment for development [8–12]. Additionally, the humid environment of neonatal incubators and the infant's thin, underdeveloped skin cause the adhesive in electric leads to fail and require frequent changing. The recurrent application of adhesive to the fragile premature skin causes breakdown and inflammation of their dermal barrier, introducing possible sources for infection [11].

Several innovations have been developed for non-contact monitoring in neonatology to minimize risk of dermal injury and alarm fatigue in infants. A variety of techniques have been studied from radio wave signaling, ultrasound, imaging photoplethysmography (PPG), and video-based respiratory monitors. In controlled environments, they have been shown to correlate with ECG monitoring with correlation coe fficients from 0.79 to 0.92 [13–16]. Video-based respiratory monitors are particularly versatile due to the accessibility of cameras. These systems could be easily integrated into current monitoring systems and potentially used in clinics and at home for virtual medical appointments. However, due to the subtle motion of neonatal respiration, these technologies have struggled to accurately determine RR as the amplification of movement also increases signal noise, making it di fficult to obtain an accurate measurement.

We conducted a course of study focused on a video-based respiratory monitoring system that extracts a respiratory waveform and rate without augmenting the infant's surrounding or attire. Unlike past studies, this system extracts data on fully clothed or swaddled infants in a neonatal intensive care unit (NICU) with a variety of lighting and camera orientations. Through an iterative approach, we demonstrate a proprietary technique that is able to compensate for background signal noise while amplifying and measuring RR.

### **2. Materials and Methods**

This study was performed at a single center, Lucile Packard Children's Hospital at Stanford University. Patients in the neonatal intensive care unit (NICU) and step-down intermediate care nursery were enrolled in this study. Institutional Review Board (IRB) approval was obtained through the Stanford University Institutional Review Board. All monitoring was performed with written informed consent from parents and guardians in the NICU. Patients who were in an open crib, without significant complications, and not already enrolled in another study were eligible to enroll. Convenience sampling was used, with subjects recruited based on attending physician referral of eligible patients. Infants with concern for active infection, need for supplemental oxygen therapy, or vasopressors were excluded. To ensure patient safety, routine medical care and protective measures were not altered during monitoring sessions. Data being obtained from the study were not made available to clinicians during the course of their care for patients.

### *2.1. Study Design and Measurements*

This manuscript describes an iterative design process. The objectives of the design were to identify and measure the respiratory rate of a preterm infant within an open crib through the analysis of camera footage and comparison to current hospital monitoring standards. The iterative process involved sequential algorithm design and application on recorded data. The principle analysis of the study investigated the consistent ability to identify patients within the frame and ability to extract the subjects' respiratory rate. The secondary analysis compares respiratory rate from our non-contact monitoring to that recorded into the electronic medical record (EMR) by the current hospital standard of respiratory monitoring by ECG impedance pneumography. Any design that could not identify the patient, extract an RR for a majority of subjects, or did not correlate with the current standard was determined to be unsatisfactory, and the study would return to the previous design phase.

Data were collected at the bedside as continuous 48-h video recordings from each infant. Footage was recorded on o ff-the-shelf IP cameras from Wansview with 320X180 resolution at ten frames per second and provided raw frames in YUV format. Cameras were placed approximately 4–6 feet from the bassinet (Figure 1). Simultaneously, vital sign data were collected for usual patient care from the ECG contact-based sensors and were extracted through the hospital's Research Data Export system. No changes were made to the patient's care while enrolled in this study.

**Figure 1.** Study setup. Red circle shows camera set up above an open neonatal intensive care units (NICU) crib. The hospital-standard monitor can be seen on the opposite side of the crib as the camera.
