Automated Detection of Intraoperative Hypoxemia

Hypoxemia was defined as SpO₂ less than 95% during general anesthesia regardless of the duration. The limit of 95% was selected in line with the definition of hypoxemia of perioperative respiratory adverse events [5 (link)]. All 13,130 vitalDB records were manually checked for the detection of hypoxic episodes by an anesthesiologist. The data retrieved from the VitalDB database were rechecked manually by the second anesthesiologist for quality control. The observed events were marked and annotated in all the databases. The true hypoxemia was verified by SpO₂ value, electronic medical records, vital signs, including heart rate, blood pressure, arterial blood pressure waveform, and pulse oximetry, plethysmographic waveform, and ventilator parameters, including EtCO₂ curve, PIP, and TV. Genuine hypoxemic events were discriminated using the following exclusion criteria: 1) when the measured pulse oximetry pulse rate differed by more than 20% from the electrocardiogram heart rate, the associated SpO₂, [2 (link)] 2) when anesthesiologists in charge recorded inaccurate measurement of oxygen saturation, and 3) when plethysmogram waveforms were severely distorted owing to position change or external pressure so that the signal quality reported in the monitor was severely low [2 (link)]. Artifacts that were mostly caused by motion, positioning, electrocautery, sensor dislodgement, and low peripheral perfusion were not annotated as hypoxemia and were retained in the dataset without further processing. Periods of hypoxemic episodes, divided into induction, maintenance, and emergence, were annotated in each database to determine the incidence of intraoperative hypoxemia in each period. We also investigated the total time of hypoxemia, duration from the minimum saturation to recovery until the saturation recovered to the initial value where saturation began to gradually decrease, and the cause of the hypoxemia episode suggested by the anesthesiologists based on various indicators. Patients were assigned to one of the four groups to confirm the incidence of the hypoxemic episodes stratified by age: 0–28 days (neonate), 29 days to 12 months (infant), 1–7 years, and 8–18 years [2 (link)]. Demographic data including age, sex, height, and weight were analyzed in each group.
Several parameters, such as SpO₂, EtCO2, FiO₂, TV, and PIP included in 1-minute length of segments from the determined hypoxemia episodes to the corresponding 1 minute ahead, were extracted. If a hypoxemic event defined as SpO₂ dropping below 95%, in the upcoming 1 min was observed, the target was assigned as 1; it was assigned 0 otherwise. Note that target 1 was assigned for the segment where the input SpO₂ was less than 95% while being included in the hypoxemia episode. Moreover, the patient’ demographic information (age, sex, height, and weight) was extracted corresponding to the segment and the target.