Children admitted to the Pediatric Intensive Care Unit (PICU) are at risk of deterioration, which if left untreated or undetected can result in cardiac arrest. Outcomes after pediatric cardiac arrest, even for witnessed, in-hospital events, remain poor. Thus, early detection of deterioration is paramount; ideally, long before the risk of harm has increased. However, a low false positive rate is also important, to avoid alarm fatigue. The purpose of this work was to develop an algorithm for early detection of deterioration by analyzing instabilities that lead to cardiac arrest in local vital signs data.
With ethics board approval, 8-hour segments of vital signs data from children admitted to BC Children’s Hospital’s PICU were extracted from a local database (n=96) and split, by patient, into training (n=67) and test (n=29) sets. A rule-based algorithm was chosen for its simplicity and transparency to gain acceptance among physicians. It detects vital sign instabilities based on variability, deviations from previous states, and the integral between baseline and trend. Thresholds were used to trigger potential events if multiple rules agreed on a potential event. A PICU attending physician provided the ground truth for episodes indicative of vital signs instability or their absence. Algorithm performance was calculated by splitting the trend data into 15-minute episodes, and counting the number of expected and predicted instabilities within these blocks.
Using the test data, the proposed algorithm detected 91.6% of episodes correctly (92.4% in the training data), with 83.5% true negatives, 8% true positives, 3% false positives and 6% false negatives. Thus, the specificity was 96.8% and the sensitivity 59.2% with the test data.
The rule-based algorithm demonstrated a clinically adequate performance. Feedback from a second physician indicated the need to include longer-term changes to anticipate future deterioration, which will be explored in future work.