Enhancing Clinical Decision-Making in Pediatric Monitoring: Learning Threshold Alarm Patterns to Predict Critical Illness.

backgroundPatient monitors assist caregivers in identifying deterioration earlier by using threshold alarms. Not all of the threshold alarms necessitate immediate action, but some are a result of the triggering of a physiological event. We aim to use pattern recognition techniques to identify threshold alarm signal patterns before the onset of critical illness, thereby enabling the faster and more effective detection of clinical deterioration and supporting better clinical decision-making.

methodSecondary data from 774 pediatric patients were extracted from the IMPALA Project conducted in the High Dependency Unit (HDU) at Queen Elizabeth and Zomba Central Hospitals in Malawi. The threshold alarm data were generated from the vital signs using WHO age cut-offs and GOAL3 age cut-offs. Time-segmented alarm analysis was conducted to examine the distribution of threshold alarms around each vital sign 8 h before the onset of critical illness events. Density-Based Spatial Clustering of Applications with Noise (DBSCAN) was used to generate threshold alarm signal patterns for each signal per individual before the onset of a critical illness event. We used three machine learning approaches, random forest, support vector machine, and decision tree, to learn threshold alarm patterns in signals preceding critical illness events.

resultsThe total threshold alarm summed up to (3,910,083) in total for WHO and (2,041,740) for GOAL3. Temporal distributions of ECGRR, ECGHR and oxygen saturation rate (SPO2) threshold alarms were observed, revealing patterns before the onset of the critical illness events. A pattern of most threshold alarms was distributed around (40-60) for ECGRR upper threshold alarms and (0-20) for ECGRR lower threshold alarms, (80-85) for ECGHR lower threshold alarms and (140-160) for ECGHR upper threshold alarms, and (85-90) for SPO2 for death (CPR and PICU), around WHO threshold alarms. For sepsis, most of these threshold alarms were distributed around (40-50) of ECGRR upper threshold alarms and (0-20) for ECGRR lower threshold alarms, (150-180) for ECGHR upper threshold alarms, and (85) for SPO2 for WHO threshold alarms, and most of the threshold alarms had a duration of less than 30 s. The results indicate that the random forest classifier performed better in learning the threshold patterns, with an accuracy of 93% and an area under the curve of 92, compared to using the support vector machine learning model and decision tree, which had an accuracy from a classification report of 85% and 94%, with low death and sepsis precision, recall, and F1-Score.

conclusionsThe analysis of threshold alarm data before critical illness events has provided valuable insights into threshold alarm patterns associated with death and sepsis. The data revealed distinct patterns in ECGRR, ECGHR, and SPO2 signals, and most of the threshold alarms were in the lower duration. The random forest classifier effectively distinguished these learned patterns around death and sepsis events compared to other algorithms. Further studies are required on the use of algorithms on all vital sign signal features in clinical settings.