Automated False Alarm Reduction in a Real-Life Intensive Care Setting Using Motion Detection.

BACKGROUND: Contemporary monitoring systems are sensitive to motion artifacts and cause an excess of false alarms. This results in alarm fatigue and hazardous alarm desensitization. To reduce the number of false alarms, we developed and validated a novel algorithm to classify alarms, based on automatic motion detection in videos. METHODS: We considered alarms generated by the following continuously measured parameters: arterial oxygen saturation, systolic blood pressure, mean blood pressure, heart rate, and mean intracranial pressure. The movements of the patient and in his/her surroundings were monitored by a camera situated at the ceiling. Using the algorithm, alarms were classified into RED (true), ORANGE (possibly false), and GREEN alarms (false, i.e., artifact). Alarms were reclassified by blinded clinicians. The performance was evaluated using confusion matrices. RESULTS: A total of 2349 alarms from 45 patients were reclassified. For RED alarms, sensitivity was high (87.0%) and specificity was low (29.6%) for all parameters. As the sensitivities and specificities for RED and GREEN alarms are interrelated, the opposite was observed for GREEN alarms, i.e., low sensitivity (30.2%) and high specificity (87.2%). As RED alarms should not be missed, even at the expense of false positives, the performance was acceptable. The low sensitivity for GREEN alarms is acceptable, as it is not harmful to tag a GREEN alarm as RED/ORANGE. It still contributes to alarm reduction. However, a 12.8% false-positive rate for GREEN alarms is critical. CONCLUSIONS: The proposed system is a step forward toward alarm reduction; however, implementation of additional layers, such as signal curve analysis, multiple parameter correlation analysis and/or more sophisticated video-based analytics are needed for improvement.

Classifying Alarms: Seeking Durability, Credibility, Consistency, and Simplicity.

Alongside the development and testing of new audible alarms intended to support International Electrotechnical Commission 60601-1-8, a global standard concerned with alarm safety, the categories of risk that the standard denotes require further thought and possible updating. In this article, we revisit the origins of the categories covered by the standard. These categories were based on the ways that tissue damage can be caused. We consider these categories from the varied professional perspectives of the authors: human factors, semiotics, clinical practice, and the patient or family (layperson). We conclude that while the categories possess many clinically applicable and defensible features from our range of perspectives, the advances in alarm design now available may allow a more flexible approach. We present a three-tier system with superordinate, basic, and subordinate levels that fit both within the thinking embodied in the current standard and possible new developments.

Using Supervised Machine Learning to Classify Real Alerts and Artifact in Online Multisignal Vital Sign Monitoring Data.

OBJECTIVE: The use of machine-learning algorithms to classify alerts as real or artifacts in online noninvasive vital sign data streams to reduce alarm fatigue and missed true instability. DESIGN: Observational cohort study. SETTING: Twenty-four-bed trauma step-down unit. PATIENTS: Two thousand one hundred fifty-three patients. INTERVENTION: Noninvasive vital sign monitoring data (heart rate, respiratory rate, peripheral oximetry) recorded on all admissions at 1/20 Hz, and noninvasive blood pressure less frequently, and partitioned data into training/validation (294 admissions; 22,980 monitoring hours) and test sets (2,057 admissions; 156,177 monitoring hours). Alerts were vital sign deviations beyond stability thresholds. A four-member expert committee annotated a subset of alerts (576 in training/validation set, 397 in test set) as real or artifact selected by active learning, upon which we trained machine-learning algorithms. The best model was evaluated on test set alerts to enact online alert classification over time. MEASUREMENTS AND MAIN RESULTS: The Random Forest model discriminated between real and artifact as the alerts evolved online in the test set with area under the curve performance of 0.79 (95% CI, 0.67-0.93) for peripheral oximetry at the instant the vital sign first crossed threshold and increased to 0.87 (95% CI, 0.71-0.95) at 3 minutes into the alerting period. Blood pressure area under the curve started at 0.77 (95% CI, 0.64-0.95) and increased to 0.87 (95% CI, 0.71-0.98), whereas respiratory rate area under the curve started at 0.85 (95% CI, 0.77-0.95) and increased to 0.97 (95% CI, 0.94-1.00). Heart rate alerts were too few for model development. CONCLUSIONS: Machine-learning models can discern clinically relevant peripheral oximetry, blood pressure, and respiratory rate alerts from artifacts in an online monitoring dataset (area under the curve > 0.87).

Association between exposure to nonactionable physiologic monitor alarms and response time in a children's hospital.

BACKGROUND: Alarm fatigue is reported to be a major threat to patient safety, yet little empirical data support its existence in the hospital. OBJECTIVE: To determine if nurses exposed to high rates of nonactionable physiologic monitor alarms respond more slowly to subsequent alarms that could represent life-threatening conditions. DESIGN: Observational study using video. SETTING: Freestanding children's hospital. PATIENTS: Pediatric intensive care unit (PICU) patients requiring inotropic support and/or mechanical ventilation, and medical ward patients. INTERVENTION: None. MEASUREMENTS: Actionable alarms were defined as correctly identifying physiologic status and warranting clinical intervention or consultation. We measured response time to alarms occurring while there were no clinicians in the patient's room. We evaluated the association between the number of nonactionable alarms the patient had in the preceding 120 minutes (categorized as 0-29, 30-79, or 80+ alarms) and response time to subsequent alarms in the same patient using a log-rank test that accounts for within-nurse clustering. RESULTS: We observed 36 nurses for 210 hours with 5070 alarms; 87.1% of PICU and 99.0% of ward clinical alarms were nonactionable. Kaplan-Meier plots showed incremental increases in response time as the number of nonactionable alarms in the preceding 120 minutes increased (log-rank test stratified by nurse P < 0.001 in PICU, P = 0.009 in the ward). CONCLUSIONS: Most alarms were nonactionable, and response time increased as nonactionable alarm exposure increased. Alarm fatigue could explain these findings. Future studies should evaluate the simultaneous influence of workload and other factors that can impact response time.