An Anesthesia alarm system based on neural networks.

The alarms in most patient monitoring devices are threshold alarms that call attention to the value of a single measured parameter. These threshold alarms do not aid in diagnosis and tend to have high false alarm rates. This dissertation describes a neural network based anesthesia alarm system that provides intelligent alarms. This system calls attention to a specific problem with a diagram and text message. Neural networks are mathematical models of interconnected neurons that can be trained to define a particular function. The neural networks in the system were trained to recognize alarm events based on input from monitored signals. A sensor array at the patient’s mouth provided CO2, flow and pressure signals from which features were extracted. These breath-to-breath features were the inputs to the neural networks. Each neural network output corresponded to one of 13 specific alarm messages. The system was trained by creating each of the alarm conditions and recording the corresponding inputs. The recorded data were used to train the alarm system neural networks using the backward error propagation algorithm. Two versions of the alarm system were developed. Version 1 of the alarm system was trained and tested using a lung simulator. Fourteen critical events were created 20 times each. After training, the system was able to correctly identify 99.5% of 20 repetitions of 13 created breathing circuit failures. The system received further training using two mongrel dogs. The system was then tested on seven mongrel dogs. The system correctly identified 89.3% of 1,629 events created during controlled ventilation and 75.8% of 236 events created during spontaneous breathing. Version 2 of the alarm system was also tested in animals. Thirteen alarm conditions were created at various ventilator settings in five mongrel dogs. Data from four of the dogs were used to train the alarm system neural networks. When the system was trained, data from the fifth dog were used to test the system. A total of 746 events was created during animal testing. The alarm system was 95.0% accurate during controlled ventilation and 86.9% accurate during spontaneous breathing. With informed consent from the patient, version 2 of the alarm system was clinically tested in the operating room. An observer watched and recorded all breathing circuit events in 20 patients. During 43.6 hours of clinical testing, 57 recognizable events were observed. Of the observed events, 94.7% were correctly reported by the system. The alarm system produced 74 false positive alarms, for which no alarm condition or event was observed.