A Knowledge-Based Approach to Automatic Detection of Equipment Alarm Sounds in a Neonatal Intensive Care Unit Environment

Abstract

A large number of alarm sounds triggered by biomedical equipment occur frequently in the noisy environment of a neonatal intensive care unit (NICU) and play a key role in providing healthcare. In this paper, our work on the development of an automatic system for detection of acoustic alarms in that difficult environment is presented. Such automatic detection system is needed for the investigation of how a preterm infant reacts to auditory stimuli of the NICU environment and for an improved real-time patient monitoring. The approach presented in this paper consists of using the available knowledge about each alarm class in the design of the detection system. The information about the frequency structure is used in the feature extraction stage, and the time structure knowledge is incorporated at the post-processing stage. Several alternative methods are compared for feature extraction, modeling, and post-processing. The detection performance is evaluated with real data recorded in the NICU of the hospital, and by using both frame-level and period-level metrics. The experimental results show that the inclusion of both spectral and temporal information allows to improve the baseline detection performance by more than 60%.

Keywords: Acoustic event detection; alarm detection; neonatal intensive care unit; neural networks; non-negative matrix factorization; sinusoid detection.

FIGURE 1.

Graphical description of terms used to denote particular alarm properties. Only fundamental frequency is depicted for clarity of presentation .

FIGURE 2.

Automatic system for alarm detection: (a) set of parallel detectors, one per class, and (b) diagram of an individual detector.

FIGURE 3.

An example of a spectrogram (a) of audio recording and the detected sinusoidal components (b).

FIGURE 4.

The output of the period probability estimation. Circles correspond to the estimated period timestamps after applying a threshold and crosses are the reference period timestamps.

FIGURE 5.

Histogram of SNR values over all labelled alarm samples.

FIGURE 6.

The DET graphs for different statistical models. Circles correspond to points closest to EER.

FIGURE 7.

PB-ERR metric result as a function of the tolerance value . The bold black line corresponds to the average over the alarm classes. The far right points of the curves are the reported PB-ERR results using 49% tolerance, by classes (%): a1 33.33, a3 27.90, a6 38.99, a7 19.51, a8 33.50, a10 8.50, a16 69.98.