While an increasing number of bedside medical devices, such as syringe pumps, have wireless connections that enable reliable data transmission to hospital information systems, many nonnetworked devices are still used in general hospital wards. Although dongles connected to the external output terminal of these devices may allow wireless connections [1,2], most devices are not equipped with external output terminals. Instead, these devices use auditory alert signals (alarm sounds) to notify medical staff of abnormal conditions. Medical staff members may not hear these alarms, especially when they are in distant areas or other private rooms. Between 2010 and 2015, the Japan Council for Quality Health Care reported 173 accidents and other incidents, including 23 cases of unnoticed alarms [3]. The report included the following comments about environmental factors:

These findings indicate the need for reliable alarm notification to ensure patient safety. Alarm sounds emitted by medical devices are regulated by the International Organization for Standardization and the International Electrotechnical Commission (ISO/IEC) 60601-1-8. This standard specifies the melodies and lengths of alarm sounds to reduce the risk of misunderstanding, confusion, and omission of alarm sounds from various medical devices, even when these sounds overlap and reverberate. This standard prescribes that the sounds should be organized based on the priority of corresponding abnormal situations, with alarm sounds for different situations varying in melody and length. Thus, sound event detection (SED) is expected to identify every kind of abnormal situation detected by medical devices [4-8]. The standard also defines visual alarm signals, but monitoring the signals of multiple devices with cameras is not feasible without blind spots.

As SED can be implemented using a single (monaural) microphone, it was selected as the approach to detect alarm status. Clinical application of SED requires robustness against noise because environmental noise in hospital wards is generally substantial. SEDs with deep learning have been found to be sufficiently robust against noise [9,10].

This study proposes a deep learning–based method for classifying patient abnormalities detected by medical devices with polyphonic alarm sounds collected with a monaural microphone. The ability of the classifier to identify abnormal states of these devices was evaluated using simulation datasets of their alarm sounds superimposed on hospital ward noise (HWN). Therefore, the objective of this study was to design and evaluate a convolutional recurrent neural network (CRNN) for accurately detecting and classifying multiple, overlapping alarm sounds from medical devices in a simulated noisy hospital environment. We hypothesized that a hybrid CRNN model could achieve high performance suitable for clinical application by effectively capturing both the spectral and temporal features of the alarm sounds.

Accurate transmission of alarms from unconnected medical devices requires precise recognition of visual and auditory alarms. Several studies have reported high accuracy in detecting simultaneous alarm sounds mixed with a substantial level of environmental noise [4-8]. For SED, deep learning is more robust against noise than conventional methods [9,10]. On the basis of these studies, an edge device was placed at the patient’s bedside to monitor abnormal conditions detected by multiple medical devices with individual alarm sounds. In our previous study, the classifier had F₁-scores of 0.727 at signal-to-noise ratios (SNRs) of 0 dB and applied a convolutional neural network (CNN) [11]. To our knowledge, no previous study has used a deep learning–based recurrent neural network (RNN) to detect polyphonic alarm sounds emitted by medical devices. Recent advances in attention-based architectures, such as the audio spectrogram transformer and conformer, have demonstrated strong performance in general SED tasks [12,13]. These models use self-attention mechanisms to capture long-range dependencies, potentially enhancing robustness in noisy environments. Exploring such transformer-based approaches for clinical alarm sound detection remains an important direction for future research.

The first approach to SED combines a Gaussian mixture model with a hidden Markov model, using features such as the mel-frequency cepstral coefficient from traditional methods of speech recognition [14,15].

Other approaches include the separation of sound sources by matching them using a template extracted from the input sound. This can be achieved by sound source separation techniques, such as nonnegative matrix factorization. Nonnegative matrix factorization monitors a single signal to create a basis matrix and identifies the separated sounds [16-18].

Recent approaches based on neural networks have significantly improved the performance of SED. One approach consists of SED of real-life sounds with feedforward neural networks based on a multilayer perceptron trained in a spectrum of mixed sounds [6]. An RNN with the ability to remember past states can process sequential information of the acoustic signal. RNNs with bidirectional long short-term memory have achieved excellent results in complex audio detection such as speech recognition and polyphonic piano note transcription [19-21].

Furthermore, CNNs commonly used in image recognition can robustly predict sounds with its filter shifted by both time and frequency axes [22]. However, long-term prediction remains difficult due to the limited width of the time window [9]. Therefore, although alarm sounds consist of relatively simple tones, it is necessary to predict not only the frequency axis but also the time axis to inform the priority with a pattern. Application of an RNN to polyphonic SED enabled long-term prediction by integrating the information over the time window. This study combines the strengths of both CNN and RNN to benefit from both approaches. A similar approach has shown excellent performance in automatic speech recognition [23,24].

Therefore, this approach was expected to achieve sufficient performance for potential clinical application in alarm sound recognition.

This study did not involve human participants or animal experiments. The recordings contained no speech, personal identifiers, or patient-related information. Therefore, ethics approval was not required in accordance with institutional and international research ethics policies.

The proposed classifier was found to successfully detect the status of nearby medical devices. Although video-based alarm detection was also considered, it was not feasible to detect the monitors with cameras without the introduction of a blind spot. Therefore, SED was considered a more feasible approach than video recognition.

This study describes the construction of a classifier using a large amount of artificial noise data. This classifier was used to evaluate polyphonic alarm sounds among a simulation dataset of HWN. Because the classification target is a sine wave, we expected that the target could be achieved with simple neural networks. However, BiSRNN did not achieve the target performance, requiring the application of BiGRU. Both the frequency and time axes required advanced processing to recognize alarm sounds. In addition, a convolutional layer with 4 layers outperformed one with 3 layers. The F₁-score of the classifier using BiGRU was 0.900 or higher, which was robust in the detection of polyphonic alarm sounds. Only VFP in each proposed classifier was unclear in the spectrogram due to VFP having the lowest sound pressure level of the 7 devices, thus resulting in low recall. In addition, masking of an impact sound over the entire frequency axis of the spectrogram would result in lower precision and overdetection.

The notification system should alert the hospital information system without missing any situations, whether alarms are sounding on a single device or on multiple devices simultaneously. Therefore, recall of class-wise metrics was considered the most important in evaluating clinical applications. In particular, recall for the ventilator, the most essential support device of the 7 evaluated, was 1.000.

CNN4+BiGRU had the highest F₁-score in the event-based overall metrics, and the recall values of ENP, CD, and the ventilator were 1.000 each. Only VFP showed a recall value below 0.950, primarily due to its lower sound pressure level. In some cases, detection occurred slightly earlier than the actual onset, which reduced the measured recall. However, the ventilator demonstrated low precision with frequent overdetections, suggesting that it could not be evaluated clinically because it could cause alarm fatigue [35]. In contrast, because the F₁-scores of ENP, VFP, and CD without external output were 0.900 or higher, the system was likely feasible to notify staff of the alarm status of medical devices that could not be connected to the hospital information system. Therefore, the proposed system demonstrated feasibility as an alarm sound detection system and can be further refined for clinical use.

This study had several limitations. First, the study evaluated only a limited range of medical devices from a single hospital, potentially limiting generalizability. Second, differences in hospital architecture (eg, room layouts, wall materials, and ambient noise) may affect sound propagation and detection. Third, despite ISO/IEC 60601-1-8 regulations, variations in alarm sounds across manufacturers and models were not assessed. Finally, large-scale live recording and annotation remain impractical; future work should use synthetic datasets and validate performance across diverse environments and device types.

Several reports have examined the classification of alarm sounds using deep learning. Evaluation of single sounds of horns and bicycle bells found that the 5-layer deep neural network that applied an integrated judgment process had F₁-scores of 0.99 or higher [4]. Because horns and similar devices do not create sine waves of digital sounds, they cannot be evaluated in the same way as alarm sounds of medical devices. A study of alarm sounds of medical devices in a neonatal intensive care unit found that most of the alarms had SNRs of 0 dB or higher [8]. That study, however, did not consider a classifier for polyphonic alarm sounds. A classifier using CRNN was found to be effective for polyphonic acoustic sounds [26,36].

Missed medical device alarms can lead to serious adverse events. To mitigate such risks, we developed a deep learning–based classifier for detecting polyphonic alarm sounds in hospital environments.

Alarm sounds emitted by medical devices are regulated by ISO/IEC 60601-1-8. Because this standard defines different tones and patterns for each device and priority, SED is expected to identify the device and priority successfully. Thus, we considered SED appropriate to determine alarm status. Automatic identification of alarm sounds in hospital rooms would facilitate safer medical care.

In the simulation experiment, the polyphonic alarm sound classifier showed excellent performance, with an F₁-score of 0.945 at an SNR of 0 dB. The proposed classifier demonstrated feasibility for clinical alarm sound detection and can be further optimized. When combined with network connectivity, this classifier could improve the notification of abnormal patient status detected by medical devices without requiring each device to be individually connected.