Perceived quality of a nighttime hospital soundscape

2.1.1 Patient questionnaire

The patients filled out a questionnaire in the morning after each night they participated in the experiment. Complete data were obtained from 32 observations. The questionnaire was conducted in the Dutch language, and for convenience, terms are translated into English. It included five self-ratings of sleep quality, eight ratings of disturbing sounds, and one overall assessment. The patients’ questionnaires used Likert scales with four levels: self-ratings: completely disagree – disagree – agree – completely agree; Environment ratings: never – sometimes – often – always; General rating: very annoying – somewhat annoying – somewhat pleasant – very pleasant. In the present exploratory analysis, we interpret the Likert scales as interval scales, with values 1, 2, 3, and 4.

2.1.2 Arousal detection

For an objective measurement of sleep quality, we employed Fitbit Inspire HR health tracking bracelets. In the sleep experiment, they were worn by three of the patients in each room. To prevent the “first-night effect” [47], bracelets were worn solely by patients who had been hospitalised for at least one night before the data collection. The device gathers time-stamped data at a 30 s resolution regarding the user’s four sleep stages – i.e., awake, rapid eye movement sleep, light sleep, and deep sleep. Fitbit technology has demonstrated a high degree of reliability in identifying transitions from deeper sleep stages to awake when compared to polysomnography, the gold-standard (although intrusive) scientific method for tracking sleep. On the other hand, it is largely ineffective at differentiating between deep and light sleep stages [48]. Consequently, starting from the raw Fitbit data, we filtered only and exclusively the events corresponding to a transition from a deep or light sleep stage to an awake state and labelled them as arousals. Data were captured from 19 patients in 30 observations, totalling 474 arousal events.

2.1.3 Statistical methods

In the post hoc analysis of data, we employed only standard statistical methods: first, multivariate analysis of variance (MANOVA) and univariate analyses of variance (ANOVAs), with assumptions tested in cases of significance; second, exploratory factor analysis (EFA) to identify latent factors in the patients’ questionnaire responses (Section 3.1.1); third, comparison between dependent (latent factors) and independent (measured) variables in a cross-correlation analysis, using a non-parametric statistic (Section 3.3).

2.2.1 Sound recordings

In parallel with capturing the sleep quality state of patients using Fitbit devices, we recorded the room’s acoustic environment using an Audio-Technica AT2020 cardioid microphone (mono, 16 bits, 48 kHz). For the microphone stand not to encumber normal movement and activities in the room, it was placed adjacent to the wall facing the entrance door (Figure 1). The microphone, powered by a Blue Icicle preamplifier, was controlled by a MATLAB script running on a common laptop. The software recorded continuously over the course of the whole night, i.e., approximately for 10 h, using the same audio recording equipment and level settings.

Figure 1

Layout of a four-patient room with microphone placement, as seen from above.

All the captured data, sleep tracking and audio recordings, were time stamped and synchronised. When triggered by a Fitbit-detected arousal, the software extracted a 30 s audio file capturing the room soundscape immediately preceding the disturbance. In 45 out of the 474 detected arousals, there was a short silent gap in the recording due to technical issues, and they were therefore excluded. The remaining 429 excerpts had a total duration of 3 h, 42 min, and 30 s. In the next step (see Section 3), this set of soundscape recordings was tagged and analysed, with the purpose of identifying which sounds and events may have acted as external causes for the detected sleep disturbances.

2.2.2 Annotation procedure

The procedure of sound source annotations had been developed by two of the authors of this article for the evaluation of recordings of urban soundscapes during the COVID lockdown [5], and it was adapted for the present study. Thirty participants, here referred to as “annotators” were recruited to evaluate the 429 soundscape recordings, all confirmed being in normal health and having no hearing loss at the time of the task. In a first round, 18 participants were recruited by snowball sampling among undergraduate and research students currently in or having completed a sound-focused class. Their mean age was 26 years, with a range between 18 and 40, and consisted of 11 females and 7 males. Their work lasted approximately 90 min. Two of them were excluded due to low consistency in performing the task. The average number of soundscape recordings by each annotator was 5.5. To assure high quality in ratings, we then recruited a second round of 12 annotators considered as expert raters (research students and assistants in the authors’ labs, together with the authors themselves). Their mean age was 30 years, ranging between 24 and 55, and comprised 10 females and 2 males. The average number of soundscape recordings by each expert was 27.5, taking 3–4 h.

Each file was opened using the Audacity^® software with settings as follows: spectrogram (80–8,000 Hz Mel scale), grayscale, and full-width and full-height display. The annotators wore headphones (Audio Technica ATH-M70X or a similar closed-back monitor headphone). Annotators were instructed to mark as many sound sources as they could (aiming at around 6–15 in each 30 s recording), focusing on “sounds that you think could disturb somebody’s sleep.” In the Audacity^® interface, they would select the start and end points of the sound event in the corresponding spectrogram using the “Label” function (Figure 2). A set of nine labels, notably “beep,” “breath,” “clothes,” “cough,” “footstep,” “furniture,” “mechanical,” “snore,” and “speech,” extracted from the thematic analysis of the medical staff’s interviews (Section 2.1) were provided to annotators to be used in the labelling task. A “other” category was added. The annotators were then asked to label each selected sound event using the nine categories provided (or the “other”) and to rate the Annoyance that each annotated sound source might cause.

Figure 2

Screenshot of Audacity^® interface showing a spectrogram of a sound file with several labelled annotations, and a window with the exported text file. Note that one label was misspelt in this example.

A three-step Likert scale was employed, with numerals 1, 2, and 3 referring to low, medium, and high perceived Annoyance, respectively. In the instructions given to the annotators, it was stressed that Annoyance should be rated within the given category of sound and independently of loudness and duration. Hence, for example, “breath 3” would indicate a highly annoying sound event, and more annoying than “snore 2” even if the latter might very well be louder than the former; meanwhile, duration was captured by the label’s end and start points. The three-step scale was adapted from the protocol developed by one of the authors in the study by Lindborg [49] to assess an estimate of the liking (where “dislike,” “neutral,” “like” were coded as numerical −1, 0, and 1, respectively) of sound sources in indoor (restaurant) soundscapes. Thus, in the present study, annotators were instructed to add a number from 1 to 3 to the word label for each sound event in the Audacity^® software and export all labels in text form.

3.1.1 Analysis of patient questionnaires

The means for self-ratings across patients, in order from agreeing more to less, were “Quiet falling asleep” (mean = 2.88), “Sleep satisfaction” (2.62), “restless sleep” (2.59), “awoken by sounds” (2.47), and “difficulty falling asleep” (2.22). The differences in means were very small between these scales. Among the environment ratings, we retained six for further analysis. In order of higher to lower frequency of causing disturbance (to the patients), they were: “Snoring [by other patients]” (mean = 2.03), “Actions of medical staff” (1.78), “Voices of medical staff” (1.62), “Personal medical devices” (1.47), “Voices of other patients or visitors” (1.47), and “Trolleys” (1.44). Here as well the differences are rather small, but there are intriguing patterns that will be investigated further below. The mean general assessment of “Soundscape quality” was 2.84. Please note that these means are reported here for the convenience of making an initial comparison. See Figure 3 for distributions on questionnaire responses. See also Datafile 1 for aggregated patient data.

Figure 3

Responses on the patient questionnaire. Histograms in top row: self-ratings; middle and bottom rows: room ratings; bottom right: overall assessment of the sonic environment. X-axes are Likert scale labels; Y-axes represent the density of response alternatives.

To explore whether the questionnaire responses depended on the experimental conditions, we conducted a MANOVA taking the questionnaire scales as dependent variables, and night, room, and patient as the independent factorial variables (with 2, 7, and 20 levels, respectively). Results revealed that responses did not significantly differ between nights (Pillai = 0.93, F = 2.3, p = 0.33) or patients (Pillai = 6.2, F = 1.3, p = 0.073), but did so between rooms (Pillai = 4.4, F = 1.6, p = 0.044*). To find out in what way, we conducted a series of univariate ANOVA, taking in turn each of the questionnaire responses as the dependent variable and only room as the independent variable. For three scales, responses varied significantly between rooms: “Soundscape quality,” “Quiet falling asleep,” and “Voices of medical staff.” For each of the three, the standard ANOVA assumptions were met: observations were independent (i.e., separate patients in separate rooms); residuals were normally distributed according to Shapiro-Wilk’s test; and there was homogeneity of variance according to Levene’s test. Furthermore, plots revealed no systematic bias between scale responses, depending on room. These results reflect the subjective experience of the patients during a night of sleep.

The 12 rating scales in the questionnaire were developed with the assumption that they would cover essential aspects of the patients’ sleeping experience. They can be understood as latent factors in the data, and EFA provides an estimate of the correlations between rating scales and the latent factor(s). In the present study, we used the nfactors() function in the psych package [50] running in R [51], specifically the output of VSS analysis [52] to identify three optimal factors. This supports the questionnaire’s division into three sections. The fa() function from the same library, with settings for ordinary least squares regression and promax rotation, then yielded a minimum residual solution explaining 59% of the variance in the data with factors that lent themselves to a straightforward interpretation (Table 1).

The first factor, explaining 26% of the data, loaded onto “Voices of medical staff,” “Actions of medical staff,” and negatively onto “Soundscape quality.” It also captured two aspects of self-ratings, “Awoken by sounds” and negatively by “Quiet falling asleep.” It appears that the underlying factors capture sound events happening in the room, and the patients’ evaluation of the resulting quality. We labelled the latent factor “Room.” The second factor, explaining 19% of the data, loaded onto “Trolleys,” “Personal medical devices,” and “Voices of others,” and we labelled it “Objects.” The intriguing case of voices by non-staff is further discussed in the section on sound source classification. Finally, the third latent factor, explaining 14% of the data, loaded onto “Difficulty falling asleep” and “Restless sleep,” thus quite clearly capturing a negative evaluation of the sleep as experienced by the patient himself or herself.

In summary, the factor analysis suggests that the nighttime soundscape, as evaluated by patients, was negatively affected by voices and actions produced by medical staff. They also noted sounds from specific objects and voices from others. Several questionnaire responses reflect that patients were disturbed during the night (in particular, “Restless sleep” and “Awoken by sounds”). To support these subjectively based findings, we also measured sleep quality in an objective way, described in the next section.

3.1.2 Analysis of arousal data

Similar to the analysis of questionnaire data described in the previous section, we conducted an ANOVA by taking the number of arousals per observation as the dependent variable and the experimental conditions as independent factors. Results revealed that arousals did not significantly differ among nights, rooms, and patients (p > 0.3 in all cases). This should be interpreted to say that sleep quality did not depend on experimental conditions.

However, the patients did report different experiences, in regards to “Soundscape quality,” “Quiet falling asleep,” and “Voices of medical staff,” which is not reflected in the Fitbit data. It might be that the multifaceted questionnaire was able to capture subjective and more subtle nuances in the sleep experience. Or, it might be that the patients’ awareness of reduced sleep quality did not come from memories (the following morning) of how many times they had been disturbed during the night, and most certainly not by what exact sound. Rather, the questionnaire responses would be more easily explained as based on observations patients had consciously made before falling asleep, such as being annoyed by the lack of quiet or staff voices or experiencing such incidents upon awakening in the morning. This is not a fault of the observational acuity of questionnaire respondents, but simply reflecting that it is unusual, to say the least, to be rational about sounds we perceive while sleeping [31] and extremely hard to remember anything external at all that might have happened during sleep. So how could we find out which noises might have contributed to nighttime arousals, stirring patients and half-waking them up? To investigate the potential causes of sleep disturbances, recordings of the sonic environment were analysed.

3.2.1 Sound source classification

The semi-structured interviews with medical staff pointed towards a rather large number of potential sources for disturbing noises. This information was used in the development of the patient questionnaire. In addition, we conducted an initial analysis of sound recordings, which yielded six categories (plus “other”) of sound sources and events [25]. These analyses constituted the starting point for a detailed classification of the sound sources, and their Annoyance level, at the nighttime hospital wards. Going from the patient questionnaire, via the initial analysis, to the second (soundscape) analysis, the following section explains how we interpreted sources and events in the sleeping environment. See Figure 4 for a schematic overview of the process of classification. Note that in the context of the present study, “sound source” and “sound event” are considered to be the same thing.

Figure 4

Schematic representation of the classification of sound sources used in the case study, integrating terms and labels from three successive steps: (1) patient questionnaire, (2) initial analysis of soundscape recordings, and (3) detailed annotations and analysis of recordings.

The overall most prevalent kind of sound was “snoring.” This category in the patient questionnaire was split into “snoring” and “coughing” in the initial analysis and then in the soundscape analysis further detailed to include “breath” as a separate category. The questionnaire’s categories for “staff voices” and “other voices” were merged in “speech,” and through detailed listening. it was determined that all voices were in fact made by personnel and categorised accordingly. The “staff actions” that the patients rated might refer to different actions and sound sources. It appeared that the most common source directly attributable to personnel actions was “footsteps.” Meanwhile, the patient questionnaire rated “trolleys,” which in the initial would be “incidental sounds” and as such fairly common in the soundscape analysis, where they were divided into two distinct categories, “mechanical” (typically, trolley sounds, also metal plates, etc.) and “furniture” (doors closing, beds creaking, etc.). The two categories “TV, telephone” and “Liquids (toilet flushing)” in the patient questionnaire were in the initial analysis merged under the “other sounds” label but were not present to the ear in the soundscape analysis. Finally, “medical devices” or “alarms” (heard as “beeps”) retained a distinct category throughout the process.

This yielded nine categories for the soundscape analysis. They were labelled as follows: beep, breath, clothes, cough, footstep, furniture, mechanical, snore, and speech. In addition, a rest category was included as: “other (specify),” though eventually, no annotated sound needed to use it. The nine basic categories were allocated to two higher-level and non-overlapping categories (i.e., forming a taxonomic clade) following a procedure previously used by the authors [49,5].

Sounds labelled as “breath,” “cough,” and “snore” were all produced by patients. Note that they are mouth-made sounds rather than voices. So, what about “patient voices”? Based on annotations and subsequent careful re-listening by the authors, there was no sound event that could clearly be understood as “patient voices,” for example, either a patient talking in their sleep, or patients chatting between themselves. There would have been events where a staff member speaking with a patient (perhaps the staff responding to a call from the latter or administering medicine); however, in the present analysis, it appeared reasonable to group such sound events under “staff actions.” Sound labelled as “footstep” and “speech” were put under “actions” by medical staff, while “mechanical,” “beep,” and “furniture” were considered “objects” that are also under the control of medical staff. Finally, sounds labelled as “clothes” could have been generated by movements either by personnel or patients. Since it was impossible to know which, “clothes” was not included either.

3.2.2 Results from annotations

In total, annotators identified 9,296 sound events in the recordings. See Datafile 2 for the complete annotations data. Raw annotations were tidied up (changing to lowercase, removing trailing blank spaces and non-letter symbols, correcting spelling mistakes, grouping by synonyms, e.g., snore-snoring or breath-breathing, and so forth) and allocated to the nine predetermined categories. In a small number of cases, the annotator had used the “other” category. These were individually listened to by the researchers and either interpreted in one of the predetermined nine categories or removed. After this, the repartition among categories was as follows: “snore” (47.0%), “breath” (25.0%), “furniture” (6.4%), “mechanical” (5.9%), “speech” (5.7%), “clothes” (4.4%), “footsteps” (2.3%), “beep” (2.1%), and “cough” (1.2%). See Table 2 for an overview of all the statistics that were derived.

From counts and percentages in the sound categories, two estimates of inter-rater agreement among the annotators (N = 28) could be calculated. For each annotator, we extracted the number of labels in each of the nine sound source categories relative to the total, and their Cumulative duration relative to the total, as captured by the labels. We then tested agreement in two ways: first, by using functions in the DescTools package [53], Krippendorff’s alpha was found to be 0.79 for label counts and 0.71 for label durations, and second, by calculating Spearman’s rho on the vectors pairwise. For the number of labels, the median correlation was 0.85 in a range [0.54, 0.90], and for durations, it was 0.80 in a range [0.37, 0.85]. While the evaluation of the Krippendorff statistic is context sensitive [54], we found that the overall level of agreement among annotators was good and decided to keep data from all 28 annotators.

Recall that the annotators gave each labelled sound a value (1, 2, or 3) to indicate its level of Annoyance (low, medium, or high) within that category of sound. We treated the three levels as numerical values on an interval scale. Error bars are plotted in Figure 5.

Figure 5

Annoyance (raw) rated by annotators (N = 28) in nine sound source categories; means with 95% confidence intervals. Compare with classification in Figure 4.

Looking at means within categories (see Datafile 2), the highest Annoyance estimates were for “cough” (mean = 2.80) and “snore” (2.09). However, these kinds of sounds were of different character: coughing appeared less frequently (1.2% of annotations) and were of shorter duration (1.42 s on average) than the very prevalent snoring (47%) that were typically longer (2.22 s). Slightly less annoying were “beep” (alarms and other signals) and “speech” (almost exclusively communication amongst the nursing personnel). Similar to the previous pair, their occurrence was different: beeps were less frequent and shorter than speech. The Annoyance levels of sounds under “furniture,” “footsteps,” and “mechanical” were comparable, but the durations differed greatly. In fact, the “furniture” category included intermittent sounds such as the closing of doors, while “mechanical” included several occurrences of continuous background noise from air conditioners. Finally, the least annoying (yet still a potential cause for sleep disturbance) were sounds from breathing and the rustling of clothes, which were similar in terms of duration and frequency. These observations caused us to carefully consider the duration of potentially disturbing sound events. As mentioned earlier, Annoyance was rated for each sound event on a scale 1–3. Noting that the annotators used this scale slightly differently, we z-scaled the values within each annotator, putting the centre at 2 (corresponding to ‘medium’ Annoyance) and giving the distribution a standard deviation of 1. Since the distribution of Raw Annoyance ratings within the annotator was always positively skewed, the variable scaled this way did not have a range of [1, 3] but instead [0.26, 10.9]. However, and importantly, the interquartile range was [1.1, 2.9], which is indeed close to [1, 3]. This motivates employing the z-scaled variable to operationally define two indices of the negative impact that sounds of different types might have on sleep. They are defined as follows:

Normalised Annoyance, A _z is the z-scaled Annoyance, A _raw, within each Annotator centred on 2:

Integrated Annoyance, A _int, is Normalised Annoyance, A _z , multiplied by the logarithm of the sound’s duration, dur :

Cumulative Annoyance, A _cum, is the sum of A _int for all sound annotations within a given time interval, t :

See Table 2 for an overview of the results, listing mean values for raw, scaled, Integrated, and Cumulative (total) Annoyance for each sound category and the three higher-level categories (as described in Section 3.3.2) in the annotations data. First, looking at the raw counts (or frequency), we note that almost half (47%) of the annotated events were “snoring.” Note that each individual snore was separately labelled and that they most often appeared in a series of three or more, sometimes but not always with more or less intensive “breaths” (25%) in between snores. “Furniture,” “mechanical,” and “speech” followed at around 6% of the annotations. Second, taking duration and rated Annoyance into account might produce a more nuanced picture of sleep disturbances. We note that “speech” had the highest Integrated Annoyance (because voices tended to be present for a longer time) followed by “mechanical,” which had the longest average duration, and “snore.” The last column in Table 2, Cumulative Annoyance, gives a measure for how much, in the whole dataset, the Integrated Annoyance for each sound category amounted to. Again, “snore” is clearly the primary culprit, followed by “breath,” “speech,” and “mechanical”; the remaining four categories, including “beep,” might be safely given less attention in future research.

3.2.3 To modelling a “typical night”

Second, we gathered information on the distribution of annoying sounds to chart out a “typical” night to identify sleep-disturbing sound sources. Estimates for Annoyance were obtained for pre-determined sound categories and post-determined higher-level categories. The number of occurrences and duration of the annotated sounds, and their Annoyance levels, were calculated to obtain Integrated and Cumulative Annoyance levels for each category as well as the three higher-level categories. Integrated Annoyance was then calculated in time windows over the nightly recordings (9:30 pm to 7:30 am) to obtain an estimation of the prevalence of sleep-disturbing sound events. Within the three higher-level categories, personnel-generated behavioural sounds have the highest level of Integrated Annoyance (1.44). Recall that Integrated Annoyance depends on the (logarithm of) duration of sound events, as well as their rated Annoyance level (Table 2). This finding is critical in supporting design-driven interventions towards the improvement of patients’ well-being in hospital wards. Sounds belonging to this category are, in fact, actively produced by nursing staff during their work routine (notably, “speech” and “footsteps”), and their potential to disrupt sleep increases as a function of their Cumulative duration over time. As such, they can be actively mitigated to reduce their negative impact. Within the object’s sounds, mechanical sounds have a higher Annoyance level. Even if sounds in this category occur less frequently than, for instance, furniture sounds, they tend to have a longer duration which is responsible for a higher Integrated Annoyance level. This might be due to the presence, within this category, of continuous sounds such as background noise generated by air conditioning and heating systems. Alarm (“beep”) sounds, very often indicated as a critical source of discomfort for hospital patients and nursing staff [55] present low Cumulative Annoyance. This is perhaps not surprising in the context of the ward under study, where alarms occur far less frequently than, for instance, intensive care units in which alarms represent a major health concern. In general, this is yet another evidence supporting the claim that the evaluation of the soundscape quality is highly context dependent [56].

See Figures 6 and 7 for illustrations of how Integrated Annoyance of different sound types developed over the course of a typical night, with “snoring” being by far the most annoying sound event, followed by “breath,” “speech,” and “mechanical.” The reader will recall that the sound samples analysed in this study immediately precede each single arousal as collected by the tracking bracelets. Therefore, the curves can be interpreted as an index of the probability that a given sound type caused a sleep disturbance at that point in time.

Figure 6

Estimated disturbance of sounds in the nine categories over the course of a typical night. Values are accumulated Integrated Annoyance in time windows of 20 min with 50% overlap.

Figure 7

Estimated disturbance by patient-generated sounds set against disturbance by personnel-generated sounds (behaviours and objects) during a typical night. Values are accumulated Integrated Annoyance in time windows of 20 min with 50% overlap.

Preliminary experimental findings highlight an accumulation of snoring sounds rated with the highest Raw Annoyance level in the interval from 12:00 am to 4:00 am. This might be due to contextual and environmental factors, i.e., it is the time of the night when patients are normally in a deeper sleep state, and there is a limited presence of other sound sources, i.e., annotators might have tended to interpret as more annoying those sound sources that “intruded” on the soundscape. This interpretation aligns with one of the co-authors’ previous interpretation that noise sensitivity (a predictor of Annoyance), as a self-report measure, captures an evaluative predisposition towards sounds rather than aspects of auditory processing or noise exposure per se [57]. The highest sound Annoyance is found between 3:00 am and 4:00 am. In this frame, a higher Integrated Annoyance for patient-generated sounds (breathing and snoring) corresponds to higher Annoyance level for speech, which is a sound voluntarily generated by nursing personnel (Table 2). This finding suggests a direct correlation between the increased occurrence of speech sounds and disturbance in patient sleep, and as mentioned, its negative impact could be prevented if nurses are made aware of the consequences of their sound-producing behaviour. Note that this time frame coincides with the last “big” night shift before the last night shift (which takes place around 6 am, see next paragraph). During this shift, nurses are active with major caregiving activities such as administering medicines. In shared patient rooms, this will create awakening and further induce more sound-producing events such as coughing or conversations between patients or patients and nurses. In addition, it is commonly known that nurses do activities to keep themselves busy and active to maintain an awake state. They may socialise in the central nurse station or in the corridor, and these activities may contribute to nocturnal noise.

Towards the end of our data collection timeframe, Cumulative Integrated Annoyance tends to grow for all the sound categories. Between 6:00 am and 7:00 am speech, furniture, beep, footsteps, and cough become more frequent and annoying (Figures 6 and 7). This is most likely due to the progressive awakening of the ward’s activity with the handover from the night shift, breakfast delivery to patients, cleaning, the start of the medical exams, and all activities that will involve more communication and action-generated sounds. It confirms prior findings that hospital patients perceive their sleep to be inadequate mostly in the early morning [58]. As we further discuss towards the end of this paper, a systemic change and a shift in caregiving activities could help postpone early awakenings. Some hospitals, for example, skipped night treatments to facilitate long sleep. Technological solutions which provide systematic knowledge on how work shifts impact on the quality of the ward’s soundscape and thus on patients’ sleep would be able to inform long-term structural interventions on the hospital organisation towards improved patients’ sleep.