Sleep Quality Index for Sensor Data in Eldercare Monitoring

The elderly population aged 65 and older is increasing from 13% in 2010 to 19% in 2030, whereas the population of age 15-65 is decreasing [1]. Since older adults prefer to live independently, many of the health changes go undetected, such as dementia, frailty, and urinary tract infections (UTI) [2]. Detecting health changes early is crucial for promoting health and controlling healthcare costs. On the other hand, late health assessments or unreported problems can lead to poor quality of life [3]. Automatic health monitoring systems are a possible solution to identify and assess problems in their early stages, which provide more time for intervention to address the problems before they become serious. Such systems can be effective in supporting aging-in-place and allowing older citizens to maintain their independent lifestyle for as long as possible.

With the aging population increasing rapidly, the demand for innovative health monitoring solutions is more needed than ever. Sleep quality, in particular, has gained more attention from healthcare professionals. Poor sleep is recognized as a symptom of various health issues and can be the root cause the progression of serious conditions, such as cognitive decline and cardiovascular disease. Hence, integrating sleep quality monitoring into automatic health systems is essential for comprehensive eldercare.

Sleep quality is an important aspect of elderly mental and physical health [4-9]. Sleep is crucial for maintaining the brain and cognitive functions [10-12]. Poor sleep quality is a common factor among older adults with mild cognitive impairment (MCI), which is a transition stage between normal cognition and dementia [13]. More importantly, older adults with MCI are more likely to experience poor sleep than healthy older adults [14], and poor sleep is associated with an increased risk of progression from MCI to dementia [15]. Studies on animal models suggest that chronic poor sleep leads to increased cortical amyloid-beta (Aβ)—a hallmark pathology of Alzheimer's disease [16]. Sleep is therefore a critical aspect through which the brain maintains its cognitive health. When poor sleep becomes chronic, an acceleration of cognitive decline may occur [17].

In addition to the established links between poor sleep and cognitive decline, recent studies emphasize the role of sleep quality in overall health outcomes for older adults. Poor sleep is associated with a higher risk of falls, reduced immune function, and greater susceptibility to chronic conditions such as hypertension and diabetes [18, 19]. Addressing sleep disturbances through reliable monitoring and timely interventions could therefore play a pivotal role in preventing or delaying the onset of these conditions, thereby enhancing the quality of life for older adults. Furthermore, it is important to note that sleep disorders are often underdiagnosed in this population, which underscores the need for more effective monitoring solutions [20].

Sleep quality can be measured subjectively through questionnaires [21, 22] such as the Pittsburgh sleep quality index (PSQI) [23]. Questionnaire-based methods are not practical for long-term monitoring of older adults due to the burden of self-reporting. There are multiple quantitative ways to monitor sleep quality. Polysomnography is one way to compute sleep quality through a comprehensive recording of physiological changes that occur during sleep [24]. It requires multiple sensors to be placed on the patient's body, which makes it unsuitable for long-term sleep monitoring in the home. Actigraphy is another method to monitor sleep using nonobtrusive acceleration sensors that capture body movement data [25]. Wearable devices may have an accelerometer and heart rate monitoring, which may report sleep stages. The main drawback of these devices is the requirement to remember to charge them and wear them during sleep, which increases the burden on older adults who may have memory problems. Comfort is also a factor. Thus, the method is a sub-optimal solution for eldercare monitoring applications. The last technique is Ballistocardiography (BCG), which is an unobtrusive method for measuring heart rate, heart rate variability, respiration rate, and relative blood stroke volume based on the body movement induced by blood flow through the cardiovascular system, due to the heart's pumping mechanism [26]. BCG sensors are usually installed on the bed or under the mattress, which makes them more suitable for longitudinal monitoring of older adults.

While subjective measures like questionnaires provide valuable insights, they are not practical for continuous monitoring due to the reliance on self-reporting, which can be burdensome and less reliable in older adults. Objective methods, such as polysomnography, although comprehensive, are not suited for long-term use in home settings due to their invasive nature. On the other hand, actigraphy offers a less invasive alternative but still involves wearing devices that may not be ideal for older individuals who might find them uncomfortable or forget to use them. In contrast, BCG technology offers a promising solution by providing a non-intrusive, reliable way to monitor sleep patterns over time. Recent advancements in BCG have improved its accuracy and reliability, making it an even more viable option for eldercare settings [27, 28].

In this paper, we investigate a sleep quality measure that uses BCG but does not require sleep stages because these are often not accurate for older adults with noisy cardiac signals due to reduced stroke volume and increased stiffness in the vessels. We also avoid wearable devices due to the reasons mentioned above. Hence, we introduce a sleep quality metric from a bed sensor that is placed under the mattress and captures heart rate, respiration rate, and restlessness.

The contributions of this paper are: 1) a new sleep index that relies on a simple hydraulic-based bed sensor and does not require heart rate variability or sleep stages; 2) a penalty is applied to the sleep index if older adults do not meet the recommended sleep hours; and 3) a sleep normality score is introduced and incorporated to take into account unusual periods of sleep.

The paper is organized as follows. In section 2, we describe the eldercare monitoring system architecture. Section 3 describes the method, and Section 4 explains the sleep normality metric. Section 5 shows the dataset used in the study. Experimental results are displayed in section 6, and case studies are in section 7. Finally, section 8 provides conclusions.

A sleep measure is needed for tracking the sleep health of older adult residents. Our sleep quality index (SQI) consists of three parts: sleep efficiency, time in bed, and sleep normality. Sleep efficiency is defined as the ratio between sleep time and total time in bed (TiB). As shown in Eq. (1), sleep efficiency is computed by taking the total amount of time spent in bed (in minutes) and subtracting the number of minutes it took to fall asleep plus the minutes spent awake.

$\begin{gathered} { sleep \ efficiency }=\frac{\text { total TiB (minutes) }}{\text { total TiB (minutes) }} -\frac{(\text { minutes to fall asleep }+ \text { minutes awake })}{\text { total TiB (minutes) }}\end{gathered}$          (1)

As a proxy that does not require a classification of sleep versus wake time, we use restlessness in bed, as captured by the bed sensor. Specifically, we use the minutes of restlessness to account for the time when the older adult is in bed and not sleeping. Hence, Eq. (1) for sleep efficiency takes the form of Eq. (2), which we call SQI_restlessness.

$S Q I_{ {restlessness }}=1-\frac{\text { restlessness time (minutes) }}{\text { total TiB (minutes) }}$          (2)

The rationale behind using restlessness as a proxy for sleep efficiency lies in its ability to provide a non-intrusive measure of sleep quality. Traditional methods, such as polysomnography, often require the classification of sleep versus wake time, which can be cumbersome and less accurate, particularly in older adults with irregular sleep patterns [24]. By focusing on restlessness, our approach circumvents the need for extensive sensor setups while still delivering meaningful insights into sleep efficiency.

However, using restlessness as a proxy has its limitations and edge cases. For instance, restlessness may not perfectly correlate with wakefulness in all cases; some individuals might experience minimal movement during wakeful periods, leading to an overestimation of sleep efficiency [33]. Conversely, some sleep disturbances, like periodic limb movement disorder (PLMD), could cause significant restlessness without fully waking the individual, potentially leading to an underestimation of sleep quality [34]. Additionally, restlessness may not capture short, subtle awakenings that are typically recorded in polysomnography-based methods [35]. These edge cases highlight the trade-off between the convenience and non-intrusiveness of using restlessness as a proxy and the detailed accuracy provided by direct sleep vs. wake classification.

Note, however, that the SQI_restlessness percentage does not reflect if someone got enough total sleep. For example, someone who sleeps for 4 out of 5 hours in bed will have the same SQI as when he/she sleeps for 8 out of 10 hours in bed. To account for total sleeping time, we use a weight that penalizes the index value if the older adult sleeps fewer than the recommended number of hours for people in this age range. Based on the latest recommendations from the National Sleep Foundation's updated report, people 65 or older are recommended to sleep seven to eight hours per day [36]. Residents who spend less than this amount of time in bed should expect to have a lower sleep quality. Therefore, for these residents, a penalty is applied to their sleep quality index as shown in Eqs. (3) and (4).

$S Q I_{T i B}=\mathrm{W}_{T i B} *\left(1-\frac{\text { total restlessness time (minutes) }}{\text { total TiB (minutes) }}\right)$          (3)

$W_{T i B}=\left\{\begin{array}{ll}-\left(\frac{\text { total TiB }(\text { hours })-8}{8}\right)^2+1 & \text { if total } \mathrm{TiB} \text { (hours) }<8 \\ 1 & \text { if total } \mathrm{TiB} \text { (hours) } \geq 8\end{array}\right\}$          (4)

This formula penalizes residents who sleep fewer than 8 hours, with greater weight given to less time in bed. We chose 8 hours as a cutoff in order to balance the formula. For example, a time in bed of 7 hours gives a weight of 0.984, which is less than a 2% adjustment. Six hours gives a weight of 0.938, a more meaningful adjustment, 5 gives 0.859, 4 gives 0.75, and so on to 0 hours giving a weight of 0. For any amount of time in bed at or greater than 8, W_TiB is 1. This formula is represented graphically in Figure 1.

1.png

Figure 1. Visualization of the effect of time in bed on the selected weight (W_TiB)

By incorporating a time-in-bed adjustment, our method ensures that the sleep quality index more accurately reflects the total sleep duration, an essential factor in evaluating sleep health. The penalty applied for insufficient sleep hours helps highlight the importance of adequate sleep duration, which is particularly relevant for older adults who may struggle to achieve the recommended 7-8 hours of sleep.

W_TiB is one of two weights that are multiplied by the restlessness percentage in order to create our sleep quality index. The second weight is calculated by using 'sleep normality,' a metric that determines how typical a resident's sleep looks compared to their prior sleep data. Detailed information on this metric is described in the sleep normality section. The sleep normality weight W_SN is computed by taking the mean of a resident's sleep normality scores from the previous two weeks and comparing it to the current date's sleep normality as described in Eq. (5).

$W_{S N}=\left\{\begin{array}{c}\frac{S N-S N_{\text {Baseline }}}{S N_{\text {Baseline }}}, \text { if exist }\left(\mathrm{SN} \text { and } \mathrm{SN}_{\text {Baseline }}\right) \\ 1, \text { otherwise }\end{array}\right\} f(x)$          (5)

where, SN is sleep normality, and SN_Baseline is the mean sleep normality from the previous two weeks.

The inclusion of sleep normality as a metric allows the SQI to account for variations in sleep patterns over time. By comparing the current night's sleep to a baseline established over two weeks, our method can detect deviations that might indicate emerging health issues. This dynamic adjustment helps ensure that the sleep quality index remains sensitive to changes in sleep behavior, which could be crucial for early detection of health deterioration.

Nights that appear more typical will give a weight greater than 1, while nights that look more unusual give a weight from 0 to 1. Since the typical range of values for sleep normality in the test cases we examined is about 60 to 90 out of 100, this weight is typically no more than a 15% adjustment to SQI_TiB. In other words, the range of values for W_SN was found to be between 0.85 and 1.15.

The combination of both weights with the restlessness percentage gives us our new SQI_Final metric, as shown in Eq. (6). This is the metric that is used in the experimental results section.

$S Q I_{F i n a l}=W_{S N} * \mathrm{SQI}_{T i B}$          (6)

The final sleep quality index, SQI_Final, integrates both time in bed and sleep normality adjustments, offering a comprehensive measure of sleep health that is tailored to the specific needs of older adults. This approach not only captures the quantity and quality of sleep but also considers how consistent sleep patterns are over time, making it a powerful tool for long-term monitoring in eldercare settings.

In this section, we introduce the dataset used to evaluate our approach in computing the sleep quality for six older adults from our eldercare monitoring system. We use bed sensor data, which captures the resident's heart rate, respiration, restlessness, and time in bed (TiB). Our goal is to develop a sleep quality index (SQI) that accurately reflects the quality of the resident's sleep in terms of the number of sleeping hours, amount of restlessness, and the normality of the sleep compared to a baseline.

To investigate the effectiveness of the proposed SQI and sleep normality metric, we pull multiple case studies for the residents in Table 1. Each case study has a time span of about 2 weeks and shows a general trend in our proposed metrics.

Table 1. Dataset description

Figure 3a shows the first case study for resident I in Table 1. On November 21, 2019, this resident had a fall incident. It is expected that the sleep quality degrades prior to a fall. We can see in the figure the previous 14 days leading to the incident date, which shows a clear decreasing trend.  The trend slope is very sharp the week before the fall. This case study shows the effectiveness of the proposed index in correlating the reduction in the SQI and the possibility of fall. This could be used by our monitoring system to generate an alert any time there is a decreasing trend or reduction in the SQI, which can help in detecting early signs of illness.

3a.png

3b.png

3c.png

3d.png

Figure 3. SQI for: a) resident I, b) resident VI, c) resident II, and d) resident IV

Figure 3b shows the second case study for resident VI in Table 1. Around August 06, 2019, the resident had to be taken to the ER. Also, the resident had lower back pain on that day. The SQI values are computed for 14 days leading to the event to check if there is a trend or noticeable change in the sleep quality. We can see that there is a clear decreasing trend captured by the SQI.  During these 14 days the resident also experienced multiple health conditions. On July 24, 2019, the resident experienced knee pain which might explain the reduction in SQI value on the following day. The resident was supplied with painkiller medications afterward, which might explain the improvement in the SQI for the next three days. It was reported in the notes that the resident was sleeping well on the nights of 27-28 of July, which showed up in the SQI value for July 27th.  The resident had a wound on July 30th and severe pain, which could help explain the reduction in SQI on that day leading to the night of the 31st. Then, a new medication for severe pain was ordered and used on the day of the 31st. After that, the resident was in bed and did not want to eat for some of the remaining days. Also, the resident was confused on the remaining days and was found on the floor multiple times which explains the poor SQI for the days leading to the ER visit.

Figure 3c depicts data for Resident II from April 18 to April 30, 2020. For about a week starting on April 18, we can see that the SQI dropped from around 0.7-0.75 down to around 0.6. On April 21, a note read, "Resident is c/o severe pain and is needing assistance getting in and out of the bed." The note suggests that the resident could be having difficulty sleeping and the drop in the SQI reflects this. A few days later, on April 27 a note read, "[Resident] stated [they] did feel somewhat better but still has quite a bit of pain in the thigh area of [their] right leg." The increase that we see in the sleep quality index around this time reflects the improvement of the resident's health as they continue to recover.

The other case study shows sleep normality data for Resident IV in Figure 3d. From October 31 to November 2, 2019, the resident had a significant drop in their sleep normality which coincides with EHR notes that give context to the chart. On November 2, the day with a local low for sleep normality at just over 40, an EHR note reads, "Resident running a fever this am 101.4 gave [them] some ibuprofen started [them] on [their] antibiotic fever broke after giving ibuprofen." Along with this note, the resident's bed sensors gave readings for RST, HR, and RSP which are highly unusual when compared to the previous 2 months of data for the resident. Later, on November 8, an EHR note read, "Resident's temp is 95.4 [They are] feeling better." At this point, the sleep normality has recovered to a stable level that seems to line up well with what we would expect from the EHR note.

During a period of flu-like symptoms where resident III experienced a noticeable drop in their sleep quality index (SQI), falling from the baseline average of 0.61 to around 0.5. According to an EHR note, the resident was "having a temp during night and cough worse. I requested nurse make SNV to assess him/her. He/she is up and dressed. States she/he had to sleep on three pillows last night to keep from coughing and had temperature. States had several coughing episodes that just couldn't stop coughing. L lung clear but has rhonchi and slight wheeze throughout R lung fields today which had been clear. SN called Dr. X office to try to get an appointment with Someone there today." After about a week of these symptoms, the resident's health began to improve, and a subsequent note stated, "Pt states he/she is feeling better." Correspondingly, the SQI gradually returned to the baseline level, reflecting the recovery of their sleep quality.

In another instance, Resident V experienced a health issue where they "came to the nurse station wanting something for cough. Slept in chair because every time he/she laid down, he/she would have a coughing fit. I called Nurse and she said he/she can go to urgent care if he/she likes." Unlike Resident III, whose similar symptoms resulted in a clear drop in their sleep quality index (SQI), Resident V's SQI remained unaffected because the resident slept in a chair rather than in bed, where our sensors are placed. This case highlights a limitation of our current approach, which relies solely on bed sensors to monitor sleep quality. When residents sleep outside of bed, the system cannot capture changes in sleep quality, leading to gaps in the data and potentially missing important health events.

As we can see, the proposed SQI is effective in capturing a variety of health issues that lead to a drop in sleep quality. However, a key limitation of this approach is its dependency on the resident being in bed. Since we do not use wearable sensors for continuous monitoring, any time spent outside of bed goes unrecorded. To enhance the robustness of our approach, integrating additional sensors, such as radar, depth imaging, or even passive infrared sensors, could allow for more comprehensive monitoring. These technologies would enable us to track the resident's activity and sleep patterns even when they are not confined to their bed, thereby improving the accuracy and reliability of the SQI in various health scenarios.

The sleep quality index (SQI) developed in this study offers several potential applications in clinical practice, particularly within the context of eldercare monitoring and early intervention. Its ability to provide continuous, non-intrusive monitoring of sleep quality makes it a valuable tool for enhancing patient care and detecting early signs of health deterioration.

One significant application of the SQI is its integration into existing eldercare monitoring systems. By continuously tracking changes in sleep quality, the SQI can serve as an early warning tool for healthcare providers. This capability allows clinicians to identify residents who may be experiencing health decline, enabling timely interventions before health issues become critical. The early detection of such changes can lead to more proactive care and potentially prevent the escalation of health problems.

Furthermore, the SQI's ability to facilitate remote monitoring presents substantial opportunities for telehealth applications. Clinicians can utilize the SQI to monitor patients from a distance, thereby reducing the need for frequent in-person visits while still ensuring that any concerning trends in sleep quality are promptly addressed. This remote capability is especially valuable in managing the health of older adults, where regular in-person check-ups might be challenging.

In addition to monitoring, the SQI can play a crucial role in preventative care and early intervention strategies. By incorporating the SQI into routine care protocols, healthcare providers can enhance the early detection of health issues, such as sleep disorders or the initial stages of cognitive decline. The continuous monitoring offered by the SQI allows for early identification of sleep disruptions, enabling healthcare providers to initiate interventions at the earliest signs of potential health problems, ultimately improving patient outcomes and quality of life.