Profiling of Sleep Models Based on Voluntary and Involuntary Sleep in Adults with Type 2 Diabetes

The purpose of this study is to generate the concept of voluntary sleep (V) and involuntary sleep (IV) in sleep, to build sleep models using them, and to profile by diabetes control in adults with type 2 diabetes. We obtained 595 nights of sleep data from 50 participants. Participants measured sleep with the sleep meter HSL-101 (Omron Healthcare, Kyoto) and answered the Pittsburgh Sleep Quality Index (PSQI). They were operationally defined as V and IV. "V1: sleep self-determination" and "V2: conscious sleep quantity" were generated from voluntary sleep, and "IV1: continuous deep sleep" and "IV2: actual sleep quantity" were generated from involuntary sleep. Using cluster analysis, they were classified into three models, "CL1: sleep satisfaction model", "CL2: short sleep model", and "CL3: dissatisfaction sleep model". When the diabetes controls in each cluster were compared by ANOVA and Bonferroni's test, HbA1c was higher in the order of CL1, CL2, and CL3, and there was a significant difference between CL1 and CL3 (p = 0.029). Similarly, age was low in the same order and BMI was high in the same order, with a significant difference between CL1 and CL3 (p = 0.030, 0.037). Sleep in adults with type 2 diabetes could be clustered into three models based on V and IV, each of which could be profiled by a significantly characteristic diabetes control index. It will be possible to identify the patient's sleep model from the diabetes control and appropriately approach voluntary sleep and involuntary sleep.


Introduction
Sleep disorders are a serious problem in today's 24-h society. Inadequate sleep can have a major impact on our society and is regarded as a new public health problem owing to its widespread effects [1,2]. Too little or too much sleep is associated with significant causes of death, such as cardiovascular disease [3,4], cerebrovascular disease [5,6], diabetes [7,8], and hypertension [3,9,10]. In addition, frequent episodes of inadequate sleep are associated with decreased mental and physical well-being [11]. These are captured by the concept of sleep quantity. In addition to quantity, quality is another way that sleep is perceived. The quality of sleep decreases with age [12,13], and times of inadequate sleep are associated with worse metabolism and decreased activity [14,15]. Indeed, sleep quantity and quality are composite elements that characterize sleep, and both need to be considered when predicting health outcomes [16].
There is also the concept of regularity in sleep. The circadian clock has been established to be a pervasive biological mechanism, and an irregular sleep schedule is the most common cause of circadian clock disruption [17]. Inconsistent and irregular sleep duration and timing increase the rates of obesity, hypertension, and dyslipidemia [18][19][20][21]. Although studies have suggested that poor sleep can impair health in a variety of aspects, most of these researches were based on individual perceptions; these studies treated sleep subjectively, but there are other ways of perceiving sleep objectively [22]. Subjective sleep is primarily evaluated through self-reports, whereas objective sleep is examined by measurements, such as polysomnography and actigraphs [23].

3
In this study, we used a sleep meter and a self-administered questionnaire to measure sleep in adults with type 2 diabetes. The purpose of this study was to define sleep that can be controlled by oneself as voluntary sleep (V) and sleep that is difficult to control by oneself as involuntary sleep (IV), to construct sleep models using these concepts, and to understand the relationships of these types of sleep with diabetic control in adults with type 2 diabetes.

Participants
We asked 91 adults with type 2 diabetes and who went to three hospitals in a prefecture to participate in measurement of sleep using a sleep meter and to answer the Pittsburgh sleep quality index (PSQI), which is a self-evaluation of sleep quality. A total of 50 subjects (38 men, 12 women) agreed to participate in the study, and a total of 595 days of sleep were analyzed. No participant had fewer than three days of measurement.

Ethical Considerations
This study was approved by the local ethics review committee (approval number: 742-1). The researcher explained to the subjects verbally and in writing that participation in the research was voluntary that there would be no disadvantage in refusing to participate, and that personal information was strictly protected. Written informed consent was obtained from all participants.

Objective Sleep Evaluation
Participants were asked to use a sleep sensor (HSL-101; Omron, Kyoto, Japan) to self-monitor their sleep for 14 nights. The sleep meter was placed next to the bed, and measurements were performed without restraining, attaching to, or contact with the participant. The participants were asked to press a button on the sleep meter when they went to bed and upon waking.
The following sleep parameters were automatically calculated using the sleep meter: bedtime (BT), total sleep time (TST), deep sleep time (DST), sleep onset latency (SOL), and wake after sleep onset (WASO). In addition, sleep efficiency (SE) was calculated as (TST/BT) × 100. Bedtimes and wake-up times were obtained from the sleep graph to calculate the bedtime standard deviation (BSD) and wake-up time standard deviation (WSD). The absolute value of the difference between self-reported sleep time and TST was designated as absTST. The absolute value of the difference between self-reported sleep efficiency and SE was designated as absSE.

Subjective Sleep Evaluation
The Japanese version of the PSQI was used to assess subjective sleep [24][25][26]. The PSQI comprises seven components, as follows: C1, sleep quality; C2, sleep latency; C3, sleep duration; C4, habitual SE; C5, sleep disturbance; C6, use of medication for sleep; and C7, daytime dysfunction. The scores for these components were added together to obtain the global score. The cutoff was 6 points, and scores ≥ 6 represented poor sleep quality. Subjective bedtime (sBT), subjective TST (sTST), and subjective SE (sSE) were extracted from the C3 and C4 scores.

Steps in Establishing Sleep Models
Adults with type 2 diabetes were classified according to their sleep types, and a three-step analysis was performed to characterize their diabetic control. As the first step, sleep elements were operationally divided into V and IV elements V was a subjective sleep item that can be consciously controlled, and its components were BSD, WSD, absTST, absSE, sBT, sTST, and sSE. IV was an objective sleep item that was difficult to control consciously, and its components were BT, TST, DST, SOL, WASO, and SE. Bartlett's sphericity test validated the principal component analysis, and a principal component analysis was performed for each. As the second step, a k-means cluster analysis was performed using the integrated principal component score and the PSQI score as the variables. The k-means method is a clustering algorithm that specifies the number of clusters and collects and classifies data with similar attributes. In this study, we increased the number of clusters in order from three and continued the analysis until the composition of one or more clusters became homogenous and interpretable. Thereafter, we obtained the sleep models. Finally, we examined whether there was a difference in diabetes control among the clustered sleep models and performed profiling. Analysis of variance and Bonferroni's test were performed to compare sleep models. Table 1 shows the participants' basic attributes and the values for each measurement. The mean total PSQI score was 8.06 ± 4.48, which was significantly higher than the cutoff point (p = 0.002).

Voluntary and Involuntary Sleep Models
On principal component analysis, sBT was excluded from V and BT was excluded from IV. V included six variables (i.e., BSD, WSD, sTST, sSE, absTST, and absSE) and IV included five variables (TST, DST, SOL, WASO, and SE).
The results for V are shown in Table 2. Kaiser-Meyer-Olkin (KMO) sampling adequacy was 0.708 and Bartlett's test of sphericity was significant (p < 0.001), demonstrating the validity of the principal component analysis. The principal components were calculated until the eigenvalue was ≥ 1. The cumulative contribution ratio was 71.412%. The main variables of the first principal component were WSD, BSD, absTST, and absSE, with principal component loadings of 0.879, 0.778, 0.752, and 0.738, respectively. These components indicated the timing of going to bed, waking, and recognizing sleep. Therefore, the first principal component was named V1 (i.e., sleep selfdetermination), which accounted for 45.605% of the total variance. The main variables of the second principal component were sTST and sSE, with principal component loadings of 0.828 and 0.737, respectively. These indicated the length and efficiency of sleep. Therefore, this second principal component was named V2 (i.e., conscious sleep quantity).
The results for IV are shown in Table 3. The KMO sampling adequacy was 0.493 and Bartlett's test of sphericity was significant (p < 0.000). The principal components were calculated until the eigenvalue was ≥ 1. The cumulative contribution ratio was 78.782%. The main variables of the first principal component were SE, WASO, and DST, with principal component loadings of 0.946, 0.825, and 0.748, respectively. These components were difficult to consciously control and included SE, depth, and persistence. Therefore, this first principal component was named IV1 (i.e., continuous deep sleep), which accounted for 53.333% of the total variance. The main variables of the second principal component

Clustering by Sleep Indicators
A k-means cluster analysis was performed based on PSQI, V1, V2, IV1, and IV2, which were indicators of sleep. The number of clusters was increased from three, and analysis continued until the composition of one or more clusters became interpretable. As a result, the best fit was classified into three clusters. These clusters were significantly different in terms of PSQI, V1, and V2 ( Table 4). The configuration of each cluster is shown in a radar chart (Fig. 1). To note the condition featuring the best sleep, sleep quality was set to the maximum PSQI score of 21 at the edge of the plot, and V1 values were multiplied by − 1.
Cluster 1 (n = 17) had good sleep quality (PSQI 4.24 ± 1.35) and tended to be good in all components of V and IV. Therefore, it was named CL1 (i.e., satisfactory sleep model). Cluster 2 (n = 23) had poor sleep quality (PSQI 8.48 ± 1.56) and was as good as cluster 1 in terms of V1 and IV1 but had low levels of sleep quantity, which were common with V2 and IV2. Therefore, it was named CL2 (i.e., poor sleep quantity model). Cluster 3 (n = 8) had extremely poor sleep quality (PSQI 16.25 ± 2.92), with V1 and IV1 being particularly poor. These patients had poor sleep quality, with no sleep satisfaction in either V or IV. Therefore, it was named CL3 (i.e., insufficient sleep quality model).
In summary, CL1 was characterized by good HbA1c, older age, and low BMI; CL2 was characterized by moderate HbA1c, age, and BMI; and CL3 was characterized by high HbA1c, young age, and high BMI.

Discussion
Sleep disorders, such as reduced sleep quantity and quality, have been reported in most patients with type 2 diabetes [7,8,27]. One of authors of this study ( [28] and unpublished results) revealed that differences in sleep perception and actuality in diabetics and variability in wake-up and bedtime affect HbA1c. These findings indicate that conscious sleep and unconscious sleep might affect diabetes control. In this study, the sleep structure of adults with type 2 diabetes was clarified by dividing it into V and IV. We found three sleep models by clustering the structure as an index and profiled the models according to diabetes control. There were three unique points in this study. First, sleep in adults with type 2 diabetes was classified as V or IV. Evidence-based sleep management skills should be handled by community health professionals, such as sleep specialists who develop care models and primary care providers who implement sleep health programs [29,30]. By doing so, patients reported a reduction in anxiety over sleep problems, feeling better in managing their sleep, and having a sense of control over their sleep [29]. The reason for classifying sleep as V or IV was to clarify the necessity of intervention. Because V can be adjusted consciously, it is expected to be improved by improving the sleep management ability of the patient. IV is beyond the conscious reach of individuals and requires appropriate intervention by medical professionals. Clarifying the responsibility for sleep management enables focused management. Sleep control may be promoted by leaving as much control as possible to the patients themselves.
In this study, V was divided into orthogonal components of V1 (i.e., sleep self-determination) and V2 (i.e., conscious sleep quantity) by principal component analysis. V1 and V2 were thought to represent the voluntary elements of sleep quality and quantity, respectively. Similarly, IV was divided into orthogonal components of IV1 (i.e., continuous deep sleep) and IV2 (i.e., actual sleep quantity), and IV1 and IV2 represented the involuntary elements of sleep quality and quantity, respectively. The results of this study suggested that the concepts of V and IV have both qualitative and quantitative aspects and that these can be applied to evaluate a wide range of sleep elements.
Clustering of clinical phenotypes and symptom patterns of diseases have been widely used [31]. Such profiling may ultimately help to better understand the disease and to predict therapeutic response [32]. In this study, the patients with type 2 diabetes were clustered into three models, as follows: CL1 (i.e., satisfactory sleep model), CL2 (i.e., poor sleep quantity model), and CL3 (i.e., insufficient sleep quality model). These were characterized by a balance of V, IV, and self-assessment of sleep. Based on these results, we believed that it is possible to stratify the risk of sleep disorders by classifying sleep conditions into more uniform clusters. This study suggested that using actual V and IV for clustering could identify patients at high risk for diabetes and may provide personalized medical treatment guidance [33]. Our results suggested that maintaining good sleep has beneficial metabolic effects, which may enhance the current prevention strategies that primarily focus on promoting sufficient sleep and other healthy lifestyles to prevent metabolic diseases [17]. Our data provided evidence that the components of V and IV were associated with a favorable metabolic profile for diabetes and suggested possible sleep interventions to promote good sleep.
In addition, cluster profiling was able to explain the three models based on clinical indicators, such as HbA1c, BMI, and age, which are important in type 2 diabetes. HbA1c and BMI are components of diabetes control and had been canonically correlated with sleep ability [28], thereby, strengthening the robustness of CL1. Our results showed that sleep quality was better with lower HbA1c and lower BMI. Therefore, the quality of diabetes control can be inferred by classification of sleep characteristics and construction of a sleep model, and diabetes may be controlled by sleep intervention. In the future, sleep control for the care of adults with type 2 diabetes should focus on voluntary and involuntary movements, determine sleep items that can be intervened on, and modeling and profiling sleep in relation to diabetes. In terms of clinical application, a patient's sleep model may be identified based on the diabetes control status and appropriately apply it to self-correctable sleep and sleep requiring intervention.

Limitation
In this study, we investigated the presence or absence of the common sleep disorder sleep apnea syndrome (SAS), but we did not consider participants with SAS during analysis. In addition, discrimination between V and IV was subjective.
In future studies, a larger sample size may enable statistical classification. Finally, this study was conducted for more than a year, and there may have been seasonal fluctuations in sleep; this is a factor that needs to be considered in future studies.

Conclusions
Using the concept of V and IV and self-evaluation, three models of sleep quality and quantity were created in adults with type 2 diabetes. The participants in the satisfactory sleep model had acceptable diabetes control, whereas those in the insufficient sleep quality model had poor diabetes control. Sleep intervention may be an important part of care for adults with type 2 diabetes.