Digital phenotyping of sleep patterns among heterogenous samples of Latinx adults using unsupervised learning

1. Background

Sleep health is recognized as a public health concern in the United States [1]. Inadequate sleep duration and poor sleep quality are associated with incident obesity, impaired physical function, hypertension, and cancer[2, 3]. Risk of poor sleep health is higher among certain populations, such as individuals with chronic conditions associated with circadian rhythm disturbance (e.g., HIV)[4, 5]. In addition, stigmatized populations, such as racial, ethnic, or sexual and gender minority (SGM) adults, may experience poor sleep health due to chronic exposure to social stressors (e.g., perceived or anticipated rejection, discrimination)[6–9]. Such physiological and psychological factors reportedly contribute to a disproportionately higher prevalence of sleep problems in these individuals[6, 9]. However, there is scarce research using objectively-estimated (e.g., accelerometer-based) data to investigate sleep health patterns and potential between-individual heterogeneity that could be indicative of sleep disturbance subtypes (i.e., “phenotypes”). Identification of possible sleep phenotypes constitutes a starting point for developing individualized prevention and treatment strategies, providing precedence for undertaking this investigation.

An increased prevalence of sleep disturbance across various groups (e.g., older adults, transgender individuals), and during life transitions (e.g., menopause), as well as those with chronic co-morbidities (e.g., HIV, disorders) is well-documented.[10–15] However, existing studies evaluating sleep disturbances in minority populations have been mostly limited to self-report [15, 16], which has low concordance with objectively-estimated sleep measures[17], and detection of sleep pathologies (e.g., sleep apnea) in samples with a formal diagnosis of sleep disorders [18]. However, investigation of sleep phenotypes based on sleep disturbance among heterogenous, at-risk individuals who have not yet received a formal diagnosis can aid in detection during earlier stages of their sleep disturbance. Such phenotyping investigations can further help elucidate the degree to which treatments or interventions for sleep problems should be tailored. For example, should we identify that sleep health patterns vary within an at-risk population but clusters of patients with similar patterns can be identified, tailoring screening and intervening approaches at the sub-population level might be suitable.

The lack of research in this area might in part be due to the complex structure of objectively-estimated sleep data and the limited applicability of traditional clustering methods for such data.[19] Most traditional clustering techniques (e.g., k-means) cannot accommodate longitudinal data that have complex covariance structures with correlated measurements across time points, and vary in type (e.g., continuous, binary, multinomial) and time intervals of measurement.[20] They require certain assumptions to be met in the structure of the data, therefore they are sensitive to outliers, and can fail to identify the optimum structure of the clusters.[21] As the data size increases, exhausting all partitioning possibilities becomes computationally prohibitive.[22] Importantly, certain clustering methods assume that each data point (e.g., participant) belongs exclusively to a single latent group,[19, 23] which is an unrealistic assumption in most scenarios involving patient-oriented research.[23, 24] In such instances, a probabilistic approach that quantifies the likelihood of a participant belonging to any of the clusters present in the dataset by considering the entire probability distribution of membership assignments is more suitable. This allows assessment of the relative probabilities, and therefore degree of confidence with which the assignment is made.[25, 26] Flexible and dynamic methods that simultaneously consider uncertainty whilst delineating variability in sleep patterns are thus needed to improve applicability in real-life settings.

Accordingly, this study investigates the question, “Are there homogeneous sub-groups of individuals with distinct sleep patterns at risk for worsened sleep outcomes but heterogeneous with respect to age, co-morbidities, lifestyle, and self-reported health?” To address the challenges related to digital sleep data described earlier, we implement an unsupervised machine learning technique that can accommodate mixed membership structures within the data.[25, 27] This approach relies on the assumption that data come from several subpopulations and that the overall population is a mixture of these subpopulations, and further allows multivariate response data. This provides an ideal phenotype analysis framework to meet the objectives of the present study.

2. Methods

This study is an analysis of objectively-estimated sleep data collected to identify determinants of sleep disturbances in Latinx adults as a part of the Precision in Symptom Self-Management (PriSSM) Center (P30NR016587) at Columbia University. The PriSSM Center aims to advance the science of symptom self-management among the Latinx population through a social ecological lens that accounts for variability in individual, interpersonal, organizational, and environmental factors across the life course. All procedures were approved by the institutional review boards of the Columbia University Irving Medical Center and the New York State Psychiatric Institute. Informed consent was obtained from all participants.

3. Results

4. Discussion

This study investigated possible subgroups (“phenotypes”) based on objectively-estimated sleep parameters among individuals at increased risk for sleep disturbance. Using an unsupervised learning approach that accommodates inherent data complexities and sample heterogeneity, we identified three distinct sleep phenotypes that were best estimated through TIB, SE, and bed time. To our knowledge, this study is the first to investigate digital sleep phenotyping leveraging accelerometry and mixture models an unsupervised learning approach.

Our analyses indicated three distinct clusters (i.e., mixtures) to be the most suitable based on model fit, which were not independently associated with age or weight. Phenotype 1 constituted the largest cluster (N=64, 56% of the sample) and was associated with significantly greater likelihood of overall adequate SE and less variability in SE and WASO. Later bed times and shorter TST were also associated with this cluster, which were statistically significant in adjusted t-tests (See Table 4). Phenotype 2 constituted the smallest cluster (N=11) and had the highest likelihood of substantial and consistent sleep disturbance characterized by lower SE, more frequency awakenings NOAs, and greater WASO and TIB than the other phenotypes. The median age associated with this phenotype was also older than those of others though it was only significant when compared to phenotype 1 (i.e., 65.0 vs 56.0 years). Phenotype 3 (N=43) was characterized by a more variable sleep profile, with lower SE compared to phenotype 1, but also lower sleep disturbance compared to phenotype 2 as determined by fewer NOAs (t=−2.52, p=0.018) and shorter WASO (t=−4.41, p<0.001). Self-reported sleep disturbance scores across the three phenotypes were comparable and within ½ SD of the population-normalized mean. This suggests low concordance between scores from self-reported vs accelerometer-based sleep measures. This is consistent with previous studies,[17, 46] and supports the need for phenotyping based on multiple variables rather than using a single score when identifying individuals at risk for sleep problems.

Previous studies have reported shorter sleep duration with increasing age.[47] However, in this study phenotype 1 was associated with shorter TST but also significantly younger age compared to phenotype 2. While this finding might appear discrepant, our results indicated a trend where phenotype 1 was also more likely to engage in daytime sleep episodes (p=0.046). Combined with the relatively greater variability in TST linked to phenotype 1, it is possible that these characteristics collectively point to a sleep profile that is flexible in nature. These “short-but-efficient” sleepers might be better at adapting their sleep behaviors as needed according to changes in their daily schedules. While the lack of contextual information on these daytime sleep periods hinders a more comprehensive interpretation of daytime sleep episodes,[48] we note their significantly longer duration (i.e., ~2.5 hours on average) than what has been recommended and considered as beneficial.[49] Prior evidence on daytime sleep architecture is mixed,[50] with some studies indicating greater daytime sleep episodes are associated with increasing age,[51] while others report age-related changes in the timing but not the duration of daytime sleep episodes.[29, 47] There were 13 daytime sleep episodes from 9 participants in total, suggesting that these individuals might not necessarily be habitual daytime nappers based on the existing definitions of habitual napping (e.g., every day or at least twice/week).[49] It is possible that the individuals in phenotype 2 have difficulty initiating sleep onset in general or have daily schedules that are not amenable to daytime naps or longer sleep periods in general, which could explain their overall shorter sleep durations. Ultimately, daytime sleep architecture differs from that of the main nocturnal sleep period and their implications are still under debate[48, 52–54]. Our findings warrant further investigation into how these daytime sleep episodes might be related to different sleep phenotypes.

Our results indicated a trend (p-adjusted=0.07) where phenotype 3 was characterized by greater variability in both bedtimes and waketimes when using hourly intervals (i.e., p-adjusted<0.0001 for both pairwise comparisons). Regularity in bed times and waketimes has been suggested to promote adequate sleep by synchronizing sleep drive and circadian rhythm.[55] Similarly, greater variability in bed and wake time has been associated with increased likelihood of insomnia.[56] This is somewhat interesting given that this phenotype was not associated with the lowest sleep quality or older age. On the other hand, phenotype 2 was characterized by shorter but less variable TST, lower SE, and increased WASO, as well as earlier bed times and older age. Variability in bed times and waketimes has been found to decrease with increasing age, [57] therefore it is possible that these older participants had well-established sleep schedules. However, future studies with large sample sizes and longer data collection periods from each individual are needed to delineate whether the within-person variability associated with these phenotypes is stable over time.

SE has been recommended as a valid and robust indicator of sleep quality,[48] and its patterns have been studied by others across the life span.[30] Increased prevalence of sleep disturbance among women, SGM adults, and older adults have been well-documented in the literature, though mostly limited to data from population-based self-reported data.[8, 11, 58, 59] For instance, Zheng et al.[12] reported naturally occurring cyclical fluctuations in SE in menstruating midlife women (mean age=51.5). The identified profile characteristics of phenotypes 2 and 3 are consistent with these previous findings with respect to increased likelihood of variable SE, TIB, bed time and awakening time, as well as female sex. However, we also identified variable TST associated with phenotype 1 and similar age ranges compared to phenotype 2 (i.e., 19–77 for phenotype 1 vs 24–77 phenotype 2). To our knowledge, this study is the first to quantify these phenotypic characteristics using objective sleep data within a sample comprising ages 19–77, 32% transgender individuals, and 22% older adults with HIV. That we were able to identify these sleep phenotypes described in the literature in a mixed sample of individuals relying on ~4 days of data underscores the promise of unsupervised machine learning techniques to better understand sleep health phenotypes with relatively minimal participant burden.

Although validating the potential prognostic value of these phenotypes is beyond the scope of this study, these preliminary findings have possible implications from a precision sleep health perspective. For example, phenotype 3 may represent an “at risk” category where someone with these characteristics could be in the early stages of chronic sleep disturbance but perhaps not as severe as someone of phenotype 2. Individuals with different sleep phenotypes would likely benefit from different sleep health interventions (e.g., increasing TST for phenotype 2 individuals vs. establishing consistent a sleep schedule for phenotype 3). There are currently no established guidelines on how such sleep health prescriptions can or should be individually tailored in clinical settings, or which sleep parameters constitute optimal intervention targets. Moreover, the lack of established sleep parameter cut-off values for determination of sleep disturbance [44, 45, 60] could partly be because using a single cut-off across all patient populations is unsuitable. Instead, multiple cut-off values depending on patient phenotypes might be a more suitable approach. Similarly, integration of variability metrics and composite scores derived from frequently-sampled and/or longitudinal data are unique advantages of not only research-oriented devices such as that used in these studies, but increasingly of consumer devices such as fitness trackers and mobile phone apps (mHealth)that can augment clinical data.[61, 62] To this end, there is great promise in interdisciplinary efforts that employ mHealth devices for data collection and machine learning techniques for optimizing and designing scalable sleep solutions.[63]

Inclusion of participants at greater risk for sleep problems across the adult life span [6, 7, 64] further expands the existing literature on sleep health parameters within marginalized populations.[65, 66] For example, despite evidence that Latinx SGM adults report a higher prevalence of short sleep duration (≤6 hours per night), evidence on sleep health in SGM adults is limited to largely White samples. Similarly, cohort studies with racial/ethnic minorities and HIV-positive patients indicate that sleep disturbances and apnea are common[67], yet largely undiagnosed and untreated [5]. A previous study reported that despite having poorer self-reported sleep quality, Black women were less likely than other women to report their sleep disturbances to a physician [68]. Evidence indicates SGM patients face significant discrimination within healthcare settings,[69, 70] which could raise hesitancy to seeking counseling for sleep issues. Passively-collected and sleep data might therefore be particularly useful as it circumvents this challenge, as well as various others including inadequate clinician documentation of patient-reported sleep problems,[71] and providers’ implicit and explicit attitudes in clinical settings [72] which can impact clinical decision-making.[73] These collectively underscore the advantages of this low-effort, low-cost, and time-efficient method for delineating sleep problems among patient populations who might be harder to engage in clinical care.

There are several methodological strengths of our study. First, MCMC-based MMGLMMs consider each participant’s probability of membership to all existing clusters in the dataset to decide on the optimal classification, based on the maximal value of the posterior component probabilities. Next, they can accommodate sample heterogeneity and do not require classical normality in the random effects distribution. Instead, they assume a normal mixture and captures any possible dependence across values of different biomarkers through a joint distribution. Thus MMGLMMs improve clustering quality when using data with repeated measures, and have been demonstrated to be superior to averaging over the repeated measurements for producing relatively more stable and accurate clusters, and increase robustness against noise [74].

5. Limitations

We acknowledge limitations of this study. First, we had a relatively modest sample size of Latinx adults with more female than male participants. Therefore, generalizability of some of these findings might be limited and we cannot ascertain if the number and size of clusters, and cluster-relevant predictors might differ if the sample included other races and ethnicities. Similarly, temporal stability of the identified clusters across time is another interesting question this study did not investigate. Next, accelerometry cannot detect sleep apnea, which can be common in HIV, and contributes to poor sleep quality. The relationship of non-nocturnal sleep to overall sleep health remains an open research question [48], and whether they provide clinically relevant information on phenotype membership warrants further investigation. Our study sample included a wide age range cross the lifespan, which allows evaluation of variability in bed time and wake time in relation to key transition periods (e.g., young adulthood).[57] However, five participants only had 2 days of data, which limits our ability to interpret within-person variability outcomes, as well as to compare differences between week days vs weekend days. Though the median number of available days was 4 for all 3 clusters, it is possible that we might not have had sufficient days of data to capture the true variability for some of the participants. We similarly did not measure chronotype (i.e., morningness-eveningness preference), which warrants further investigation. An interesting future direction would be to conduct a phenotyping analysis that assesses the stability of the chronotypes associated with each phenotype.

6. Conclusion

In conclusion, digital data-driven approaches integrating robust data analytic techniques can aid in early detection and management of sleep problems. Our limited understanding of how sleep patterns unfold among different patient populations hinders the development of tailored interventions (e.g., behavioral intervention for promoting patient engagement in their care to attenuate the negative effects of discrimination-related stress on sleep health). These results have implications for informing individually-tailored strategies and precision sleep health applications across diverse patient populations.

Supplementary Material