Identification of sleep phenotypes in COPD using machine learning-based cluster analysis (2024)

Associated Data

Supplementary Materials

Abstract

Keywords: COPD, Sleep disorders, Phenotypes, Comorbidities

1. Introduction

Chronic Obstructive Pulmonary Disease (COPD) is a prevalent and preventable disease with a course that is often complicated by comorbidities, such as sleep disorders. Persistent symptoms of COPD are often associated with poor sleep quality [1,2]. Indeed, up to 78 % of patients with COPD report sleep disturbance which significantly impacts their quality of life [1–3]. Such sleep disturbances include problems initiating or maintaining sleep, increased sleep, reduced rapid eye movement (REM) sleep and frequent sleep stage shifts, sleep fragmentation and micro-arousals [4]. Considering that patients with COPD typically spend a third of their lives sleeping, deviations from healthy sleep patterns in this population may correlate to adverse long-term outcomes. A recent large-scale study investigating the links between sleep quality and COPD exacerbations revealed that poor sleep may indeed be a better predictor of exacerbation than a subject’s smoking history [5]. Furthermore, poor sleep has been associated with a 25 %–95 % increased risk of COPD exacerbation [6].

In the general population, abnormal objective polysomnographic metrics have been linked to poor long-term outcomes. For example, the Sleep Heart Health Study cohort demonstrated that alterations in polysomnography measures, such as total sleep time, were associated with a higher risk of various cardiovascular outcomes, including mortality [7, 8]. Moreover, a population-based cohort study in older men indicated that sleep fragmentation, a lower percentage of REM sleep, and severe reduction in blood oxygen saturation were independently associated with mortality [9]. Consequently, understanding the polysomnographic factors that impact long-term outcomes especially mortality in COPD, is crucial for developing effective management strategies and improving patient outcomes.

Studying big data related to sleep parameters provides valuable insights into the sleep health of patients in a real-world setting. However, existing studies are limited in terms of sensitivity and specificity as they used imprecise cohorting methods that relied strictly on International Classification of Diseases (ICD) coding to determine eligibility. To overcome these limitations, we developed and validated advanced data science modalities, including natural language processing, to enable us to achieve high precision, recall, and F-1 scores (>0.9) for specific objective sleep metrics [10]. This has allowed us to determine a comprehensive set of variables derived from physician report of a polysomnography with a high degree of precision.

To identify distinct clinically important phenotypic clusters of COPD patients based on the sleep parameters, we performed a longitudinal observational study spanning over 20 years among the U.S. Veterans population. Sleep parameters for this study were obtained by incorporating a comprehensive set of polysomnographic variables. Utilizing unsupervised modeling methods, we sought to identify patient clusters in an unbiased manner, without relying on pre-existing knowledge or assumptions. Furthermore, we evaluated the associations of identified phenotypic clusters with all-cause mortality, while accounting for potential confounding factors such as comorbidities, to better understand the prognostic implications of these COPD clusters in the real-world clinical setting. The identification of these sleep-related specific clusters may complement currently available disease assessment tools in improving the overall risk stratification and informing future targeted interventions for this vulnerable patient population.

2. Study Design and Methods

3. Results

Among 4,237,444 users of sleep services, 82,782 had baseline PSGs with CPT codes of 95810 (Fig. 1). Of those, we identified 10,447 patients with a COPD diagnosis within the three years of index PSG. From these, 455 patients were excluded based on the judgment of board-certified sleep specialists (AS and RA) to have very unrealistic high total sleep time with a very low sleep efficiency, e-Fig. 1. Thus, the analysis reported in this work is based on data from 9992 patients with COPD and a baseline PSG.

We identified five clusters (e-Fig. 2). Table 1 shows the demographic characteristics on all the study subjects and in patients within each of the five clusters. The two clusters with the lowest and highest mortality differed from the other three clusters on two significant variables. Cluster 1 included distinctly and significantly younger patients than other clusters. Cluster 5 included patients who had higher comorbidity burden as identified by a markedly higher CCI of 5.5. Further, Cluster 1 had higher proportion of female veterans compared to other clusters. Cluster 1 and 5 had higher proportions of Black veterans compared to the other three clusters. Clusters 2, 3, and 4 had very similar demographic characteristics but distinct sleep parameters.

Table 2 shows the sleep parameters extracted from the physician reports of baseline sleep studies of the participants. Cluster 1 showed the longest TST, highest SE, and shortest SOL and WASO. Compared to Cluster 1, TST and SE diminished and SOL and WASO increased in clusters 2–5. Percent of subjects with AHI of five or higher slightly increased from cluster 1 to five (p-value <0.0001). Similarly, the prevalence of moderate or severe sleep apnea significantly increased moving from Cluster 1 to the other clusters. E-Table 1 provides demographic data comparing the subjects who died during the study follow-up to those who remained alive. Those who died were older, more likely to be White males, had higher proportion of BMI less than 18.5, and had higher comorbidity burden. E-Table 2 provides sleep parameters comparing the subjects who died during the study follow up compared to those who remained alive. Subjects who died had more abnormal sleep parameters including shorter TST and SE. SOL and WASO did not differ between the two groups. A higher percentage of patients who remained alive had TST >300 min (54.8 %) compared to those who died (42.3 %). The same significant trend was observed for SE > 80 % (alive, 44.4 %; dead, 33.6 %), SOL 8–30 min (alive, 30.0 %; dead, 28.5 %). We didn’t observe significant differences in percent of participants with mild, moderate, and severe AHI between the alive and dead groups (p values > 0.05).

Overall mortality (mortality anytime during the study period) and short term (≤5.3 year) mortality increased incrementally and significantly moving from Cluster 1 to Cluster 5. The overall mortality ranged from 9.4 % in Cluster 1–42 % in Cluster 5. The short-term mortality ranged from 3.4 % in Cluster 1–24.3 % in Cluster 5. Fig. 2 depicts the time to death in the five clusters. Time to death diverged consistently during the follow up period. Using Cluster 1 as the reference, adjusted odds ratio (for age, BMI, sex, race and CCI) were 3.42 (2.9–4.0), 4.03 (3.4–4.7), 5.15 (4,3–6.1) and 5.75 (4.8–6.8) for Clusters 2, 3, 4, and 5 respectively (Fig. 2)

To further understand the relationship among the four determinants of longevity (age, CCI, TST and SE), we created three-dimensional graphs. E-Fig. 3A explores the relationship of age, TST and SE in the five clusters. Cluster 1 (green) clearly differs from the other clusters because of younger age while has similar sleep related parameters and comorbidity burden as Cluster 2 (blue). Cluster 2 has similar age to clusters 3 (orange) and 4 (gray) with longevity advantage aligned with more favorable sleep parameters. Cluster 5 (red) shows a more dispersed sleep parameters’ values and does not conform to the other three clusters with similar age distribution. E-Fig. 3B depicts the relationship among the comorbidity burden and sleep parameters as it relates to the clusters. In Cluster 5 (red), with a high comorbidity burden, sleep data is scattered while the clusters 2–4 gather tightly along the sleep parameters.

Fig. 3 displays the contribution of the four main factors including age, CCI, TST and SE on longevity. To enhance the pictorial presentation and the figure’s simplicity, we scaled TST (in minutes) by dividing to 10 and SE dividing by two. We used 100-Age in place of age so the younger age in our figure will be presented by a higher number. CCI maximum in our subjects was 7. Thus, we used the equation of (7-CCI) *10 as a higher number indicating a lower comorbidity burden. With these conversions in mind, the area of each quadrangle indicates the overall longevity and the factors mainly contributing to the longevity in each cluster. Cluster 1 has the largest area and thus the best longevity while the Cluster 5 has the smallest area and hence the lowest longevity. In contrast to Cluster 1, the driver of lower longevity in Cluster 2, 3 and 4 are TST and SE, while for Cluster 5 comorbidity burden plays the main role in lower longevity.

Our data clearly identified two clusters that reflect the major determinants of mortality: age (Cluster 1) and comorbidity burden (Cluster 5). To determine phenotypes based on sleep characteristics, we therefore used the data from clusters 2, 3 and 4 to generate a decision tree. Fig. 4 shows the output of the decision tree using the cluster 2, 3 and 4. The entry point for the phenotypic characterization is total sleep time. The decision tree identified a TST threshold of ≥278 min to create the subtypes. The decision tree used the SE to further identify phenotypes. For example, patients with high TST and high sleep efficiency cluster to patients with COPD and favorable longevity profile. While patients with low TST and low SE have more shortened longevity. Interestingly the group with only reduced TST or reduced SE fit into the middle cluster with longevity profile between clusters 2 and 4.

4. Discussion

Our study represents the first application of unsupervised machine learning to investigate a comprehensive set of objective sleep parameters extracted from physician interpretation notes, age, and comorbidity burden (CCI) to identify distinct COPD phenotypic clusters. By utilizing the largest dataset of clinical polysomnograms performed on patients with a concomitant clinical diagnosis of COPD, our data suggests significant effects of total sleep time and sleep efficiency on mortality differences within this population. Notably, the machine learning-derived clusters aligned with clinical intuition, thereby supporting the validity of our methodology.

We observed that the cluster with the lowest risk of mortality encompassed the youngest patients with the least burden of comorbidities (Cluster 1) and better sleep parameters, while the cluster with the highest risk of mortality included older patients with the highest comorbidity burden (Cluster 5). Although the findings in these two clusters were expected, the remaining three clusters (2, 3, and 4) demonstrated that sleep parameters were useful in determining mortality risk (both short term and overall mortality). These three clusters likely represent most patients encountered in clinical practice, making them particularly informative in a clinical context. Both sleep quantity, as measured by total sleep time (TST), and sleep continuity measure, represented by sleep efficiency (SE), predicted short-term and overall mortality in this cohort of patients with COPD. Specifically, our decision tree analysis revealed that a total sleep time cutoff of ≥278 min (normal recommendations of 420–540 min of total sleep time per night) and a sleep efficiency cutoff of ≥43 % were crucial variables differentiating between these three clusters. Our findings suggest that these vital sleep parameters may have significant clinical implications and contribute further to the phenotypic characterization of COPD.

In our study, we highlight the significant influence of sleep quality and duration on the longevity of patients with COPD. Our decision tree analysis demonstrated that patients who slept for over 278 min experienced improved longevity. Consistently, another study from the Women’s Health Initiative indicated that shorter total sleep time (TST) predicts reduced longevity [19]. However, the relationship between inadequate sleep duration and COPD is bidirectional, as each condition may exacerbate the other. Patients with COPD frequently experience symptoms such as coughing, wheezing, and dyspnea, which can hinder their ability to fall asleep or maintain sleep. Moreover, many patients suffer from anxiety and/or depression and often use medications with stimulating effects, all of which are known risk factors for decreased sleep duration [20,21]. Sleep continuity measure, on the other hand, is evaluated by sleep efficiency (SE), which measures the proportion of time spent in bed that is dedicated to sleep. For adults undergoing attended sleep studies, normal sleep efficiency typically falls within the range of 80–90 % [22]. In line with this, our findings suggest that poor sleep continuity measure, as indicated by reduced SE of <43 %, is associated with an increased risk of mortality.

There are other factors that contribute to poor sleep continuity measure in COPD patients. Nocturnal hypoxemia is seen in 27–70 % of patients [23]. Nocturnal hypoxemia can disrupt sleep architecture, leading to reduced total sleep time, sleep efficiency, and REM sleep, as well as increased arousal index and sleep fragmentation [24]. Furthermore, restless leg syndrome (RLS) in patients with COPD, with rates up to 36 % [25]. RLS can contribute to sleep disruption and reduced sleep quality. COPD and obstructive sleep apnea (OSA) can often coexist in a phenomenon known as “overlap syndrome” which leads to worsened nocturnal hypoxia, hypercapnia, and more severe sleep disturbances [26]. CODP related symptoms clearly and robustly associate with poor sleep. Thus, lack of COPD optimal management may result in disturbed sleep due to nocturnal symptoms. Several physiological changes in sleep may worsen respiratory function at night and thus not only disturb sleep but also have nighttime adverse consequences like nocturnal increased risk of death or exacerbation [27–29]. It is important to note that among various sleep medicine metrics in our study, only total sleep time and sleep efficiency were of high predictive value towards clustering. Due to poor quality of sleep, patients report daytime excessive daytime sleepiness and fatigue, which can exacerbate COPD symptoms [30]. Apnea hypopnea index (AHI), the clinical marker for obstructive sleep apnea did not turn out to be an important predictor.

The strengths of our study are underpinned by its rigorous methodology, robust dataset, and comprehensive statistical analysis. With a sample of veterans that was longitudinally followed up for over two decades, our study encompassed a large and diverse population of patients with sleep problems and confirmed COPD diagnoses. We utilized a validated natural language processing (NLP) algorithm for accurate data extraction and an unsupervised machine learning algorithm, K-means clustering, to identify distinct phenotypic clusters. The elbow method minimized subjectivity in cluster selection, while Cox regression analysis adjusted for potential confounders, and Kaplan-Meier estimates strengthened the validity of our findings. Our study also considered relevant demographic and clinical variables, providing a comprehensive understanding of the impact of these factors on the study outcomes. Lastly, the non-parametric supervised learning algorithm with decision tree analysis further reinforced the reliability of our findings, given the extended longitudinal follow-up period.

While the study adds to the literature in meaningful ways, it also has limitations. Inherent to studies using dataset, a diagnosis made by automated measures is often less accurate than a clinical chart review. Furthermore, granular details such as the role of duration and adherence of therapeutic interventions such as cognitive behavioral therapy, medication, CPAP therapy and confounders such as history of smoking, and ejection fraction were not determined. Objective sleep parameters measured in a laboratory may differ significantly from those during a night of sleep at home. In addition, data extracted from one night of sleep may not represent a valid evaluation of sleep quality and quantity over an extended period. Many patients with more severe forms of OSA are diagnosed through split night PSG (baseline and titration done on the same night) or home sleep testing. Use of only baseline PSG in our study may have affected the AHI relationship in our analysis. Unmeasured confounders also potentially played a role and were not measured in this study. The male predominant veteran dataset also limits generalizability to the population. Several other clinically relevant variables such as respiratory distress index, pulmonary functions, oxygen saturation, carbon dioxide levels and frequency of exacerbation were not included in this study due to methodological limitations. Because of the observational nature of our analysis, our results will need to be validated using other cohorts. In addition, while we have demonstrated the importance of sleep parameters in determining mortality in our cohort, it remains unclear whether the lower TST and SE should be considered sleep biomarkers associated with increased mortality or whether they indeed have a direct causal relationship contributing to increased mortality.

If confirmed, our findings may have significant implications. GOLD therapeutic approach proposes utilizing a comprehensive evaluation that considers a person’s symptoms and risk of exacerbation as a basis for pharmacologic treatment guidance [1]. We propose that information related to total sleep time and sleep efficiency may complement a COPD patient’s risk stratification. With the widespread availability of high-tech wearable tools such as smart watches, the clinical incorporation of sleep parameters is within easy reach. Less resource intensive methods such as a sleep diary can also provide a good insight into a patient’s sleep.

In conclusion, our study utilized unsupervised machine learning to identify distinct COPD phenotypic clusters, highlighting the significant impact of total sleep time and sleep efficiency on mortality differences in this population. Despite its limitations, our findings suggest potential clinical implications for COPD phenotypic characterization and could inform future studies on incorporating sleep parameters into risk stratification.

Supplementary Material

Support

The study was supported by seed funding from Baylor College of Medicine, Houston, Texas, United States, the Center for Innovations in Quality, Effectiveness and Safety (CIN 13-413), Michael E. DeBakey VA Medical Center, Houston, TX, United states and a national institute of health (NIH), National Heart, Lung, and Blood Institute (NHLBI) K25 funding (#:1K25HL152006-01), VA Clinical Science Research & Development (IK2 CX001981), and Artificial Intelligence/Machine Learning Consortium to Advance Health Equity and Researcher Diversity (AIM-AHEAD) funding (OD032581-01S1).

Abbreviations list

Footnotes

References