PSQI-based severity-ordered sleep-quality strata and associated clinical and psychosocial factors among adults with untreated obstructive sleep apnea: a multicenter cross-sectional study

Abstract

Background

Obstructive sleep apnea (OSA) is prevalent and is associated with impaired functioning, cardiometabolic morbidity, and mental health problems. Subjective sleep quality is an important patient-reported outcome in OSA, yet most studies rely on a single Pittsburgh Sleep Quality Index (PSQI) global score, potentially masking heterogeneity in domain-level complaints. We aimed to identify PSQI-based sleep-quality strata among adults with untreated OSA and to examine their clinical and psychosocial correlates.

Methods

We conducted a multicenter cross-sectional study in two sleep monitoring centers in China (January-October 2025). Adults with newly diagnosed, untreated OSA completed the PSQI and questionnaires assessing depressive symptoms, anxiety symptoms, daytime sleepiness, perceived stress, and social support. Latent profile analysis using the seven PSQI component scores was used to summarize PSQI component patterns. Because the retained solution was expected to be graded rather than sharply separated, the resulting strata were ordered by increasing subjective sleep burden. Refit multinomial models based on most-likely stratum assignment were used descriptively to examine correlates, and a three-step classification-uncertainty-adjusted analysis was performed in parallel to account for uncertainty in latent stratum assignment. Recruitment center was included as a design-related covariate, and additional sensitivity analyses excluded participants with diabetes and restricted the analysis to men.

Results

Four graded PSQI-based strata were identified and presented as minimal (22.0%), mild (27.1%), moderate (24.7%), and severe (26.2%) sleep disturbance. Their component means were largely parallel and differed mainly in overall level, indicating severity ordering rather than decisively distinct symptom structures. Individual-level violin and boxplot displays showed within-stratum heterogeneity and partial overlap between adjacent strata, reinforcing the need to interpret the retained strata as severity-ordered PSQI strata rather than discrete clinical phenotypes. Greater subjective disturbance co-occurred with higher depressive and anxiety symptom scores, greater daytime sleepiness and perceived stress, and lower social support. In the severe-versus-minimal comparison, severe OSA, higher depressive symptom scores, higher anxiety scores, and lower social-support scores showed the most consistent associations, but the per-point psychosocial odds ratios were modest and should be interpreted as exploratory correlational estimates rather than clinically actionable causal effects.

Conclusions

Among adults with untreated OSA referred for laboratory polysomnography, PSQI-based latent profile analysis identified four severity-ordered descriptive strata of subjective sleep disturbance. The study did not provide strong evidence for qualitatively distinct symptom architectures beyond a graded PSQI severity pattern. More severe subjective burden co-occurred particularly with severe OSA, higher depressive and anxiety symptom burden, and lower social support. These findings remained materially similar after accounting for recruitment center and under classification-uncertainty-adjusted three-step analysis, but they should be interpreted cautiously in light of multiple exploratory comparisons, modest effect sizes, and the clinic-based sample.

Supplementary Information

The online version contains supplementary material available at 10.1186/s12889-026-27834-y.

Background

Obstructive sleep apnea (OSA) is a common chronic sleep-related breathing disorder characterized by recurrent upper-airway obstruction during sleep, resulting in intermittent hypoxemia and sleep fragmentation [1]. With population aging and rising obesity prevalence, the burden of OSA is increasing worldwide [2]. Untreated OSA is associated with cardiovascular and metabolic diseases and reduced health-related quality of life, creating substantial public health and healthcare-system burden [3].

Subjective sleep quality is an important patient-centered outcome in OSA management. The Pittsburgh Sleep Quality Index (PSQI) is widely used to assess perceived sleep quality across multiple dimensions, including sleep latency, sleep duration, sleep efficiency, sleep disturbances, sleep medication use, and daytime dysfunction [4–6]. Because the PSQI is a self-report instrument, it primarily captures subjective symptoms rather than objective polysomnographic signs. Poor subjective sleep quality has been linked to worse symptom burden and adverse clinical outcomes in people with OSA [7, 8]. However, most studies summarize PSQI results using a global score or a single cut-off, which may obscure whether patients with similar total scores show different domain-level complaint patterns.

OSA is increasingly recognized as a heterogeneous syndrome, and person-centered approaches can help describe clinically relevant symptom patterns [9]. Latent class analysis and latent profile analysis (LPA) identify unobserved patterns based on observed indicators rather than assuming that all individuals lie on a single continuum [10]. In the present context, however, the value of a PSQI-based person-centered analysis was not to assume a priori that qualitatively distinct subtypes must exist, but to test whether the seven PSQI components contained informative structure beyond a single global score while also determining whether the retained solution was better understood simply as an ordered severity gradient.

This distinction matters because subjective sleep complaints do not map perfectly onto traditional OSA severity indices such as the apnea–hypopnea index (AHI) or oxygen saturation metrics [11]. Patients with similar objective respiratory severity may report different constellations of insomnia-like complaints, daytime dysfunction, and sleep dissatisfaction, and those differences may coexist with varying levels of psychological distress and social support. Thus, examining PSQI component patterns may clarify how subjective symptom burden clusters within untreated OSA, even if the retained strata ultimately prove to be largely severity-ordered.

Psychological distress is common in OSA and may interact with sleep complaints through bidirectional pathways [12]. Depressive and anxiety symptoms, perceived stress, and social support are plausible correlates of subjective sleep quality and may influence health behaviors and treatment engagement [13, 14]. Accordingly, this study aimed (1) to identify PSQI-based subjective sleep-quality strata among adults with newly diagnosed, untreated OSA using the seven PSQI component scores, and (2) to examine sociodemographic, clinical, and psychosocial correlates of these strata while explicitly distinguishing subjective symptom measures from objective polysomnographic signs. We hypothesized that higher PSQI-based subjective burden would co-occur with worse physiological severity and greater psychological distress, but we treated these associations as exploratory and non-causal.

Materials and methods

Study design and participants

This multicenter cross-sectional study was conducted from January 2025 to October 2025. Participants were consecutively recruited from the Sleep Monitoring Center at the Pulmonary Oncology Department of the Affiliated Hospital of Changchun University of Chinese Medicine (Changchun, China) and the General Practice Department of the Affiliated Luohu Hospital of Shenzhen University Medical School (Shenzhen, China). A recruitment-center indicator (Center 1 vs Center 2 in the analytic dataset) was retained to examine potential site heterogeneity in descriptive and multivariable analyses.

Consecutive eligible patients were invited to participate. Inclusion criteria were: (a) OSA diagnosed by overnight polysomnography (apnea–hypopnea index [AHI] > 5 events/h) [15]; (b) age ≥ 18 years; (c) newly diagnosed and not receiving OSA-specific treatment at the time of recruitment (e.g., CPAP/APAP or mandibular advancement devices); (d) able to complete questionnaires; and (e) able to provide informed consent. Exclusion criteria were: (a) previously diagnosed psychiatric disorders documented in the medical history at recruitment, or cognitive impairment that could compromise questionnaire completion; (b) pregnancy; and (c) severe comorbidities judged by the treating clinicians to preclude safe participation or valid completion of study procedures. The latter criterion was pragmatic and referred to conditions that would interfere with participation rather than to chronic disease burden per se. Detailed counts by individual exclusion reason were not retained in the analytic report and should therefore be considered unavailable when interpreting the recruitment process.

Sample size

Currently, there is no universally accepted formula for a priori sample size calculation in latent profile analysis (LPA) [16]. We therefore adopted a pragmatic approach, aiming to recruit a sample size large enough to support stable estimation of latent profiles and subsequent multivariable analyses. As a general principle in multivariable regression modeling, small samples can yield unstable estimates and overfitting, whereas larger samples improve precision [17]. Considering 16 candidate predictors and allowing for incomplete questionnaires, we targeted at least 300 participants and ultimately enrolled 400 participants with complete data.

Outcome variable

Sleep quality over the past month was assessed using the PSQI. Originally developed by Buysse et al. in 1989, the PSQI is a comprehensive instrument for evaluating subjective sleep quality in both clinical and research settings [4]. The Chinese version of the PSQI has also been reported to have acceptable reliability and validity in Chinese populations [6]. The scale comprises 24 items, of which 18 are self-reported and contribute to scoring. These items are categorized into seven components: subjective sleep quality, sleep latency, sleep duration, habitual sleep efficiency, sleep disturbances, use of sleeping medication, and daytime dysfunction. Each component is rated on a scale from 0 to 3, and the scores are aggregated to produce a global PSQI score ranging from 0 to 21. In the present study, the seven PSQI component scores were treated as subjective symptom indicators and were used as continuous variables in the LPA.

Predictor variables

Sociodemographic variables

Sociodemographic characteristics were gathered through a self-constructed questionnaire. The variables assessed encompassed age, gender, educational attainment, marital status, monthly income per capita, duration of OSA, smoking status, alcohol consumption, coffee consumption, frequency of weekly exercise, recruitment center, and medical history of hypertension, diabetes, and heart disease. Smoking status was classified into three categories: never, ever, and current. Alcohol consumption was categorized as never, occasional, and daily. Coffee consumption was divided into almost never, occasional, and frequent. Weekly exercise was categorized as infrequent, occasional, and regular. Recruitment center was coded according to the site of enrollment and treated as a design-related covariate rather than a substantive exposure of primary interest.

Objective clinical signs and polysomnographic indicators

Clinical and polysomnographic parameters were extracted from medical records and sleep monitoring reports and were treated as objective clinical signs rather than self-reported symptoms. The parameters assessed included BMI, presence of cardiovascular disease, hyperlipidemia, AHI, OSA severity classification, ODI, total sleep time, MSaO₂, and LSaO₂. BMI was calculated using the formula weight divided by height squared (kg/m²) and was classified into categories of normal, overweight, or obese. The severity of OSA was categorized as mild, moderate, or severe.

Depressive symptoms

Depressive symptoms were assessed using the 9-item Patient Health Questionnaire (PHQ-9), a tool developed in accordance with the Diagnostic and Statistical Manual of Mental Disorders, Fourth Edition (DSM-IV), and widely used as a concise self-report screening measure for depression [18]. The PHQ-9 is available in the public domain. The Chinese version of the PHQ-9 has demonstrated robust reliability and validity among older Chinese adults in primary care, evidenced by a Cronbach’s α of 0.91 [19]. Each item on the scale is scored from 0 to 3, yielding a total possible score ranging from 0 to 27, with higher scores indicating more severe depressive symptoms. Because the PHQ-9 includes a sleep-related item, it was interpreted here as a brief depressive-symptom screening measure rather than as a perfectly unidimensional measure of depression.

Anxiety symptoms

Anxiety symptoms were evaluated utilizing the 7-item Generalized Anxiety Disorder scale (GAD-7), a publicly accessible instrument developed in accordance with DSM-IV criteria [20]. Chinese-language validation studies have reported acceptable psychometric performance for the GAD-7 in Chinese clinical and student populations [21]. Each item on the scale is rated on a four-point Likert scale, ranging from 0 ("not at all") to 3 ("nearly every day"). The total score ranges from 0 to 21, with higher scores indicative of more severe anxiety symptoms.

Daytime sleepiness

Daytime sleepiness was assessed utilizing the Epworth Sleepiness Scale (ESS), a widely recognized eight-item self-report instrument designed to evaluate the propensity for dozing off in various daytime scenarios [22]. The Chinese ESS has been evaluated in Chinese patients with OSA and healthy controls [23]. Each item is rated on a scale from 0 to 3 (0 indicating "would never doze" and 3 indicating a "high chance of dozing"), resulting in a cumulative score ranging from 0 to 24, with higher scores indicating increased levels of daytime sleepiness. The application of the Chinese version of the ESS in this study was authorized by Mapi Research Trust (Authorization No. 113533).

Social support

Perceived social support was assessed using the Social Support Rating Scale (SSRS), a tool developed by Shuiyuan Xiao in China to assess social relationships across three dimensions: objective support, subjective support, and utilization of social support [24]. The SSRS consists of 10 items distributed across these dimensions and has been extensively employed within Chinese populations, exhibiting strong reliability and validity. Authorization to use the SSRS in this study was secured from the original developer, and the scale was administered in compliance with the developer’s stipulations, with appropriate citation of the original publication. Because the total SSRS score combines support availability, perceived support, and support utilization, higher or lower total scores should be interpreted as reflecting an overall pattern of more or less favorable social-support resources and engagement, rather than a single homogeneous social-support construct.

Perceived stress

Perceived stress was assessed utilizing the 14-item Perceived Stress Scale (PSS)developed by Cohen et al. and subsequently translated into Chinese as the Chinese Perceived Stress Scale (CPSS) in 2003 [25, 26]. Authorization for the use of the Chinese version in this study was granted by Mapi Research Trust (Authorization No. 113489). The CPSS encompasses 14 items that address the dimensions of tension and loss of control. Responses are recorded on a five-point Likert scale ranging from 0 to 4, with seven items reverse-scored. Higher aggregate scores reflect elevated levels of perceived stress.

Data collection procedures

Eligible patients were approached during their attendance at the sleep monitoring centers. Trained research nurses provided detailed explanations of the study objectives and procedures, obtained written informed consent, and then invited patients to complete self-administered questionnaires in a quiet setting. For participants with limited literacy, investigators read the items aloud and recorded the responses verbatim. To minimize information bias, another researcher extracted clinical and polysomnographic data from the electronic medical records. Questionnaires with substantial missing data or evidently inconsistent responses were reviewed promptly and, when feasible, completed or corrected at the point of collection. In the final analytic files used for the reported models, all PSQI component indicators and all covariates required for the regression analyses were complete for all 400 participants.

Statistical analysis

All analyses were performed using R (version 4.4.2), Python, and Mplus (version 8.3). Continuous variables with approximately normal distributions are presented as mean ± standard deviation (SD) and were compared across strata using one-way analysis of variance (ANOVA). Skewed continuous variables are summarized as median and interquartile range (IQR) and were compared using Kruskal–Wallis H tests. Categorical variables are expressed as frequencies and percentages and were compared using χ² tests or Fisher’s exact tests, as appropriate. Self-report questionnaire measures were interpreted as subjective symptom indicators, whereas PSG-derived and other clinical parameters were interpreted as objective signs. Because many descriptive and regression comparisons were examined, P values were treated as exploratory indicators rather than as stand-alone evidence. No formal Bonferroni correction was applied; interpretation emphasized effect sizes, 95% confidence intervals, directionality, clinical plausibility, and consistency across sensitivity analyses. To make within-stratum overlap transparent, individual-level violin plots with embedded boxplots and jittered observations were generated for the global PSQI score and for each of the seven PSQI component scores. No missing values were present in the variables used for the LPA and multinomial regression analyses; therefore, no imputation procedure was required for the reported models.

Latent profile analysis (LPA) was conducted in Mplus to summarize PSQI component-score patterns among patients with OSA using the seven PSQI component scores as continuous indicators. Unlike variable-centered approaches that model only a single global PSQI score, LPA models the multivariate distribution of the seven PSQI components, yields profile-specific means, and assigns posterior membership probabilities. In this study, the person-centered approach was used to assess whether the PSQI components carried clinically informative structure beyond a single total score, while recognizing that the resulting solution might primarily reflect an ordered severity continuum rather than qualitatively distinct symptom architectures. Models with one to five profiles were estimated. Model selection was based on the Akaike Information Criterion (AIC), Bayesian Information Criterion (BIC), sample-size-adjusted BIC (aBIC), entropy, the Lo-Mendell-Rubin adjusted likelihood ratio test (LMRT), and the bootstrap likelihood ratio test (BLRT), together with profile sizes and pragmatic interpretability. Average posterior probabilities were examined to evaluate classification precision. For presentation, the retained profiles were ordered from minimal to severe disturbance according to their PSQI component means.

After the optimal LPA solution was determined, differences in sociodemographic, clinical, and psychosocial variables across the PSQI-based strata were examined using ANOVA, Kruskal–Wallis tests, χ² tests, or Fisher’s exact tests, as appropriate. Variables associated with the PSQI-based strata in univariate analyses (P < 0.05) were entered into multinomial logistic regression models to examine correlates of membership. Recruitment center was entered into the refit multinomial model as a design-related covariate regardless of its univariate P value. Multicollinearity was assessed using tolerance and variance inflation factor (VIF), and highly collinear variables were not included simultaneously. To address uncertainty in latent profile assignment, we performed a three-step classification-uncertainty-adjusted latent-profile regression as the principal corroborative analysis. This approach uses information from the posterior classification probabilities/classification-error structure rather than assuming that each modal profile assignment is error-free. The same covariates were included as in the refit minimal-reference model. For ease of clinical reading, we also report a descriptive refit multinomial logistic regression based on participants’ most likely stratum assignment, with the minimal-disturbance stratum as the reference category and recruitment center included as a design-related covariate. Additional sensitivity analyses repeated the most-likely-class model after excluding participants with diabetes and within the male subgroup because of the marked sex imbalance. Results are reported as odds ratios (ORs) with 95% confidence intervals (CIs); the classification-uncertainty-adjusted corroborative analysis is presented in Supplementary Table S6.

Reporting guidelines

This cross-sectional observational study was designed and reported in alignment with the guidelines outlined in the Strengthening the Reporting of Observational Studies in Epidemiology (STROBE) statement for cross-sectional studies [27].

Ethics approval and consent to participate

Ethics approval for this study was obtained from the Ethics Committee of the Affiliated Hospital of Changchun University of Chinese Medicine (Approval No. CCZYFYLL-SQ-2025-289). All procedures were conducted in accordance with the ethical standards of the 1964 Declaration of Helsinki and its later amendments. Written informed consent was obtained from all participants before enrolment. Each participant was fully informed about the study objectives, procedures, potential risks and benefits, and was advised of their right to withdraw from the study at any time without any negative consequences.

Results

Participant characteristics

A total of 400 patients diagnosed with OSA were included in this complete-case analysis. The mean age of the participants was 44.85 years (± 11.87), and 90.5% were male. No missing values were present in the variables used for the LPA or the reported multinomial regression models. The distributions of the demographic and clinical characteristics are presented in Table 1, with the retained strata shown in increasing severity order (Stratum 1 minimal, Stratum 2 mild, Stratum 3 moderate, Stratum 4 severe). Exploratory univariate analyses revealed differences across strata for BMI category (P = 0.022), alcohol consumption (P = 0.026), coffee consumption (Fisher’s exact P = 0.005), weekly exercise (P = 0.011), hypertension (P = 0.009), diabetes (P = 0.032), OSA severity (P < 0.001), AHI (P < 0.001), ODI (P < 0.001), MSaO₂ (P < 0.001), and LSaO₂ (P < 0.001). These P values were interpreted descriptively in view of multiple comparisons. The distribution of strata did not differ significantly between the two recruitment centers (χ² = 1.15, P = 0.766). No significant differences were observed for age, sex, education, marital status, income, smoking status, history of heart disease, cerebrovascular disease, hyperlipidemia, OSA duration, or total sleep time (all P > 0.05).

Table 1.

Sociodemographic and clinical characteristics by PSQI-based severity-ordered stratum (n = 400)

Characteristic	Total (n = 400)	Stratum 1 (minimal, n = 89)	Stratum 2 (mild, n = 107)	Stratum 3 (moderate, n = 101)	Stratum 4 (severe, n = 103)	Test statistic	P value
Age, Mean ± SD	44.85 ± 11.87	43.96 ± 11.44	45.20 ± 12.49	43.91 ± 12.19	46.18 ± 11.27	F = 0.84	0.471
Sex, n (%)						χ2 = 1.20	0.753
Male	362 (90.50)	82 (92.13)	96 (89.72)	93 (92.08)	91 (88.35)
Female	38 (9.50)	7 (7.87)	11 (10.28)	8 (7.92)	12 (11.65)
Education level, n (%)						χ2 = 7.36	0.289
High school below	119 (29.75)	26 (29.21)	32 (29.91)	25 (24.75)	36 (34.95)
High school/technical secondary education	115 (28.75)	31 (34.83)	27 (25.23)	26 (25.74)	31 (30.10)
College degree or above	166 (41.50)	32 (35.96)	48 (44.86)	50 (49.50)	36 (34.95)
Marital status, n (%)						χ2 = 6.37	0.702
Married	239 (59.75)	51 (57.30)	66 (61.68)	57 (56.44)	65 (63.11)
Unmarried	72 (18.00)	20 (22.47)	17 (15.89)	22 (21.78)	13 (12.62)
Divorced	56 (14.00)	13 (14.61)	13 (12.15)	15 (14.85)	15 (14.56)
Widowed	33 (8.25)	5 (5.62)	11 (10.28)	7 (6.93)	10 (9.71)
Per capita monthly income (yuan), n (%)						χ2 = 13.84	0.128
< 3000	171 (42.75)	46 (51.69)	41 (38.32)	38 (37.62)	46 (44.66)
3000–5000	67 (16.75)	10 (11.24)	20 (18.69)	23 (22.77)	14 (13.59)
5000–8000	113 (28.25)	22 (24.72)	38 (35.51)	24 (23.76)	29 (28.16)
> 8000	49 (12.25)	11 (12.36)	8 (7.48)	16 (15.84)	14 (13.59)
Recruitment center, n (%)						χ2 = 1.15	0.766
Center 1	244 (61.00)	54 (60.67)	63 (58.88)	66 (65.35)	61 (59.22)
Center 2	156 (39.00)	35 (39.33)	44 (41.12)	35 (34.65)	42 (40.78)
BMI, n (%)						χ2 = 14.77	0.022
Normal	86 (21.50)	27 (30.34)	23 (21.50)	23 (22.77)	13 (12.62)
Overweight	107 (26.75)	27 (30.34)	31 (28.97)	19 (18.81)	30 (29.13)
Obese	207 (51.75)	35 (39.33)	53 (49.53)	59 (58.42)	60 (58.25)
Smoking status, n (%)						χ2 = 5.69	0.459
Never	107 (26.75)	22 (24.72)	32 (29.91)	25 (24.75)	28 (27.18)
Former	143 (35.75)	25 (28.09)	39 (36.45)	39 (38.61)	40 (38.83)
Current	150 (37.50)	42 (47.19)	36 (33.64)	37 (36.63)	35 (33.98)
Alcohol consumption, n (%)						χ2 = 14.39	0.026
Never	104 (26.00)	35 (39.33)	30 (28.04)	18 (17.82)	21 (20.39)
Occasional	90 (22.50)	17 (19.10)	25 (23.36)	26 (25.74)	22 (21.36)
Daily	206 (51.50)	37 (41.57)	52 (48.60)	57 (56.44)	60 (58.25)
Coffee consumption, n (%)						Fisher’s exact	0.005
Almost none	226 (56.50)	62 (69.66)	65 (60.75)	47 (46.53)	52 (50.49)
Occasional	23 (5.75)	6 (6.74)	4 (3.74)	4 (3.96)	9 (8.74)
Frequent	151 (37.75)	21 (23.60)	38 (35.51)	50 (49.50)	42 (40.78)
Weekly exercise, n (%)						χ2 = 16.66	0.011
Rare/none	194 (48.50)	30 (33.71)	57 (53.27)	54 (53.47)	53 (51.46)
Occasionally	163 (40.75)	41 (46.07)	44 (41.12)	38 (37.62)	40 (38.83)
Regular	43 (10.75)	18 (20.22)	6 (5.61)	9 (8.91)	10 (9.71)
Hypertension, n (%)						χ2 = 11.47	0.009
Yes	226 (56.50)	38 (42.70)	58 (54.21)	64 (63.37)	66 (64.08)
No	174 (43.50)	51 (57.30)	49 (45.79)	37 (36.63)	37 (35.92)
Diabetes, n (%)						χ2 = 8.78	0.032
Yes	46 (11.50)	5 (5.62)	20 (18.69)	10 (9.90)	11 (10.68)
No	354 (88.50)	84 (94.38)	87 (81.31)	91 (90.10)	92 (89.32)
Heart disease, n (%)						χ2 = 5.40	0.145
Yes	73 (18.25)	9 (10.11)	21 (19.63)	20 (19.80)	23 (22.33)
No	327 (81.75)	80 (89.89)	86 (80.37)	81 (80.20)	80 (77.67)
Cerebrovascular disease, n (%)						χ2 = 3.61	0.307
Yes	82 (20.50)	14 (15.73)	27 (25.23)	23 (22.77)	18 (17.48)
No	318 (79.50)	75 (84.27)	80 (74.77)	78 (77.23)	85 (82.52)
Hyperlipidemia, n (%)						χ2 = 1.78	0.619
Yes	140 (35.00)	29 (32.58)	43 (40.19)	33 (32.67)	35 (33.98)
No	260 (65.00)	60 (67.42)	64 (59.81)	68 (67.33)	68 (66.02)
OSA severity, n (%)						χ2 = 41.64	< 0.001
Mild	78 (19.50)	34 (38.20)	14 (13.08)	12 (11.88)	18 (17.48)
Moderate	175 (43.75)	44 (49.44)	43 (40.19)	46 (45.54)	42 (40.78)
Severe	147 (36.75)	11 (12.36)	50 (46.73)	43 (42.57)	43 (41.75)
OSA duration (years), median (IQR)	15.00 (10.00, 20.00)	15.00 (6.00,20.00)	15.00 (9.00,20.00)	15.00 (10.00,20.00)	15.00 (10.00,25.00)	H = 5.42	0.144
AHI (events/h), median (IQR)	27.10 (21.32, 45.23)	21.70 (12.70,27.10)	29.40 (23.70,45.50)	27.80 (23.50,55.80)	27.60 (23.50,52.70)	H = 40.53	< 0.001
ODI (events/h), median (IQR)	24.55 (17.95, 41.40)	18.90 (10.70,25.00)	26.40 (20.70,41.40)	25.60 (21.20,52.40)	25.10 (20.95,48.45)	H = 39.49	< 0.001
Total sleep time (min), median (IQR)	408.35 (394.02, 480.00)	408.30 (379.10,433.50)	409.50 (393.35,500.05)	408.20 (394.70,466.50)	408.20 (398.50,488.50)	H = 3.14	0.371
MSaO₂ (%), median (IQR)	92.80 (90.20, 94.10)	93.80 (92.40,95.00)	92.90 (90.30,94.00)	91.90 (88.40,93.70)	92.20 (89.50,93.80)	H = 30.42	< 0.001
LSaO₂ (%), median (IQR)	76.00 (66.75, 83.00)	81.00 (76.00,86.00)	75.00 (67.00,83.00)	72.00 (60.00,81.00)	74.00 (63.00,81.00)	H = 38.09	< 0.001

Values are mean ± SD, median (IQR), or n (%). P values were calculated using one-way analysis of variance (ANOVA) for normally distributed continuous variables, Kruskal–Wallis H tests for skewed continuous variables, chi-square tests for most categorical variables, and Fisher’s exact test for coffee consumption because of small expected cell counts

OSA obstructive sleep apnea, BMI body mass index, AHI apnea–hypopnea index (events/h), ODI oxygen desaturation index (events/h), MSaO₂ mean oxygen saturation (%), LSaO₂ lowest oxygen saturation (%), Center 1 Changchun, China, Center 2 Shenzhen, China

Bold P values indicate nominal statistical significance at P < 0.05

Latent profile model fit

Table 2 presents the fit indices for the LPA solutions ranging from one to five profiles. The AIC, BIC, and aBIC values decreased substantially from the one-profile to the four-profile solutions, indicating progressively improved fit. For the two-, three-, and four-profile models, both the LMRT and BLRT supported the k-profile solution relative to the k-1 model. Although the five-profile solution showed slightly higher entropy, it did not converge properly because of insufficient E-steps; its parameter estimates and fit indices were therefore not interpreted. We retained the four-profile solution on the basis of statistical fit, balanced profile sizes, and pragmatic interpretability. The Mplus output for the four-profile model used 20 initial random starts and 4 final-stage optimizations, and the best loglikelihood value was replicated across four solutions, which supports local stability of the retained model. To strengthen this point further, an additional high-start rerun reproduced the same optimum to three decimals and yielded virtually identical ordered class proportions and profile-specific means (Supplementary Table S5). Accordingly, the four-profile solution appears locally stable within the current analytic framework, but it should be interpreted as a severity-ordered descriptive summary of PSQI burden rather than as definitive evidence of a unique latent structure or sharply separated clinical subtypes. For ease of interpretation in the manuscript, the four retained profiles were renumbered in increasing severity order. Average posterior probabilities for most likely profile membership are shown in Table 3. The diagonal probabilities were high for all four assigned profiles (0.966, 0.958, 0.942, and 0.989 for Profiles 1–4, respectively), with only small off-diagonal probabilities, indicating good classification precision for the retained four-profile solution.

Table 2.

Model fit indices for latent profile solutions based on the seven PSQI components (n = 400)

Number of profiles	AIC	BIC	aBIC	LMRT P value	BLRT P value	Entropy	Profile proportions (%)
1	8016.540	8072.420	8027.997	—	—	—	100
2	5822.760	5910.572	5840.765	< 0.001	< 0.001	0.942	51.617/48.383
3	5072.943	5192.687	5097.495	< 0.001	< 0.001	0.949	28.233/35.478/36.289
4	4776.505	4928.181	4807.604	0.02	< 0.001	0.930	21.970/27.092/24.717/26.221
5	—	—	—	—	—	0.982	31.750/18.500/13.411/16.339/20.000a

AIC Akaike information criterion, BIC Bayesian information criterion, aBIC sample-size-adjusted BIC, LMRT Lo–Mendell–Rubin adjusted likelihood ratio test, BLRT bootstrap likelihood ratio test. LMRT and BLRT compare a k-profile model with a (k − 1)-profile model. For the retained four-profile solution, the manuscript orders the profiles by increasing severity (minimal, mild, moderate, severe). The five-profile solution did not converge and is shown for descriptive purposes only

Bold values indicate the retained four-profile solution selected for interpretation

Table 3.

Average posterior probabilities for most likely profile membership (four-profile solution)

Assigned profile	Profile 1	Profile 2	Profile 3	Profile 4
Profile 1	0.966	0.034	0.000	0.000
Profile 2	0.018	0.958	0.024	0.000
Profile 3	0.000	0.028	0.942	0.030
Profile 4	0.000	0.000	0.011	0.989

Rows indicate assigned profile (most likely class) and columns indicate the model-based latent profile. Higher diagonal values indicate better classification accuracy. Profiles are shown in manuscript order: Profile 1 minimal sleep disturbance, Profile 2 mild sleep disturbance, Profile 3 moderate sleep disturbance, and Profile 4 severe sleep disturbance

Description of PSQI-based severity strata

Four PSQI-based strata were identified from the seven PSQI components and, for presentation, were ordered by increasing subjective burden (Figs. 1 and 2). Stratum 1, the minimal-disturbance stratum, included 22.0% of the sample (n = 89) and showed the lowest scores on all PSQI components, with a mean global PSQI score of 2.15 ± 1.51. Stratum 2, the mild-disturbance stratum, comprised 27.1% of the sample (n = 107) and showed modestly elevated scores across components, with a mean global PSQI score of 7.47 ± 1.59. Stratum 3, the moderate-disturbance stratum, accounted for 24.7% of participants (n = 101) and displayed broader elevations across domains, with a mean global PSQI score of 13.16 ± 1.76. Stratum 4, the severe-disturbance stratum, comprised 26.2% of the sample (n = 103) and had consistently high component means (approximately 2.5–2.8) and a mean global PSQI score of 19.16 ± 1.43. Figure 1 now displays both stratum-specific component means with 95% CIs and individual-level global PSQI distributions using violin plots, embedded boxplots, and jittered observations. Figure 2 further shows the individual-level distributions of each PSQI component within the retained strata. These participant-level displays indicate that the strata are separated mainly by overall PSQI level, while within-stratum component scores remain heterogeneous and adjacent strata show partial distributional overlap. Thus, the retained solution is best understood primarily as a graded PSQI severity continuum rather than as sharply distinct qualitative symptom configurations.

Fig. 1.

PSQI-based severity-ordered strata derived from latent profile analysis. Panel A shows mean PSQI component scores (0–3) across the seven PSQI domains for each retained stratum among adults with untreated OSA (n = 400). Error bars in Panel A indicate 95% confidence intervals around the stratum-specific mean scores and should be interpreted as uncertainty around the estimated mean centers, not as the full within-stratum distribution. Panel B displays individual-level global PSQI distributions (0–21) within each severity-ordered stratum using violin plots, embedded boxplots, and jittered observations; numeric annotations show mean ± SD. Higher scores indicate worse subjective sleep disturbance. Strata are presented in ascending order of severity: minimal disturbance (22.0%, n = 89), mild disturbance (27.1%, n = 107), moderate disturbance (24.7%, n = 101), and severe disturbance (26.2%, n = 103). The figure is intended to show both stratum centers and within-stratum dispersion rather than to imply non-overlapping clinical phenotypes. PSQI: Pittsburgh Sleep Quality Index; OSA: obstructive sleep apnea

Fig. 2.

Individual-level distributions of the seven PSQI component scores by PSQI-based severity-ordered stratum. Violin plots with embedded boxplots and jittered observations show within-stratum distributions for each PSQI component (0–3). Higher scores indicate worse subjective sleep disturbance. The component-specific displays show within-stratum heterogeneity and partial overlap between adjacent strata, supporting cautious interpretation of the retained strata as severity-ordered PSQI strata rather than discrete clinical phenotypes. PSQI: Pittsburgh Sleep Quality Index

Differences in psychosocial and sleep-related questionnaire scores across PSQI-based strata

Table 4 summarizes the questionnaire-based comparisons across the four PSQI-based strata. Global PSQI scores differed markedly across strata (F = 2087.47, P < 0.001), consistent with the severity ordering described above. Exploratory between-stratum comparisons also suggested higher depressive symptoms (F = 13.29, P < 0.001), anxiety symptoms (F = 6.53, P < 0.001), daytime sleepiness (F = 9.21, P < 0.001), perceived stress (F = 3.80, P = 0.010), and lower social support (F = 2.73, P = 0.044) with increasing subjective sleep burden. However, these questionnaire contrasts should be read as descriptive patterns rather than as multiplicity-adjusted confirmatory findings. Patients in the minimal-disturbance stratum had the lowest mean scores for depression (5.55 ± 3.93), anxiety (8.61 ± 5.14), daytime sleepiness (7.92 ± 6.02), and perceived stress (29.30 ± 13.36), along with the highest social-support score (41.42 ± 12.98). Patients in the severe stratum had the highest PSQI score and relatively elevated depression, anxiety, and daytime sleepiness scores, along with the lowest social-support score (35.28 ± 15.68). Thus, greater subjective sleep burden co-occurred with greater psychological distress and lower social-support resources/utilization.

Table 4.

Questionnaire scores by PSQI-based severity-ordered stratum (n = 400)

Measure	Total (n = 400)	Stratum 1 (minimal, n = 89)	Stratum 2 (mild, n = 107)	Stratum 3 (moderate, n = 101)	Stratum 4 (severe, n = 103)	F	P value
Global PSQI score, mean ± SD	10.73 ± 6.45	2.15 ± 1.51	7.47 ± 1.59	13.16 ± 1.76	19.16 ± 1.43	2087.47	< 0.001
PHQ-9 score (depressive symptoms), mean ± SD	8.27 ± 5.04	5.55 ± 3.93	8.35 ± 5.13	9.33 ± 5.05	9.50 ± 4.95	13.29	< 0.001
GAD-7 score (anxiety symptoms), mean ± SD	10.88 ± 5.66	8.61 ± 5.14	11.42 ± 5.86	11.79 ± 5.75	11.39 ± 5.33	6.53	< 0.001
ESS score (daytime sleepiness), mean ± SD	11.04 ± 6.54	7.92 ± 6.02	12.02 ± 6.31	11.74 ± 6.13	12.01 ± 6.87	9.21	< 0.001
CPSS score (perceived stress), mean ± SD	33.31 ± 13.13	29.30 ± 13.36	35.15 ± 12.77	34.26 ± 12.72	33.94 ± 13.18	3.80	0.010
SSRS score (social support), mean ± SD	38.08 ± 15.19	41.42 ± 12.98	38.58 ± 15.39	37.46 ± 15.87	35.28 ± 15.68	2.73	0.044

Values are mean ± SD. P values were calculated using one-way analysis of variance (ANOVA)

PSQI Pittsburgh Sleep Quality Index, PHQ-9 Patient Health Questionnaire-9, GAD-7 Generalized Anxiety Disorder-7, ESS Epworth Sleepiness Scale, CPSS Chinese Perceived Stress Scale, SSRS Social Support Rating Scale

Bold P values indicate nominal statistical significance at P < 0.05

Multicollinearity diagnostics

The findings of the multicollinearity assessment are presented in Table 5. The majority of predictors exhibited acceptable tolerance values and VIFs below 3, indicating a minimal risk of collinearity. However, AHI and ODI demonstrated exceptionally low tolerance values (0.005) and exceedingly high VIFs (greater than 180), indicating severe collinearity between these two indices. As a result, AHI and ODI were excluded from the multinomial logistic regression models, while the other predictors were retained for further analysis.

Table 5.

Multicollinearity diagnostics for candidate predictors

Predictor	Tolerance	VIF
BMI (continuous)	0.535	1.87
Alcohol consumption (categorical)	0.777	1.287
Coffee consumption (categorical)	0.484	2.065
Weekly exercise (categorical)	0.861	1.162
Hypertension (yes/no)	0.837	1.195
Diabetes (yes/no)	0.893	1.12
AHI (events/h)	0.005	182.316
OSA severity (categorical)	0.309	3.237
ODI (events/h)	0.005	181.945
MSaO₂ (%)	0.676	1.479
LSaO₂ (%)	0.54	1.851
PHQ-9 score	0.543	1.842
GAD-7 score	0.716	1.396
ESS score	0.592	1.688
CPSS score	0.664	1.506
SSRS score	0.955	1.047

Tolerance and variance inflation factor (VIF) were calculated for candidate predictors. Lower tolerance and higher VIF indicate greater multicollinearity

Abbreviations: AHI apnea–hypopnea index, ODI oxygen desaturation index, MSaO₂ mean oxygen saturation, LSaO₂ lowest oxygen saturation, PHQ-9 Patient Health Questionnaire-9, GAD-7 Generalized Anxiety Disorder-7, ESS Epworth Sleepiness Scale, CPSS Chinese Perceived Stress Scale, SSRS Social Support Rating Scale

Bold values indicate severe multicollinearity, defined as tolerance < 0.10 and VIF > 10

Correlates of PSQI-based stratum membership

For descriptive comparison against the minimal-disturbance stratum, we refit a multinomial logistic regression using Stratum 1 as the reference category (Table 6). Because this approach is based on most-likely stratum assignment, we additionally evaluated the same covariate pattern using a classification-uncertainty-adjusted three-step latent-profile regression (Supplementary Table S6). ORs > 1 indicate higher odds of membership in the comparison stratum (Strata 2–4) rather than the minimal stratum, whereas ORs < 1 indicate lower odds relative to the minimal stratum. Compared with the minimal stratum, the mild stratum was associated with less frequent regular exercise (OR = 0.21, 95% CI: 0.07–0.65, P = 0.007), more frequent diabetes (OR = 5.25, 95% CI: 1.50–18.43, P = 0.010), severe OSA (OR = 8.99, 95% CI: 2.97–27.19, P < 0.001), and a modestly higher GAD-7 score (OR = 1.09 per point, 95% CI: 1.01–1.17, P = 0.020). The moderate stratum was associated with severe OSA (OR = 6.54, 95% CI: 1.99–21.45, P = 0.002), higher depressive symptom score (OR = 1.12 per point, 95% CI: 1.02–1.23, P = 0.023), and higher anxiety score (OR = 1.14 per point, 95% CI: 1.06–1.23, P < 0.001). In the severe-versus-minimal comparison, severe OSA (OR = 3.73, 95% CI: 1.23–11.29, P = 0.020), higher depressive symptom score (OR = 1.16 per point, 95% CI: 1.05–1.28, P = 0.003), higher anxiety score (OR = 1.12 per point, 95% CI: 1.04–1.20, P = 0.003), and lower social-support score (OR = 0.97 per point, 95% CI: 0.95–1.00, P = 0.028) showed the most consistent pattern. To avoid overstating small per-point effects, these estimates are best interpreted over clinically meaningful score differences: for example, a 5-point higher GAD-7 score corresponds approximately to OR = 1.54 for the mild-versus-minimal contrast and OR = 1.76 for the severe-versus-minimal contrast, whereas a 5-point higher PHQ-9 score corresponds approximately to OR = 2.10 for the severe-versus-minimal contrast. Recruitment center was not significantly associated with mild-versus-minimal (OR = 1.00, 95% CI: 0.52–1.95, P = 0.989), moderate-versus-minimal (OR = 0.74, 95% CI: 0.37–1.49, P = 0.401), or severe-versus-minimal (OR = 1.09, 95% CI: 0.55–2.13, P = 0.811) membership. The full exact reciprocal contrasts of the current minimal-reference model are provided in Supplementary Table S1, the exact direct severe-versus-minimal model is presented in Supplementary Table S2, and the diabetes-exclusion and male-only sensitivity analyses are shown in Supplementary Tables S3 and S4, respectively. These analyses were broadly concordant, and the classification-uncertainty-adjusted three-step analysis remained directionally and inferentially consistent with the refit most-likely-class model (Supplementary Table S6). Nevertheless, all regression findings should be interpreted as exploratory correlations in a cross-sectional clinic-based sample rather than as causal or immediately actionable clinical thresholds.

Table 6.

Multinomial logistic regression examining correlates of PSQI-based stratum membership (reference: Stratum 1, minimal disturbance; n = 400)

Predictor	Stratum 2 vs Stratum 1		Stratum 3 vs Stratum 1		Stratum 4 vs Stratum 1
	OR (95% CI)	P value	OR (95% CI)	P value	OR (95% CI)	P value
BMI
Normal	1.00 (Reference)		1.00 (Reference)		1.00 (Reference)
Overweight	1.61 (0.63–4.15)	0.322	0.77 (0.28–2.13)	0.613	1.90 (0.68–5.33)	0.220
Obese	1.17 (0.40–3.47)	0.775	0.78 (0.24–2.55)	0.687	2.99 (0.96–9.25)	0.058
Alcohol consumption
Never	1.00 (Reference)		1.00 (Reference)		1.00 (Reference)
Occasional	1.02 (0.38–2.78)	0.964	1.36 (0.47–3.92)	0.573	1.17 (0.42–3.27)	0.768
Daily	1.43 (0.60–3.40)	0.414	1.72 (0.68–4.32)	0.249	1.67 (0.69–4.02)	0.255
Coffee consumption
Almost none	1.00 (Reference)		1.00 (Reference)		1.00 (Reference)
Occasional	0.35 (0.07–1.76)	0.204	0.70 (0.13–3.80)	0.683	1.38 (0.33–5.81)	0.664
Frequent	1.41 (0.46–4.30)	0.550	1.92 (0.57–6.42)	0.292	0.77 (0.25–2.31)	0.637
Weekly exercise
Rare/none	1.00 (Reference)		1.00 (Reference)		1.00 (Reference)
Occasional	0.95 (0.45–2.03)	0.895	1.23 (0.56–2.70)	0.612	0.92 (0.42–2.00)	0.827
Regular	0.21 (0.07–0.65)	0.007	0.56 (0.19–1.68)	0.303	0.57 (0.20–1.61)	0.292
Hypertension
No	1.00 (Reference)		1.00 (Reference)		1.00 (Reference)
Yes	1.00 (0.49–2.02)	0.996	1.52 (0.73–3.16)	0.267	1.39 (0.68–2.84)	0.368
Diabetes
No	1.00 (Reference)		1.00 (Reference)		1.00 (Reference)
Yes	5.25 (1.50–18.43)	0.010	2.17 (0.56–8.40)	0.260	1.90 (0.52–6.98)	0.334
OSA severity
Mild	1.00 (Reference)		1.00 (Reference)		1.00 (Reference)
Moderate	1.88 (0.79–4.50)	0.154	2.43 (0.94–6.26)	0.067	1.21 (0.51–2.87)	0.666
Severe	8.99 (2.97–27.19)	< 0.001	6.54 (1.99–21.45)	0.002	3.73 (1.23–11.29)	0.020
Recruitment center
Center 1	1.00 (Reference)		1.00 (Reference)		1.00 (Reference)
Center 2	1.00 (0.52–1.95)	0.989	0.74 (0.37–1.49)	0.401	1.09 (0.55–2.13)	0.811
MSaO₂ (%) per 1% increase	0.91 (0.77–1.07)	0.254	0.94 (0.79–1.10)	0.421	0.92 (0.78–1.08)	0.315
LSaO₂ (%) per 1% increase	0.99 (0.94–1.05)	0.834	0.95 (0.90–1.01)	0.088	0.97 (0.92–1.03)	0.364
PHQ-9 score	1.04 (0.95–1.15)	0.377	1.12 (1.02–1.23)	0.023	1.16 (1.05–1.28)	0.003
GAD-7 score	1.09 (1.01–1.17)	0.020	1.14 (1.06–1.23)	< 0.001	1.12 (1.04–1.20)	0.003
ESS score	1.00 (0.94–1.07)	0.941	0.99 (0.92–1.06)	0.696	1.00 (0.93–1.06)	0.894
CPSS score	1.01 (0.98–1.04)	0.615	0.99 (0.96–1.02)	0.581	0.98 (0.96–1.01)	0.260
SSRS score	0.99 (0.96–1.01)	0.281	0.98 (0.96–1.00)	0.106	0.97 (0.95–1.00)	0.028

Strata are presented in increasing severity order: Stratum 1 minimal sleep disturbance, Stratum 2 mild sleep disturbance, Stratum 3 moderate sleep disturbance, and Stratum 4 severe sleep disturbance. Odds ratios (ORs) and 95% confidence intervals (CIs) are shown. ORs > 1 indicate higher odds of being classified into the comparison stratum rather than Stratum 1 (reference), whereas ORs < 1 indicate lower odds. Categorical predictors were entered with the following reference levels: normal BMI, never alcohol, almost no coffee, rare/none exercise, no hypertension/diabetes, and mild OSA. Continuous predictors are scaled per 1-unit increase (for oxygen saturation, per 1% increase). AHI and ODI were excluded from the multivariable models due to severe collinearity. The regression used most-likely stratum assignment and therefore does not fully account for latent-class uncertainty

OSA obstructive sleep apnea, BMI body mass index, MSaO₂ mean oxygen saturation (%), LSaO₂ lowest oxygen saturation (%), Center 1: Changchun, China, Center 2 Shenzhen, China, PHQ-9 Patient Health Questionnaire-9, GAD-7 Generalized Anxiety Disorder-7, ESS Epworth Sleepiness Scale, CPSS Chinese Perceived Stress Scale, SSRS Social Support Rating Scale

Bold P values indicate nominal statistical significance at P < 0.05

Discussion

Using LPA of the seven PSQI component scores, we identified four severity-ordered PSQI strata among 400 adults with newly diagnosed, untreated OSA. For clarity, these were presented as minimal, mild, moderate, and severe disturbance strata. The four-stratum solution showed high posterior classification precision, although the stratum mean patterns were largely parallel and differed mainly in overall level. Accordingly, the key negative finding is that the analysis did not provide strong evidence for decisively different symptom architectures beyond a graded PSQI severity pattern. The solution is therefore best interpreted as a descriptive summary of subjective sleep burden within this clinic-based sample, rather than as proof of sharply distinct natural subtypes.

The fact that the retained strata were primarily severity-ordered does not make the analysis uninformative; rather, it helps show how multidimensional subjective burden co-occurs with clinical and psychosocial correlates across increasing PSQI burden. At the same time, the added value of the LPA should be described modestly. In the present sample, the person-centered solution mainly organized patients by overall level of PSQI disturbance, and the findings should not be viewed as demonstrating qualitatively different symptom architectures that would automatically justify different management pathways. In this sense, the PSQI should be regarded here as a pragmatic, notional severity measure of subjective sleep quality, analogous to how brief symptom scales such as the PHQ-9 provide a rough indication of depressive-symptom severity rather than demonstrating mechanistic or diagnostic subtypes. The very high mean global PSQI score in the severe stratum (19.16 of 21) also suggests possible ceiling compression within that subgroup: 23 of 103 patients (22.3%) scored the maximum value of 21 and 47 of 103 (45.6%) scored 20 or above, which may reduce sensitivity to within-stratum variation and accentuate between-stratum separation.

Another important observation is the imperfect correspondence between subjective symptoms and objective OSA signs. Although the minimal-disturbance stratum generally showed more favorable PSG and oxygenation indices, subjective burden did not map monotonically onto AHI or oxygenation across every contrast. This symptom–sign discordance is consistent with the idea that patient-reported sleep complaints capture only one dimension of OSA burden and should be interpreted alongside, rather than as a substitute for, physiological severity measures [28–30].

The adjusted correlates of stratum membership were clinically plausible but should be interpreted cautiously. In the severe-versus-minimal comparison, severe OSA, higher depressive symptom scores, higher anxiety scores, and lower social-support scores remained associated with more severe subjective sleep disturbance. These findings are compatible with the idea that physiological severity and psychosocial burden jointly contribute to how untreated OSA is experienced [31–35]. However, several observed odds ratios were modest or imprecisely estimated, and the large number of exploratory comparisons increases the chance of chance findings. For the psychosocial measures, per-point ORs such as 1.12 or 1.16 should not be read as large effects; over a 5-point difference, they correspond roughly to ORs of 1.76 and 2.10, respectively. These magnitudes may be clinically suggestive at the group level but do not define individual treatment thresholds. The obesity signal, for example, was only borderline in the full severe-versus-minimal model but became stronger in the male-only sensitivity analysis, suggesting that precision was limited by sample composition and that some covariate effects may be unstable. Likewise, the previously highlighted occasional-coffee and LSaO₂ findings were not robust once the model was refit using the minimal stratum as the direct reference. Recruitment center was not associated with stratum membership in the refit model, which reduces concern that the main contrasts were driven by a single site, although this should not be interpreted as definitive proof of transportability across settings. These results are therefore better viewed as clues to correlational patterning than as definitive effect estimates. Importantly, the classification-uncertainty-adjusted three-step analysis yielded directionally and inferentially consistent estimates, indicating that the main severe-versus-minimal signal was not driven solely by modal class assignment.

The psychosocial findings also require nuanced interpretation. Depressive symptoms were consistently higher in the more disturbed strata, but the PHQ-9 includes a sleep-related item and was used here as a brief depressive-symptom screen rather than as a diagnostic instrument or a perfectly unidimensional depression measure [36]. We therefore do not interpret the PHQ-9 association as showing that depression independently defines the PSQI-based strata; rather, it indicates co-occurring questionnaire-based symptom burden that partly overlaps with sleep complaints. Shared symptom variance and common-method reporting may therefore contribute to the observed association between PHQ-9 and PSQI [37–39]. Likewise, the SSRS total score combines support availability, perceived support, and support utilization, so lower SSRS scores in the higher-burden strata should not be read as a simple one-dimensional lack of support. In Chinese clinical settings, family structure, help-seeking norms, and patterns of support mobilization may all shape how social support is reported and used. Accordingly, the psychosocial results are best understood as indicating co-occurring symptom burden and social-contextual strain, rather than as isolating a single mechanism.

Lifestyle behaviors also warrant cautious interpretation. Coffee consumption showed an association with the severe stratum, especially for occasional consumption compared with almost none. Given potential reverse causality, it is plausible that patients with poorer sleep and greater daytime impairment used caffeine as compensation, while caffeine timing and dose may also have worsened sleep initiation or quality [40]. Because the present measure did not capture dose or timing, this finding should be treated as hypothesis-generating. Regular exercise was more common in the minimal-disturbance stratum and was associated with lower-burden membership, consistent with literature suggesting beneficial effects of physical activity on sleep and mood [41], but exercise may also proxy for broader health status and functional capacity.

An unexpected inverse association between diabetes and the higher-disturbance strata was present when the mild stratum was used as the reference category in the original model orientation. This finding should not be interpreted as protective. More plausible explanations include selection and referral processes, differential healthcare contact, residual confounding, and sampling variability [42–44]. Reassuringly, when participants with diabetes were excluded, the direct severe-versus-minimal pattern was materially unchanged: severe OSA, depressive symptoms, anxiety symptoms, and lower social support remained associated with severe disturbance. The diabetes result should therefore remain explicitly hypothesis-generating rather than clinically actionable.

This study has several limitations. Its cross-sectional design precludes causal inference, and the clinic-based sample–drawn consecutively from two PSG centers in China and composed predominantly of men–limits transportability to broader OSA populations, including women, community settings, primary care, and home-testing pathways. Residual referral, site-level, and treatment-related confounding cannot be fully excluded because we lacked an AHI < 5 comparison group, detailed exclusion counts, center-specific re-estimated LPA solutions, and medication-level data. In addition, although the hard-assignment results were corroborated by diabetes-exclusion, male-only, and classification-uncertainty-adjusted three-step analyses, we did not fit a full one-step joint mixture-with-covariates model; accordingly, some uncertainty related to class enumeration and residual classification error remains. Multiple exploratory comparisons were conducted without formal multiplicity correction, so nominal P values should be read alongside effect sizes, uncertainty intervals, and consistency across sensitivity analyses. Finally, although the revised Figs. 1 and 2 now display individual-level global and component-specific PSQI distributions, these visualizations remain descriptive and do not remove uncertainty in latent-profile assignment or the need for external replication. Symptom reporting, support mobilization, and referral pathways may also reflect cultural or health-system factors that were not directly measured in the present study.

Future work should therefore prioritize replication in broader and more sex-balanced samples, explicit modeling of site and referral processes, and inclusion of comparison groups beyond diagnosed OSA. Longitudinal designs are needed to clarify temporal relationships among physiological severity, psychological distress, and subjective sleep quality. Methodologically, future studies could compare the present person-centered approach with dimensional alternatives–such as general-distress factor models, two-factor symptom frameworks, or item-response-theory analyses of PSQI items–to evaluate whether the PSQI is better conceptualized as a graded continuum, a multidimensional construct, or both. Analyses that directly model latent-class uncertainty, re-estimate the PSQI-based strata within additional centers, incorporate richer medication and comorbidity data, and extend participant-level distributional visualizations to independent samples would further improve interpretability.

Conclusions

In this multicenter study of adults with newly diagnosed, untreated OSA, latent profile analysis identified four largely graded PSQI-based strata of subjective sleep quality. The results did not support decisively distinct symptom structures beyond severity ordering. More severe subjective sleep disturbance co-occurred most consistently with severe OSA, greater depressive and anxiety symptom burden, and lower social support, while some other associations were less robust to alternative model orientation and sensitivity analyses. The main pattern of findings remained materially unchanged after accounting for recruitment center and classification uncertainty, but the results should still be interpreted cautiously in light of the clinic-based sample, potential selection bias, multiple exploratory comparisons, modest per-point effect sizes, and the predominantly severity-ordered nature of the retained solution.

Abbreviations

AIC

Akaike information criterion

AHI

Apnea–hypopnea index

ANOVA

Analysis of variance

APAP

Automatic positive airway pressure

AUC

The area under the receiver operating characteristic curve

aBIC

Sample-size–adjusted Bayesian information criterion

BIC

Bayesian information criterion

BLRT

Bootstrap likelihood ratio test

BMI

Body mass index

CI

Confidence interval

CPAP

Continuous positive airway pressure

CPSS

Chinese Perceived Stress Scale

DSM-IV

Diagnostic and Statistical Manual of Mental Disorders, Fourth Edition

ESS

Epworth Sleepiness Scale

GAD-7

Generalized Anxiety Disorder-7

LMRT

Lo–Mendell–Rubin adjusted likelihood ratio test

LPA

Latent profile analysis

LSaO₂

Lowest oxygen saturation

MSaO₂

Mean oxygen saturation

OR

Odds ratio

OSA

Obstructive sleep apnea

PHQ-9

Patient Health Questionnaire-9

PSG

Polysomnography

PSQI

Pittsburgh Sleep Quality Index

PSS

Perceived Stress Scale

SD

Standard deviation

SSRS

Social Support Rating Scale

STROBE

Strengthening the Reporting of Observational Studies in Epidemiology

VIF

Variance inflation factor

Authors’ contributions

EL：Conceptualization, Methodology, Formal analysis, Data Curation, Writing—Original Draft and Writing—Review & Editing. FA：Conceptualization, Methodology, Formal analysis，Data Curation, Writing—Original Draft and Writing—Review & Editing. KW：Investigation, Data Curation, Validation. YT：Software, Investigation, Formal analysis. PT：Formal analysis, Data Curation，Conceptualization, Methodology, Writing—Original Draft and Writing—Review & Editing. HW：Conceptualization, Formal analysis, Supervision, Writing—Original Draft and Writing -Review & Editing. BG ：Conceptualization, Formal analysis, Supervision, Writing—Original Draft and Writing -Review & Editing.

Funding

This study was supported by the Science and Technology Project of the Jilin Provincial Administration of Traditional Chinese Medicine (No.2024260), the Changchun University of Traditional Chinese Medicine Theme Case Project (No.2024YJ03), the Shenzhen Key Medical Discipline Construction Fund (No. SZXK062), the 2025 Thematic Case Project of the Development Center for Degree and Graduate Education, Ministry of Education (No. ZT-2510199001), and the Shenzhen Philosophy and Social Science Planning Project (No. SZ2024C018).

Data availability

The datasets generated and/or analysed during the current study are not publicly available due to concerns regarding patient privacy and confidentiality. However, de-identified data may be made available from the corresponding author upon reasonable request.

Declarations

Ethics approval and consent to participate

Ethics approval for this study was obtained from the Ethics Committee of the Affiliated Hospital of Changchun University of Chinese Medicine (Approval No. CCZYFYLL-SQ-2025–289). All procedures were conducted in accordance with the ethical standards of the 1964 Declaration of Helsinki and its later amendments. Written informed consent was obtained from all participants before enrolment. Each participant was fully informed about the study objectives, procedures, potential risks and benefits, and was advised of their right to withdraw from the study at any time without any negative consequences.

Consent for publication

Not applicable.

Competing interests

The authors declare no competing interests.