Nocturnal sleep phenotypes in idiopathic hypersomnia – A data-driven cluster analysis ☆

Introduction: The diagnostic process for idiopathic hypersomnia (IH) is complex due to the diverse aetiologies of daytime somnolence, ambiguous pathophysiological understanding, and symptom variability. Current diagnostic instruments, such as the multiple sleep latency test (MSLT), are limited in their ability to fully represent IH ’ s diverse nature. This study endeavours to delineate subgroups among IH patients via cluster analysis of poly-somnographic data and to examine the temporal evolution of their symptomatology, aiming to enhance the granularity of understanding and individualized treatment approaches for IH. Methods: This study included individuals referred to the Uppsala Centre for Sleep Disorders from 2010 to 2019, who were diagnosed with IH based on the International Classification of Sleep Disorders-3 (ICSD-3) criteria, following a thorough diagnostic evaluation. The final cohort, after excluding participants with incomplete data or significant comorbid sleep-related respiratory conditions, comprised 69 subjects, including 49 females and 20 males, with an average age of 40 years. Data were collected through polysomnography (PSG), MSLT, and standardized questionnaires. A two-step cluster analysis was employed to navigate the heterogeneity within IH, focusing on objective time allocation across different sleep stages and sleep efficiency derived from PSG. The study also aimed to track subgroup-specific changes in symptomatology over time, with follow-ups ranging from 21 to 179 months post-diagnosis. Results: The two-step cluster analysis yielded two distinct groups with a satisfactory silhouette coefficient: Cluster 1 (n = 29; 42 %) and Cluster 2 (n = 40; 58 %). Cluster 1 exhibited increased deep sleep duration, reduced stage 2 sleep, and higher sleep maintenance efficiency compared to Cluster 2. Further analyses of non-clustering variables indicated shorter wake after sleep onset in Cluster 1, but no significant differences in other sleep parameters, MSLT outcomes, body mass index, age, or self-reported measures of sleep inertia or medication usage. Long-term follow-up assessments showed an overall improvement in excessive daytime sleepiness, with no significant inter-cluster differences. Conclusion: This exploratory two-step cluster analysis of IH-diagnosed patients discerned two subgroups with distinct nocturnal sleep characteristics, aligning with prior findings and endorsing the notion that IH may encompass several phenotypes, each potentially requiring tailored therapeutic strategies. Further research is imperative to substantiate these findings.


Introduction
Idiopathic hypersomnia (IH) is a sleep-wake disorder characterized by excessive daytime sleepiness (EDS) and sleep inertia despite adequate nocturnal sleep.According to the International Classification of Sleep Disorders, Third Edition (ICSD-3), the diagnostic criteria for IH include the presence of EDS for at least three months, the absence of cataplexy, and confirmatory objective sleep testing that does not show sleep-onset REM periods.This testing must include a mean sleep latency below 8 min on the multiple sleep latency test (MSLT) and/or a total sleep duration of more than 660 min per 24 h, as estimated by 24-h PSG or 1week actigraphy/sleep logs [1].
However, EDS, the main symptom of IH, is prevalent in the general population, with rates reaching up to 33 %, though the prevalence can vary widely depending on the population studied and the assessment methods used [2,3].EDS can result from various causes, including insufficient sleep due to factors such as shift work and circadian rhythm sleep-wake disorders, physical conditions like obesity, sleep apnea, diabetes, and gastroesophageal reflux, as well as psychiatric disorders, particularly mood disorders like depression and bipolar disorder [4,5].While low or absent cerebrospinal fluid hypocretin-1 is highly sensitive and specific for diagnosing Type 1 narcolepsy [6,7], the absence of other reliable biomarkers, the unclear underlying pathophysiology, and the variability in clinical manifestations across other hypersomnolence disorders complicate differential diagnosis [8].Many patients exhibit clinical symptoms that are not specific to IH, and the diagnosis often relies on excluding other potential causes [8][9][10][11].This is problematic as numerous somatic and psychiatric conditions, such as insomnia, circadian rhythm disorders [12], narcolepsy, affective disorders [13,14], behavioural sleep deprivation [11], and ADHD [15], share symptoms with IH and may be comorbid.Additionally, other symptoms of IH, such as sleep inertia, prolonged night-time sleep, cognitive impairment, and functional consequences, have been rarely assessed in clinical studies due to the recent availability of validated instruments [16].
Given these complexities, management strategies for IH have largely been derived from expert consensus, with a reliance on case series and infrequent randomized controlled trials [29].There is also a need to develop non-pharmacological approaches, such as cognitive behavioural therapy [30].Previous data-driven cluster analyses in central hypersomnolence have primarily focused on differentiating between narcolepsy with and without cataplexy and other hypersomnolence disorders [31,32].One exception is a study where patients with other known reasons for hypersomnolence, such as narcolepsy and obstructive sleep apnoea, were excluded before clustering [33].
In the present study, we address the following questions.
1. Can relevant data-driven subgroups be distinguished within the patient group diagnosed with IH using cluster analysis on polysomnography variables?2. If relevant subgroups emerge, do the symptom profiles change in a diverse manner over a longer follow-up period, as measured by subjective rating scales?

Population
The population sample consisted of individuals diagnosed with IH who were referred to the Uppsala Centre for Sleep Disorders between 2010 and 2019.All diagnoses were initially made by experienced sleep specialists.Out of 225 individuals invited to participate, 91 provided informed consent.The study was approved by the Swedish Ethical Review Authority (Dnr 2021-03458), ClinicalTrials.govID: NCT05150977.

Inclusion criteria
After obtaining informed consent, medical records were re-evaluated by A.M. and P.C.B.We included only patients diagnosed with IH according to ICSD-3 diagnostic criteria at our sleep centre.The diagnostic process involved a semi-standardized medical history, a sleep log, videopolysomnography (vPSG) following American Academy of Sleep Medicine (AASM) standards, and a subsequent multiple sleep latency test (MSLT) [1].The diagnosis of IH requires a combination of positive and negative features: the absence of sleep-onset rapid eye movement sleep periods (SOREMPs), cataplexy, hypocretin deficiency, or other conditions that better explain the symptoms, and hypersomnolence must be objectified by one of three criteria according to ICSD-3: (1) an MSLT sleep latency less than or equal to 8 min with fewer than two SOREMPs; (2) at least 660 min (11 h) of sleep time measured over up to 24 h of PSG monitoring after correcting chronic sleep deprivation; or (3) at least 660 min of sleep time per 24 h, estimated by wrist actigraphy averaged over at least one week of ad libitum sleep.All patients included in this study were diagnosed according to these criteria; however, sleep duration and exclusion of habitual sleep deprivation were evaluated using sleep diaries rather than actigraphy.The majority of participants received their diagnosis based on MSLT results, although 10 participants (14,5 %) exhibited a mean sleep latency of more than 8 min but fulfilled the 660-min requirement for prolonged nighttime sleep.Notably, none of the patients demonstrated sleep onset REM during nocturnal polysomnography.
Twenty-two participants were excluded, either due to missing relevant data (n = 12) or the presence of a significant sleep-related respiratory disorder (n = 10).Thus, data from 69 participants, 49 women and 20 men with a mean age of 40 years (SD = 11.2, range = 24-66) were used in the analyses.

Template
Items from the medical journals were surveyed according to the following template.
(A) Demographic parameters: Gender, age at examination, weight, height, body mass index (BMI), medications at the time of vPSG and MSLT.The diagnostic protocol in our clinic required patients to discontinue the use of pure stimulants, such as modafinil or lisdexamfetamine, 14 days prior to undergoing polysomnography, while maintaining stable doses of all other medications.Medications were categorized as follows: no medication, medications without central nervous system (CNS) effects (e.g., antihypertensive drugs or L-Thyroxine), stimulating drugs (e.g., SSRIs or SNRIs), sedating drugs (e.g., benzodiazepines, opioids, or gabapentin), or a combination of both stimulating and sedating drugs. of SOREMPs (0 or 1).(D) Standardized questionnaires: Epworth Sleepiness Scale (ESS) [34]; Insomnia Severity Index (ISI) [35]; Montgomery Åsberg Depression Rating Scale (MADRS) [36] Diurnal Type Scale (DTS) [37].The presence of sleep inertia was determined by the response to the question, "How significant are your difficulties with becoming fully awake in the morning?" in an 87-item [38] general sleep questionnaire, with severe or very severe indicating sleep inertia.
For the second research question, responses from the ESS, ISI, and MADRS assessments at the time of diagnosis were used, with a follow-up conducted for these assessments.Additionally, the following questionnaires were administered only during the follow-up assessment: Ullanlinna Narcolepsy Scale (UNS) [39] Sleep Problem Acceptance Questionnaire (SPAQ) [40], Adult ADHD Self-Report Scale (ASRS-v1) [38] and Brief Pain Inventory (BPI) [41].

Cluster analytic approach
We assumed heterogeneity in sleep parameters within the diagnostic group of IH.Identifying such heterogeneity is complex, but cluster techniques can effectively address this challenge [42][43][44].Cluster-guided classification techniques, which are hypothesis-independent methods, group variables into mathematically definable and homogeneous subsets.These techniques have proven valuable for creating a more robust hypersomnolence classification system [31].For this study, a two-step cluster analysis was chosen because it can effectively handle clinical longitudinal data and is a hybrid between hierarchical and non-hierarchical cluster analysis [45].We used log-likelihood distance measure and Schwarz's Bayesian Criterion (BIC) as validation measure.Statistical analyses were conducted using SPSS v.28 (IBM SPSS Statistics).
Considering the above constraints, the time spent in various sleep stages was selected for cluster analysis, along with sleep maintenance efficiency (SE, %).In this context, SE is defined as the ratio between TST and SPT, quantifying how effectively a person maintains sleep once they have initially fallen asleep, thus excluding the time taken to fall asleep.Consequently, five variables were included in the cluster analysis: time spent in REM sleep and stages N1, N2, and N3, as well as SE.These variables were chosen for their diagnostic utility, lower risk of multicollinearity, and clinical and theoretical importance.Although correlations between some variables were significant (N1 and N2, REM and N3, N2 and N3, and N3 and SE), as shown in Table 1, none exceeded the threshold (r < 0.5), and the Variance Inflation Factor (VIF) was below 1.34, indicating no multicollinearity issues.

Additional statistical analysis
To investigate the subgroups identified through cluster analysis, research question two was examined using a comprehensive statistical approach.Four primary methods were employed: repeated measures ANOVA, independent t-tests, chi-squared tests, and multiple regression analysis.These analyses were conducted using SPSS v.28 (IBM SPSS Statistics) and JASP version 0.16.3.Each method was selected to address specific aspects of the research question: Repeated Measures ANOVA: This analysis assessed differences over time between clusters, focusing on the ESS, ISI, and MADRS scores.The independent variable was time (diagnosis and follow-up) and cluster affiliation, with the dependent variables being the scores on the selfassessment forms.
Chi-squared test: This test examined potential differences in ESS remission, prevalence of sleep-onset REM, medication usage, and the presence of sleep inertia.
Multiple regression analysis: Three analyses used significant cluster variables (N2, N3, SE) to predict long term changes in ESS, ISI, and MADRS scores, examining whether cluster variables could predict changes in daytime sleepiness, insomnia, and depression.

Cluster analysis
The exploratory two-step cluster analysis resulted in two clusters with acceptable cluster cohesion and separation, as indicated by a silhouette coefficient (0.3) in the "fair" range, see the left panel (A) in Fig. 1.Cluster 1 consisted of 29 participants (42 % of the sample), while Cluster 2 comprised 40 participants (58 %), with a cluster ratio of 1.38.Evaluation of the agglomeration coefficient across cluster combinations, when using hierarchical cluster analysis only, supported either a 2 or 3 cluster solution.The suggested two-cluster solution was deemed to be more valid in comparison to a three-cluster solution in which cluster 2 was split into a very small (n = 3) and a bigger (n = 37) one while cluster 1 remained intact (largest-to-smallest ratio 12.33, silhouette coefficient 0.4).The two-cluster solution was the same when using standardized and non-standardized variables.
The variable with the greatest predictive importance in determining the two clusters was time spent in deep sleep (N3), followed by N2, SE, N1, and REM, which had the lowest predictive weight, see the right panel (B) in Fig. 1.For three of the variables (N1, N3, and SE), heterogeneity in variance was found according to Levene's test, which was accounted for in the significance tests.Based on independent t-tests and descriptive statistics, Cluster 1 was characterized by longer time spent in N3, higher SE, and shorter time spent in N2 compared to Cluster 2, see Table 2.
When comparing the two clusters on other clinical/descriptive variables that were not used for clustering (for details see Table 3), it was found that individuals in Cluster 1 had significantly less WASO and a significantly higher prevalence of one SOREMP.
The clusters did not differ significantly on other objective or subjective variables.In particular, there were no significant differences in the presence or absence of sleep inertia or in the intake of medication between the two clusters.

Follow up
Upon identifying relevant subgroups, Research Question 2 assessed wheteher symptom profiles in these subgroups changed over time using subjective rating scales (ESS, ISI, MADRS).Data were collected at diagnosis and followed up in October 2022, with follow-up duration ranging from 21 to 179 months.Independent t-tests confirmed no significant differences in follow-up duration between clusters or any effect of follow-up time in subsubsequent ANOVAs.Repeated measures ANOVA indicated a significant main effect of time on ESS, suggesting improvement (F = 7.00; df = 1.48; p = 0.01; η 2 = 0.13).However, no significant differences were found between clusters 1 and 2 in terms of changes on ESS, as indicated by a non-significant Group × Time interaction (F = 0.28; df = 1.48; p = 0.60; η 2 = 0.006), see Fig. 2. Chi-square tests revealed no significant differences in remission rates of daytime sleepiness between the clusters (χ2 = 0.02, df = 1, p = 0.90).
For ISI, neither a main effect of time nor a significant Group × Time interaction was observed, indicating uniform but non-significant changes across clusters.In contrast, MADRS showed a significant worsening over time (F = 5.65; df = 1.44; p = 0.02; η 2 = 0.11), with no differences between clusters.Follow-up questionnaires, including UNS, SPAQ, ASRS-v1, and BPI, administered solely at follow-up, showed no significant between clusters, with small effect sizes across 10 to 32 participants per cluster.
Multiple regression analysis assessed whether specific cluster variables could predict changes in ESS, ISI, and MADRS scores from diagnosis to follow-up.Notably, no variables significantly precicted changes in ESS, though SE showed a marginally significant result, see Table 4. Higher SE was associated with lower ISI and MADRS scores, suggesting an inverse relationship.Conversely, increased N3 duration was positively correlated with ISI scores, indicating more time in this stage correlated with higher scores.Multicollinearity was not an issue, as the VIF did not exceed 1.30 for any variable.

Discussion
The exploratory two-step cluster analysis yielded two distinct groups with acceptable cluster homogeneity: Cluster 1 (n = 29; 42 %) and Cluster 2 (n = 40; 58 %).Cluster 1 exhibited increased N3, reduced N2, and higher SE, indicating greater sleep pressure compared to Cluster 2. Further analyses of non-clustering variables indicated shorter WASO in Cluster 1.However, there were no significant differences in other sleep parameters, body mass index, age, or self-reported measures of sleep inertia or medication usage.While sleep latency in the MSLT was similar between clusters, the prevalence of one sleep onset REM period was higher in Cluster 1, suggesting an overlap with NT2.
This finding aligns with the observations of Fronczek and colleagues [46].Furthermore, in the first data-driven clustering analysis, Sonka and co-workers differentiated three main clusters and concluded that narcolepsy type 1 and polysymptomatic hypersomnia are independent sleep disorders.In contrast, NT2 and IH with long sleep formed a single cluster, referred to as "combined monosymptomatic hypersomnia/narcolepsy type 2" [32].The ICSD-3 classifies three primary, non-recurrent central disorders of hypersomnolence: IH, NT1 and NT2.Unlike NT1, which can be convincingly explained by hypocretin

Table 2
Mean values and comparative statistics (t-test) for the five variables that were included in the cluster sample.deficiency [47], NT2 and IH lack known pathophysiology and validated, reliable biomarkers [46].In a previous clustering analysis by Cook and co-workers on 62 patients diagnosed with IH, two clusters similar to those identified in our study were found: one cluster with higher sleep pressure, prolonged sleep phases, and deeper sleep, and another with shallow sleep and lower sleep efficiency [33].However, in contrast to this we did not see differences in self-reported depressive symptoms or use of psychotropic substances.It remains unclear, whether drug use is a confounding factor or reflects different pathomechanisms [5].Additionally, Gool et al. (2022), using agglomerative hierarchical clustering on 97 variables from the European Narcolepsy Network database in a sample of 1078 unmedicated patients with hypersomnolence, identified seven distinct clusters.In four clusters, NT1 was very probable or definite in a large proportion of the subjects.Three clusters were more heterogeneous and included about 30 % of patients diagnosed with definite NT2.Clusters 5 and 6 contained roughly 30 % of subjects diagnosed with possible to definite IH.Although our data did not include other patients than those diagnosed with IH, the similarities between our two clusters and clusters 5 and 6 in their study are striking [31].
An important unresolved question in IH is the significance of distinguishing between those with and without increased sleep pressure [48].Long and short sleep duration were used as markers in the second edition of the International Classification of Sleep Disorders [24] but were subsequently removed in the third edition [23].Habitual long sleep of 10 h per night or longer is associated with difficulty waking and persistent brain fog.In line with previous clustering studies, our results support the distinction between hypersomnolence with increased sleep pressure and often habitual long sleep (quantitative hypersomnia) and hypersomnolence resulting from a qualitative sleep problem.The latter group might include or be confounded wit ADHD [15], atypical depression, or circadian disorders, and may require different treatment approaches.The ambiguity in treatment studies might be due to this heterogeneity.The second research question aimed to determine if symptom profiles change in a diverse manner over a longer follow-up period, as measured by subjective rating scales in the identified subgroups.Long-term follow-up assessments revealed a slight, though clinically insignificant, reduction in in excessive daytime sleepiness, with no significant differences observed between the clusters.Additionally, the lack of change in the ISI -a tool primarily validated for measuring insomnia but also capable of assessing the impact of a "sleeping problem" on daytime functioningsuggests that the effect of hypersomnia on daytime functioning remained largely unchanged over time in our study population.The value of this observation is limited by the lack of data on treatment between visits and treatment data at follow-up in our database.
This study demonstrates some strengths.It employs a data-driven approach using cluster analysis to identify subgroups within IH, offering valuable insights into the disorder's heterogeneity.The comprehensive assessment includes objective sleep measures such as vPSG and MSLT, as well as subjective rating scales.The identification of clinically relevant clusters, characterized by different sleep parameters, suggests potential for tailored treatment approaches.The study adheres to established diagnostic criteria and integrates its findings with previous research, enhancing its validity and relevance.However, the study also has several notable limitations.The small sample size (n = 69) limits the generalizability of the findings.Collecting data from a single sleep disorders centre may introduce centre-specific biases.The cross-sectional design, with follow-up data collected at a single time point, limits understanding of the temporal dynamics of symptom changes and treatment effects.While longitudinal follow-up provides insights into symptom progression over time, the absence of detailed treatment data between visits and at follow-up restricts the ability to assess the impact of various treatments on symptom progression.Moreover, potential confounding factors, such as the use of psychoactive medications, were not fully accounted for, which could influence the results.Measurement limitations, such as relying on sleep diaries instead of actigraphy for sleep duration, may affect the accuracy of the findings.Finally, the limited statistical power due to the small sample size and the lack of consistent long-term follow-up data further constrain the findings.Addressing these weaknesses in future research could provide more robust and generalizable insights into IH subtypes and their treatment needs.

Conclusion
Including this study, four data-driven studies [31][32][33] have identified clusters that separate IH into at least two distinct entities: one characterized by increased sleep pressure and the other by qualitative sleep disturbances.Further prospective research is needed to determine whether these subgroups require different treatment strategies.We propose that future studies aimed at developing individualized treatment approaches should prioritize wake-promoting agents for IH patients exhibiting increased sleep pressure, while focusing on pharmacological and behavioral interventions to enhance sleep quality in patients demonstrating impaired nocturnal sleep.

Funding sources
This research did not receive any specific grant from funding agencies in the public, commercial, or not-for-profit sectors.

Fig. 1 .
Fig. 1.Cluster solution.The left panel (A) displays the results of the exploratory two-step clustering approach, resulting in two clusters.The right panel (B) shows the relative predictive importance of the five input variables in distinguishing between the clusters, with time spent in deep sleep (N3) being the most important.Note.Abbreviations: N1 = Non-Rapid Eye Movement Sleep Stage 1; N2 = Non-Rapid Eye Movement Sleep Stage 2; N3 = Non-Rapid Eye Movement Sleep Stage 3; SE = Sleep Maintenance Efficiency; REM = Rapid Eye Movement Sleep.

Fig. 2 .
Fig. 2. Average score in Epworth Sleepiness Scale (ESS) at the time for diagnosis, i.e. initial assessment, and follow-up within cluster 1 (blue, squares) and cluster 2 (red, circles).(For interpretation of the references to colour in this figure legend, the reader is referred to the Web version of this article.)

Table 1
Correlation between cluster variables.

Table 3
Mean values and comparative statistics (t-test) for the profile variables that were included in the profile analysis.