Obstructive sleep apnea (OSA) is a common sleep disorder characterized by recurrent episodes of collapse of the upper airway during sleep, which is estimated to occur in nearly 1 billion adults aged 30–69 years worldwide and in more than 5 million in Portugal.1 Although commonly observed in clinical practice, OSA which is a heterogeneous condition with various predisposing factors, pathophysiological mechanisms, clinical manifestations and consequences of respiratory events; has not been formally characterized, posing critical challenges to its clinical recognition and resulting in missed or delayed diagnosis.2 A diagnosis is established when a patient has an apnea-hypopnea index (AHI) ≥5 with associated symptoms or an AHI ≥15 regardless of associated symptoms.3 Nevertheless, it has long been recognized that AHI alone does not capture OSA's patients’ heterogeneity, that one-size does not fit all.4 The review of Zinchuk et al., presents numerous studies that used analytic approaches to gain advantage of this heterogeneity.4 In addition, Allan Pack studies discuss P4 medicine and how it might be applicable to OSA, specifically if OSA diagnosis or treatment were personalized, providing several examples of this personalized, predictive, preventative and participatory medicine.5 Cluster analysis, a hypothesis-generating strategy (or unsupervised learning), is being used to identify subtypes of patients with unique characteristics, classifying patients with OSA into smaller and more homogeneous disorder phenotypes. This is a statistical approach that studies the relationships present among groups of participants or variables in a population, aiming at the benefit of allowing more specific diagnosis and treatment strategies. This is accomplished by assessing similarity (or dissimilarity) between subjects using metrics such as correlation or distance based on the features used to characterize each individual. Ideally, each cluster member is as similar as possible to each other and as different as possible from those in other clusters. One approach in cluster analysis is the k-modes algorithm.6 This algorithm extends the k-means paradigm to cluster categorical data by using a simple matching dissimilarity measure for categorical objects, modes instead of means for clustering, and a frequency-based method to update modes in the k-means fashion clustering process.7 The dissimilarity measure can be defined by the total mismatches of the two objects’ corresponding variable categories: the smaller the number of mismatches, the more similar the two objects are. To better understand OSA heterogeneity, this study applies categorical cluster analysis to various observable and measurable OSA characteristics, such as signs, symptoms, demographics, and comorbidities. It also provides a variety of cluster visualizations to communicate this information clearly and efficiently.

A literature review was previously conducted to identify the most relevant OSA variables to be collected from the medical records and a total of 51 variables were noted, such as: demographic variables (e.g., gender); physical examination (e.g., body mass index (BMI)); clinical history (e.g., snoring); and comorbidities (e.g., stroke).

In line with other studies, this research contributes to understanding OSA heterogeneity through exploring possible phenotypes by applying categorical cluster analysis (k-modes) in all severity levels. Our findings confirm that patients’ phenotypes can be identified in the OSA population referred to the sleep laboratory (three in our case). Cluster 1 and 3 were middle-aged males or elderly, while Cluster 2 was middle-aged women. We verified that physical examination aspects, such as BMI and NC, were lower in Cluster 1 but extremely high in Cluster 3, especially NC. Regarding clinical history, the overall percentage of modified Mallampati was placed in the lower levels (Mallampati 1 and 2) and driving sleepiness percentages were low (higher value of 13% in Cluster 1). Alcohol consumption pre-sleep was widely described in Cluster 1 and 3. Reported witnessed apneas were higher in Cluster 3 and 1, while nonrepairing sleep and morning headaches in Cluster 2. All the chosen comorbidities (congestive heart failure, arrhythmias, and pulmonary hypertension) were low and without statistical significance in our analysis. The percentage of the outcome measure (AHI) was demonstrated in each cluster; Cluster 2 had 61% [46%−75%] of mild severity, followed by Cluster 1 (56% [47%−66%]) and Cluster 3, with a significantly smaller proportion (45% [32%−59%]), and high percentage in the severe level (29% [18%−43%]. When we observe the probability of belonging to a cluster given the presence of a factor, we can notice that factors such as male gender, lower levels of age and modified Mallampati, witnessed apneas, nonrepairing sleep, morning headaches, driving sleepiness, alcohol consumption, congestive heart failure and arrhythmias put the patient in Cluster 1. On the other hand, elderly obese adults with increased NC, a level 4 modified Mallampati, and pulmonary hypertension, are allocated to Cluster 3. As a result, when applying our approach to the classification of a new case, physicians should choose the cluster with the greatest number of similar variables states, considering the ten statistically significant variables identified: gender, age, BMI, NC, modified Mallampati, witnessed apneas, nonrepairing sleep, morning headaches, driving sleepiness, and alcohol consumption. The validation cohort of OSA patients, confirmed the previous statistically significant variables as gender, age, witnessed apneas, nonrepairing sleep, morning headaches, and driving sleepiness. Additionally, BMI and NC can be considered in Cluster 2 (P = 0.051) and 3 (P≥0.999), while modified Mallampati and alcohol consumption only in Cluster 2 (P = 0.417 and P = 0.085), respectively.

Several authors have hypothesized the possible presence of OSA phenotypes, with the expected benefit of allowing a more precise diagnosis and treatment strategy, leading to improve patients outcomes, as described in the review of Zinchuk et al..4 The referred studies were performed between 2012 and 2019 (a total of 17), with eight studies applying latent class analysis (LCA) as the clustering method, three applying hierarchical clustering (with multiple correspondences analysis for feature selection or principal component analysis (PCA)), one utilizing hierarchical clustering and k-means, two using k-means (with multiple correspondence for feature selection or PCA), one with k-means, one with time-series analysis and dynamic cluster analysis, and one exploiting k-modes (our previous work). Furthermore, the authors referred that these studies differ in terms of individuals included, sample size, patients features and outcomes, presented in the review. We also would like to point out that while our work chose PSG, as the standard diagnosis tool, other studies elected home studies or PSG mixed with home studies (seven studies). In the studies that selected PSG as the gold standard, different AHI cut-offs were applied, namely one study employed AHI ≥518 and two AHI ≥15,19,20 while two other studies do not define the cut-off.21,22 Analysing our results to the one that selected the same AHI cut-off, we noticed some differences. First, the study collected data from PSG, namely position, sleep state and arousals, while ours did not pre-select these variables as we focused our analysis only on pre-diagnostic data. Also, in our work we analysed the data as categorical while this study was numerical, reaching 6 clusters, all related to sleep position or sleep state. The only common aspect is that, like our study, it did not report outcomes. The two studies that reported a cut-off higher than 15 events per hour collected data from PSG in a supine position (not collected in our study) or symptoms and ESS. In this case, LCA was performed, and 4 clusters were found: 1) disturbed sleep; 2) minimally symptomatic; 3) excessively sleep, and 4) moderately sleepy. Although we could compare our results to this study, it is not our aim to assess the association with prevalent and incident cardiovascular disease. The main limitation of cluster analysis is that the phenotypes are always dependent on the number and the quality of the selected variables. Nevertheless, we believe that the process and the results are relevant. The inclusion of a comprehensive number of risk and diagnostic factors, as well as the observed robustness of the cluster-defined phenotypes in the validation cohort, enhances our understanding of OSA heterogeneity, emphasizing that the stereotypical OSA patient needs to be redefined.