Coronavirus disease 2019 (COVID-19) is a respiratory disease caused by severe acute respiratory syndrome coronavirus (SARS-CoV-2). This novel coronavirus was first reported in Wuhan, China, in December 2019 and quickly spread worldwide (1). Belonging to the same coronavirus family as SARS-CoV and MERS-CoV, SARS-CoV-2 also exhibits a remarkable infectivity power (2). Immediate practices have been taken to allocate medical resources and plan for treatments. However, they might not be as effective as expected due to a lack of knowledge about SARS-CoV-2. A better understanding of the pathogenesis of COVID-19 and its different clinical phenotypes and risk groups is essential to address the immunopathology of the infection. The accumulated observational data on COVID-19 positive patients available to date serve as valuable resources. To probe this large amount of clinical data, supervised machine learning approaches have been applied to the diagnosis and prognosis of COVID-19, risk stratification, and the prediction of different outcomes (3–9). Unsupervised machine learning is an alternative approach that does not require specific labels. It avoids using preconceived knowledge or assumptions that may be subjected to biases and unknown confounders. To this end, unsupervised clustering techniques are often used for exploratory analysis to probe the underlying patterns within big data sets, enabling identification of latent clinical phenotypes and potentially deriving novel insights from the associated correlations. For example, it has been applied to derive the phenotypes of COVID-19 patients using electronic health record (EHR) data on admission of 6,000 COVID-19–positive adults at the Mount Sinai Health System in New York in the United States (10), 413 patients from an individual-level published study (11), and 213 patients in Wuhan Pulmonary Hospital (12). Hong Kong's health system is unique when dealing with the COVID-19 pandemic. Due to a government-wide policy, all COVID-19–positive patients were admitted to public hospitals regardless of their disease severity or symptoms. Therefore, the data in Hong Kong can capture the varying presentations of COVID-19. This study aims to use unsupervised clustering analysis to discover different phenotypic presentations across the disease spectrum of COVID-19 in Hong Kong based on demographics and laboratory information on admission to hospital.

Initial data retrieval yielded a total of 8,562 patients. After data preparation to deal with missing variables and inconsistent data, a final number of 7,606 patients was kept (see Figure 1). The demographics and clinical characteristics are shown in Table 1.

Four clusters were identified using demographics and laboratory variables on admission. A separate analysis was done on the deceased cohort (see Table 2). The clusters were compared with one another. Each cluster was compared with the derived population and the most used normal reference range in the cohort. Each cluster was also compared with the deceased cases. A value was marked “high/low” if still within the reference range but considered close to the cutoff values. A value was marked “elevated/reduced” if it was outside the normal reference range (see Table 2, Figure 2; Supplementary Table 3).

Unsupervised clustering was used to probe the latent phenotypes of 7,606 COVID-19–positive patients in Hong Kong across 2020. Based on age, sex, and 20 laboratory blood tests on admission, one of the four generated clusters aggregated 86.6% of deceased cases without prior information of their clinical outcomes. The clinical characteristics of this cluster, including the oldest median age; highest aggregated comorbidity; and the laboratory tests of hemoglobin, hematrocrit, and lymphocyte were observed to be in the lower range of normal, whereas neutrophil, C-reactive protein, and lactate dehydrogenase were observed to be in the higher range of normal. Findings are comparable with poor prognostic features based on contemporaneous literature. Notably, when applied to a separate validation cohort, cluster 4 was still able to identify seven out of eight (87.5%) deceased cases. A potential clinical utility may be to call for early medical attention and resource to patients that belong to cluster 4 at the initial diagnosis at hospitalization. The clinical characteristics of cluster 4 and deceased cases in this study aligned with previous findings (23–25). First, old age reflects a weaker immune system to fight against pathogens, including pneumonia (26). It is a well-established risk factor correlated with chronicity and comorbidity in COVID-19 prognosis (10, 27). On the other hand, increased neutrophil count and high neutrophil-to-lymphocyte ratio can be elicited by deep airway and alveolar damage. They represent an acute inflammatory response and are indicators of a poor prognosis and higher disease severity for COVID-19 (28–31). In line with cluster 4 and the deceased cases in this study, at an early stage of COVID-19, lymphocyte counts typically decrease, whereas white blood cell count may or may not decrease (32). In addition, patients with severe COVID-19 outcomes have exhibited abnormally high C-reactive protein levels, lactate dehydrogenase, and neutrophils, implying an inflammatory response and severe pneumonia (33). Although C-reactive protein level is not affected by physical status, age, and sex (34), it may be used to early diagnose severe pulmonary disease secondary to bacterial infection (35).

A greater proportion of males than females were infected with COVID-19 in contemporary literatures as well as MERS-CoV and SARS-CoV infection (24, 36, 37). It can be a cofounder in evaluating the sex-based difference in the susceptibility of COVID-19 (38). In this retrospective study, the proportion of males and females was roughly equal across all age groups (see Figure 4) on admission. No significant difference in disease prevalence existed between males and females. Under this condition, a higher proportion of males was still observed in cluster 4 (65.0%) and deceased cases (61.3%). It indicates being male as a risk factor for mortality (37, 39, 40). This reduced disease susceptibility of females is hypothesized to be related to the major roles of the X chromosome and sex hormones in innate and adaptive immunity (41). Nonetheless, this correlation needs further exploration and investigation.

The complications of COVID-19 include severe pneumonia, septic shock, acute respiratory distress syndrome (ARDS), and multiple organ failure (31, 42). Although lung has been the primary organ involved in the infection, elevated creatinine and urea levels were also detected in cluster 4 and in the deceased cases, which implies that impaired renal function was a potential early indicator of poor prognosis. Additionally, a low albumin level, reflective of liver synthetic function, was also observed. These abnormal laboratory variables imply that indicators for early multiorgan failure were important and is likely a reflection of systemic inflammatory response. For the reduced hemoglobin and hematocrit levels detected in cluster 4 and the deceased, a prior study suggests that SARS-CoV-2 may attack the heme on 1-beta chain of hemoglobin through CD147, and CD26, ACE2 receptors or by simulating hepcidin to increase tissue ferritin, block ferroprotein, causing the iron deficiency, and thus lower hemoglobin level (43).

In previous clustering analysis (10–12), both three and four clusters have been identified, which were similar to the number chosen by this study. Previous clusters have also found some correlations for a poor prognosis, such as old age and high comorbidity scores (12, 28), being male, high lymphocytes high neutrophil count (11), and albumin level (12). In respect to the clinical outcomes, our study shows the greatest power to differentiate a cluster by aggregating 86.6% deceased cases and capturing most of their clinical characteristics and correlations.

There were several key strengths to our study. First, a large sample size of 7,606 COVID-19–positive patients was included. This adds to greater statistical power and generalizability. The data were collected from a single integrated EHR system fully covering the Hong Kong population, meaning the clinical data set is representative of this cohort. Second, all patients that were tested positive despite mild or no symptoms were nevertheless hospitalized in Hong Kong. Our study is unique capturing a wide spectrum of clinical severity. On the contrary, most other countries only tend to admit patients with severe symptoms to the hospitals. Third, we validated our clusters on a separate temporal validation with further feature explanation using SHAP. It was demonstrated that criteria from PCA-based k-prototype clustering were correctly captured, used, and inferred. Granular analysis of these distinct clusters should give a better understanding of different clinical representations of COVID-19 on admission. Besides this, delineating interactions between specific chronicity and poor prognosis regarding different clinical groups can help understand the mechanism of COVID-19.

There were a few limitations to our study. First, for a retrospective study using observational data, the sampling was not random. There may be the potential of selection and recall bias. Although all age groups were represented, they were not evenly distributed. The results and interpretations will be best applied to the Hong Kong cohort with race dominated by patients of Chinese ethnicity although there were several imported cases of different ethnic origins. Owing to a lack of accurate data on ethnicity, we could not perform further sub-analysis. Second, as with any large-scale studies covering multiple hospitals, there may be different standards and measurement protocols. It was observed that the normal reference ranges of laboratory variables varied mildly across institutions and according to age and sex. Because no significant discrepancy was detected, the data were put together directly using their original values. Although a linear mapping could have been used to reduce the systematic noise, variations due to age and sex need to be further distinguished. Finally, some potential important clinical and laboratory data with more than 50% of missing data were excluded from the study. They include BMI, blood pressure, oxygen saturation, D-Dimer, need for assisted ventilation, ITU admission, etc., which are potentially useful for severity assessment and prognostication. These may additionally aid prediction and more accurate allocation of medical resources in the future. Owing to a lack of testing or accurate documentations, these were not able to be included in this large scale multi-institutional study. These variables are recommended to be documented or performed in the future. However, they were not routinely recorded nor taken in usual clinical practice in Hong Kong at the time of the study period.

Unsupervised clustering was used to probe the latent phenotypes of 7,606 COVID-19–positive patients in Hong Kong across the year 2020. Based on age, sex, and 20 laboratory variables on admission, one of the four generated clusters aggregated 86.6% deceased cases without prior knowledge of their clinical outcomes. Further understanding of the different COVID-19 clinical phenotypes may pave the way for more individualized patient risk stratification and treatment.

Data can be provided upon reasonable request and inquiries can be directed to the corresponding author

The studies involving human participants were reviewed and approved by Cluster Research Ethics Committee/Institutional Review Board (REC/IRB). Written informed consent for participation was not required for this study in accordance with the national legislation and the institutional requirements.

KL, K-SN, and VV were involved in study conception, statistical analysis, and drafting of the initial version of the manuscript. K-SN and VV were involved in data collection. K-WK, KT, C-FS, C-WL, and VV supervised the work and offered significant intellectual contribution. All authors discussed the data analysis, results, and approved the final version.

The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.

All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article, or claim that may be made by its manufacturer, is not guaranteed or endorsed by the publisher.