The first outbreak of an acute respiratory disease caused by a novel coronavirus (SARS-CoV-2) was identified in Wuhan, China in December 2019. Following the continued spread of the coronavirus across several countries, the World Health Organization (WHO) declared a pandemic on 11^th March 2020 (1). As of 26^th March 2022, Canada has reported 3.45 million confirmed cases and 37,449 deaths (2). Globally the COVID-19 pandemic has resulted in approximately 479 million cases and 6.12 million deaths (1).

In Alberta, as of 26^th March 2022, approximately 536,166 confirmed cases and 4,044 deaths have been reported. Although most COVID-19 cases have presented with mild symptoms, there has been a relatively low rate of hospitalization (4.1%) and ICU admission (0.7%) noted in the province due to improving vaccination rates. However, there remains a critical need for understanding the epidemiological factors and clinical characteristics associated with severe outcomes in the planning and management of the disease.

Studies have reported that age, sex, smoking status, underlying health conditions, and clinical presentation including breathing difficulties, fever and cough are important risk factors associated with severe outcome(s) (3–8). Despite an increase in well-conducted prediction analysis with regression modeling, a direct translation of the research findings into patient care remains challenging. Decision tree modeling is a promising strategy for risk predictions based on a set of sequential rules. Decision tree models are intuitive, easy to interpret by non-statisticians and offer a new way to visualize complex datasets (9). This approach has demonstrated enhanced clinical utility in identifying high-risk individuals and aiding in subsequent intervention measures (10). Also, recent literature demonstrate that machine learning approaches have successfully predicted COVID-19 infection (11, 12) and outcomes based on clinical factors (disease severity), epidemiological factors age (13), sex, socioeconomic status, public health measures (lockdowns, restriction on gatherings, quarantine, social distancing measures), underlying health conditions, and presenting symptoms (14, 15).

Finally, the province of Alberta has experienced five waves of COVID-19 to date, with high numbers of new infections and deaths occurring in each wave. It is expected that there will be additional waves. It is, therefore, important to study the factors associated with severe outcomes to determine which factors are the most important to target for prevention and mitigation. Therefore, the aim of this study was to carry out a decision tree modeling approach to identify the risk factors associated with severe outcomes among COVID-19 patients in Alberta, Canada between March 2020 and May 2021 to support future clinical decision making and planning.

Approximately 167,000 cases were identified in the study period prior to the spread of the Delta variant in Alberta. After excluding pediatric cases, a total of 140,182 adult cases who tested positive for COVID-19 were considered for further analysis. About 95% of cases recovered without any severe outcome. However, 6.37% were hospitalized, 1.39% were admitted to ICU and 1.57% died due to the disease. The frequency of variables used in the modeling is described in Table 1 including demographics, smoking status, underlying symptoms, and pre-existing conditions.

All variables (except deprivation index and ethnicity) were tested using for the ordinal outcome of disease severity (admission to hospital or ICU, or death) in the decision tree models. There were data quality issues with ethnicity information and missing deprivation index values for certain postal codes. Postal codes for long term care facilities/senior care facilities, which had higher case numbers and mortality rates, were excluded from the deprivation index value calculations. Therefore, we could not assign a deprivation index to cases that may have been residents of elder care facilities. These two variables were excluded in decision tree modeling to minimize bias in the models.

In this population-based retrospective study of 1,40,182 cases of COVID-19, about 7% experienced severe outcomes such as hospitalization, ICU admission or death. We have adopted the Charlson comorbidity index approach to group the underlying conditions and use in the model to evaluate the association with COVID-19 disease severity. Similar approach has been adopted in other studies to identify the comorbidities associated with COVID-19 in different populations (22–24).

There is emerging literature evidence related to implementing machine learning approaches to understand the pattern and key contributors to COVID-19 infections and prognosis. This current study's findings that included one of the largest population-based cohorts and explored the predictors of specific outcomes such as hospitalization, ICU admission or death are consistent with literature (8, 25).

In the CTREE model including cases in all age groups, the key factors influencing severe outcomes were age, breathing difficulties, and the presence of pre-existing conditions, including cardiovascular disease, neurological conditions, diabetes, hypertension, and renal disease. Similarly, the random forest model identified that age and breathing difficulties were the two primary risk factors in predicting poor prognosis accounting for up to 28% and 22% of outcomes, respectively. Cardiovascular disease (8%) and neurological conditions (7%) were also noted as key predictors of outcome. About 65% of the outcomes could be predicted based on these four factors. The presence of diabetes (5%), hypertension (5%) and renal disease (4%) together with the top four factors predicted about 79% of the outcomes. Therefore, these variables are key to developing a decision rule. The results from the model can help clinicians quickly identify and prioritize high risk patients and allocate health care resources more effectively.

We also adopted age-stratified modeling to get a deeper insight into the risk factors within specific age groups. We consistently noticed that irrespective of age-based stratification, that breathing difficulties and age were critical in predicting severe outcomes. Interestingly, among the younger age group (18–40 Y) obesity was among the top predictors, consistent with literature findings (26) and adding critical knowledge. Diabetes, cardiovascular and neurological conditions were identified as predictive across the age stratified models.

Our findings are consistent with the current literature, where the majority of poor outcomes are noted among the older populations (3, 8, 25) and individuals with underlying conditions (6). A model based on age and other clinical variables predicted death among COVID-19 patients in China (8). Population prevalence of neurological conditions, which includes cerebrovascular disease, dementia, paraplegia, the chronic neurological disorder is higher among older adults (27). This is consistent with our findings, older age and neurological conditions together had the highest proportion of severe outcomes. Cardiovascular conditions and diabetes, which were among the top list of predictive variables in our models, have also been reported to play an important role in predicting patient outcomes in several independent studies (28).

In this study, older age with the presence of breathing difficulties demonstrated a higher rate of hospitalization when compared to older age without breathing difficulties. Similarly, older age with any of the above-mentioned preexisting conditions showed a clear pattern of increased hospital admission. Cases that are younger or without any preexisting conditions or breathing difficulties had the best prognosis. Also, consistent with the literature, cases in our cohort who had breathing difficulties irrespective of age or presence of preexisting conditions had higher rates of hospitalization (7, 29). Modeling studies based on different cohorts have shown age is a key factor associated with poor outcomes. Our findings are consistent with the literature evidence (8, 25) in that older male have a higher risk of dying compared to their female counterparts.

This study's identification of key factors in predicting COVID-19 disease severity is limited by the available data on variables of interest. In spite of evidence in the literature suggesting the importance of ethnicity and material/social deprivation index of cases in the context of COVID outcomes, we were unable to include these variables due to data quality issues related to missing data. Ethnicity information was inconsistently collected from cases, therefore excluded. In general, deprivation index calculation excludes postal codes of long-term care facilities or other seniors' facilities due to higher mortality rates compared to the rest of the geographies. Due to the unavailability of the deprivation indices for those postal codes, (older adults), the use of the deprivation index in the analysis may have biased the models. Therefore, we excluded the deprivation index in the model prediction analysis.

We also note that subsequent to the period of analysis in our study, the Omicron wave started in late November 2021 in Alberta, Canada. The transmissibility of the Omicron variants was known to be high compared to the variants of previous waves. There are studies suggesting the clinical profile of the Omicron infected patients showed decreased severity of disease compared to the previous waves. However, we were not able to confirm this using our public health data. As such, future modeling approaches should consider the variant type as a potential predictor of outcomes.

Vaccination has reduced severe outcomes of COVID-19 in the population and is also an important factor to consider in predictive models. Vaccine rollout in Alberta began in December 2020 based on strict eligibility criteria. Nearly 50% of cases in our study population had acquired COVID-19 infection in the initial stages of the pandemic before the vaccine rollout in Alberta, so vaccination status was not used as a predictor in our analysis. Interestingly, extraction of vaccination status among the cases in the period of the study showed that only a small proportion of cases had COVID-19 infection after immunization and also had an event of hospitalization, ICU admission or death. Among the cases with severe outcomes, approximately 3% had acquired infection following the first dose and 0.3% cases following the second dose.

Overall, the study model including all age groups as well age-stratified models clearly demonstrate that age and breathing difficulties are key predictive factors leading to severe outcomes among the COVID-19 cases. Underlying conditions including neurological conditions, diabetes, cardiovascular disease, hypertension, and renal disease increase the risk for the severe outcome. Interestingly among the younger adult's cohort, obesity was identified as an important predictor of severe outcomes. Ultimately, we found decision tree modeling approaches have great promise in identifying and informing the clinical management of high-risk COVID-19 patients.

The datasets presented in this article are not readily available because the data set from this study is held securely in coded and de-identified form at Alberta Health Services (AHS). Although data sharing agreements prohibit AHS from making the data set publicly available, access may be granted to those who meet pre specified criteria for confidential access. Please contact research.administration@ahs.ca for more information. Requests to access the datasets should be directed to research.administration@ahs.ca.

The studies involving human participants were reviewed and approved by Conjoint Health Research Ethics Board of Alberta at the University of Calgary. Written informed consent for participation was not required for this study in accordance with the national legislation and the institutional requirements.

KS conceptualized, designed the study, acquired the data, and provided inputs on the analysis. MK conducted the analysis, interpreted the findings, and drafted the manuscript. KW and T-MP provided critical feedback on the analysis and drafting of the manuscript. VS, JM, GO, CN, and HU revised the manuscript critically for important intellectual content. All authors of this study contributed to this manuscript, have approved the manuscript for submission and have granted the corresponding author, and permission to submit the article for publication.

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.