(C-031) Understanding the age-specific risk factors for COVID-19: A machine learning approach

Background: While many risk factors for COVID-19 are well studied, age-specific risk stratification remains underexplored, despite age being a key determinant of outcomes. Appropriate stratification is essential for effective triage and care, particularly in resource-limited settings. Machine learning (ML) applied to large datasets has potential to uncover complex patterns and improve predictive accuracy beyond traditional models.

Objectives: To leverage ML to identify key clinical risk factors for COVID-19 outcomes, with a focus on their age-specific relevance, particularly for comorbidities and drug treatments.

Methods: Using linked nationwide Swedish population data, we assessed associations between pre-pandemic clinical and demographic risk factors and COVID-19 outcomes (test positivity, hospitalization, and death) in 2020, stratified by 10-year age groups. Risk factors included sex, prescription drugs (from the Prescribed Drug Register), and pre-existing conditions (from the National Patient Register). Variables with correlation ≥ 0.8 were excluded. Gradient boosting models were trained on 75% and evaluated on 25% of the data using the area under the receiver operating characteristic curve (AUC-ROC). Model hyperparameters included a shrinkage rate of 0.01 and tree depth to 5 (capturing interactions between up to five variables). Number of trees was optimized via tenfold cross-validation. Risk factor importance was assessed based on its impact on model performance improvement.

Results: We included 11,871,833 individuals, with 3.8 % COVID-19 test-positives, 0.3 % hospitalizations, and 0.09 % deaths during follow-up. Prediction models performed worse for test-positivity (AUC-ROC ~0.5 in younger age groups), reaching 0.72 in the 80–89 age group and 0.80 in those ≥ 90. Hospitalization predictions ranged from 0.53 (0–9 years) to 0.80 (≥ 90 years). Mortality predictions performed overall best, peaking at 0.82 (70–79 years) and 0.81 in ≥ 90 years. Most important risk factors for infection, hospitalization, and death varied across age groups but were generally related to cardiovascular, metabolic, and respiratory conditions. In younger groups, respiratory conditions and anemia were important, while antihypertensives became an increasingly important predictor of infection and hospitalization with age. For COVID-19 mortality, additional risk factors such as pneumonia, seizures, and dementia were prominent, particularly in older populations.

Conclusions: The gradient boosting models showed limited predictive ability (AUC-ROC < 0.7) for test-positivity but moderate to strong (≥ 0.7) for hospitalization and mortality, especially in those aged ≥ 70. Risk factor importance varied by age, highlighting the need for age-specific assessments.