JMIR Publications

ABSTRACT

Background:

Multimorbidity holds paramount importance in public health, representing a multidimensional state where multiple pre-existing medical conditions coexist and interact. This condition has been linked to an elevated risk of COVID-19. Those with multimorbidity who succumb to COVID-19 experience a substantial loss of years. The post-pandemic period also sees an acceleration of frailty. Therefore, it is imperative to incorporate existing multimorbidity details into epidemiological risk assessments. Handling clinical data with medical history poses significant challenges, notably the data's sparsity due to the rarity of multimorbidity conditions and the intricate enumeration of combinatorial multimorbidity features, which introduces a combinatorial explosion issue.

Objective:

This study is distinguishing the severity of COVID-19 on infected people who have multiple medical conditions alongside their demographic characteristics, age, and sex. We propose a sparsity-addressing evolutionary Machine Learning model for analyzing pre-existing multimorbidity in COVID-19 hospitalized patients using their medical history. We attempt to discover the optimal set of multimorbidity feature combinations that are highly associated with COVID-19 severity. We also perform the Apriori algorithm in the evolutionarily obtained such predictive feature combinations to discover combinations with high support.

Methods:

Utilizing data from three administrative sources in Piedmont, Italy, encompassing 12,793 individuals aged 45 to 74 who tested positive for COVID-19 between February and May 2020, we extracted multimorbidity features from 5-year pre-COVID medical histories, including drug prescriptions, disease diagnoses, sex, and age. Our analysis focused on hospitalization due to COVID-19. The data was divided into four cohorts based on age and sex. After addressing data imbalance through random resampling, we compared Machine Learning algorithms to select the optimal classification algorithm for our evolutionary approach. Employing 5-fold cross-validation, we evaluated each model's performance. Subsequently, our evolutionary algorithm utilized a deep learning classifier to generate prediction-based fitness scores for identifying multimorbidity combinations associated with COVID-19 hospitalization. We then employed the Apriori algorithm to identify frequent combinations with high support.

Results:

We identified multimorbidity predictors linked to COVID-19 hospitalization, which serves as an indicator of more severe COVID-19 outcomes. Morbidity features that occurred frequently in final evolved bins or combinations are Age>53, R03BA (glucocorticoid inhalants), N03AX (other antiepileptics) in Cohort 1, A10BA (biguanide or metformin), N02BE (anilides) in Cohort 2, N02AX (other opioids), M04AA (preparations inhibiting uric acid production) in Cohort 3 and G04CA (Alpha-adrenoreceptor antagonists) in Cohort 4.

Conclusions:

When combined with other multimorbidity features, even less prevalent medical conditions demonstrate associations with the outcome. This study provides insights beyond COVID-19, showcasing the adaptability of repurposed administrative data and contributing to enhanced risk assessment for vulnerable populations.