Demystifying COVID-19 mortality causes with interpretable data mining

Abstract

While COVID-19 becomes periodical, old individuals remain vulnerable to severe disease with high mortality. Although there have been some studies on revealing different risk factors affecting the death of COVID-19 patients, researchers rarely provide a comprehensive analysis to reveal the relationships and interactive effects of the risk factors of COVID-19 mortality, especially in the elderly. Through retrospectively including 1917 COVID-19 patients (102 were dead) admitted to Xiangya Hospital from December 2022 to March 2023, we used the association rule mining method to identify the risk factors leading causes of death among the elderly. Firstly, we used the Affinity Propagation clustering to extract key features from the dataset. Then, we applied the Apriori Algorithm to obtain 6 groups of abnormal feature combinations with significant increments in mortality rate. The results showed a relationship between the number of abnormal feature combinations and mortality rates within different groups. Patients with "C-reactive protein > 8 mg/L", "neutrophils percentage > 75.0 %", "lymphocytes percentage < 20%", and "albumin < 40 g/L" have a 2 $\times$ mortality rate than the basic one. When the characteristics of "D-dimer > 0.5 mg/L" and "WBC > $9.5\times {10}^9$ /L" are continuously included in this foundation, the mortality rate can be increased to 3 $\times$ or 4 $\times$ . In addition, we also found that liver and kidney diseases significantly affect patient mortality, and the mortality rate can be as high as 100%. These findings can support auxiliary diagnosis and treatment to facilitate early intervention in patients, thereby reducing patient mortality.

Figure 1

Frame diagram. The diagram illustrates a comprehensive architecture designed to analyze the combined mortality indicators of COVID-19 patients. This overall mining process can be broken down into four primary stages: data collection, data preprocessing, feature selection, and association rule mining. In the initial stage, a detailed review of the examination data of COVID-19 patients is conducted, including screening of basic disease information and demographic details, and data integrity checks. Following data preprocessing, the processed dataset was utilized to identify key features affecting COVID-19 patient mortality. This selection process leveraged both machine learning clustering techniques and established medical knowledge. Lastly, employing the association rule mining method, we derived specific mortality patterns for COVID-19 patients. These patterns not only align with existing research but also reveal fresh insights, thereby supporting clinical in diagnosis and treatment protocols.

Figure 2

The mortality rate and the proportion of deaths in six groups.

Figure 3

Survival curves of six groups.