Data on COVID-19 mortalities were retrieved from Jaber Hospital's mortality Electronic Health Records (EHR) for all patients admitted to the ICU between March 7, 2020, and August 19, 2021, and death reported between March 7, 2020, and August 27, 2021. The data set contains 1691 cases after excluding 12 children (<17 years old) and 46 with no data entries. The monthly total death rate in Kuwait is depicted in Worldometer cite (10). On the final day of data collection for this study, the total number of COVID-19 deaths was reported to be 2415; thus, our sample size covers 70% (1691/2415) of the COVID-19 mortality population. We also covered all death peaks and pandemic main waves during this time.

Initially, the data was extracted as a pdf file and then converted to an Excel spreadsheet. Patients' demographics (age, gender, and residency), date of ICU admission, date of death, reasons for admission, admission diagnosis, final diagnosis, cause of death, brief history, brief summary, and contributing factors are all included in each record. To ensure confidentiality, all data was anonymized and all patient identifiers were removed. Additional data cleansing were also performed to ensure data accuracy.

The data sheet obtained from EHR has mainly eleven columns: date of admission to ICU, date of death, age at death, admission diagnosis, reason for admission, final diagnosis, cause of death (COD), brief history, brief summary, and contributing factors. Except for the first three, all remaining columns are text features.

The underlying cause of death, such as “COVID-19” or “COVID 19 pneumonia,” was listed in the COD column in many records, whereas the primary/intermediate causes were found explicitly or indirectly in the brief summary or brief history columns. Furthermore, there were two major flaws in the free-text notes in the mortality EHR. The first issue is that many terminologies have misspellings or improper forms. For example, multiorgan failure is referred to as “Multi-Organs” or “Multiorgan Failure.” The second issue is inconsistency in the reporting of text notes. The causes of death are not always listed in the data columns that you would expect. Comorbidities, on the other hand, are not consistently included in the list of contributing factors. As a result, we are unable to use the existing NLP tools or Unified Medical Language System (UMLS). We had to develop our own system to extract concepts, knowledge, and relationships from the mortality EHR at hand.

We created an NLP method to identify, extract, and automatically encode natural language from mortality EHRs into structured clinical data. Tables 1, 2 are used as keywords to extract primary and intermediate CODs, while Table 3 presents keywords to extract comorbidities. Method created in Python. Figure 2 shows our algorithm.

In this method, text is stripped of punctuation, special characters, capitalization, stop words, and tokenization. Used EHR variables include cause of death, final diagnosis, brief history, and brief summary. To create a case/COD term occurrence matrix, binary variables must be created for each primary/intermediate COD listed in Tables 1, 2. Initial occurrence matrix setting is zero. CODs or equivalents are compatible with tokens. The case/term occurrence matrix cell is set to 1 upon a match. Every case applies (rows). A COD abbreviation was not mistaken for a term, as PE is not present in hypertensive or hyperthyroid. Negation was also carefully handled; if a term is preceded by a negative or conditional word, it will not match. Exclusion words consist of (no, not, no sign of, non, no history of, no active, no previous medical, not known to have, no indications of, previous condition, old condition). Text format is used to list the final primary and intermediate CODs. The pseudocode used to extract the final primary COD is depicted in Figure 3.

Determining the actual intermediate CODs are handled differently. Multiple intermediate CODs are reported as a group. Our clinicians manually validated and separated the correct outcome to determine which disorders were terminal. A counter matching the extracted causes is also computed to help identify the terminal cause based on the most common causes to cross-check the accuracy of the findings.

The comorbidities for each case are identified using Table 3 in the same manner that CODs are identified. Preprocessed word tokens are extracted from the EHR reason for admission, contributing factors, admission diagnosis and brief summary.

Original EHR mortality data had two sets of variables. First set included seven categorical and quantitative variables. Second set included eight free-text variables. The pdf data sheet was converted to an Excel sheet for data manipulation and cleaning. The second set of data was used to generate 70 variables using Python to determine death causes and comorbidities. During exploratory data analysis, we generated appropriate graphs (bar, pie, boxplots) and summary statistics (mean, median, SD, IQR). Hypothesis tests included Chi-square, TURF, ANOVA, and Kruskal Wallis. Finally, we built our prediction model with a decision tree. SPSS V23 and R were used for the statistical analysis.

The majority of the 1,691 anonymous COVID-19 decedents were male 963 (56.9%). The age at death ranges between 19.8 and 103.2 years with 63.8 years (SD 14.4). On the average the duration stay in ICU prior to death was 18.5 days (SD 12.8). Two or more comorbidities were present (mean 2.5, SD 1.9) with hypertension and diabetes mellitus shared among more than half of them (Table 4). Since these patients died in the intensive care unit, COVID-19 pneumonia was mainly the underlying cause of death that resulted in intermediate and thus primary causes of death. COVID-19 pneumonia was detected in 94 percent of cases (1592/1691).

When the mean ICU stay was compared across the three age groups of <50, 50–64, and 65 or more, no significant difference (Table 5) using the ANOVA F-test (p-value = 0.903). On the other hand, testing for mean total comorbidities across these three age groups was significant (p-value <0.0001), and the Tukey B multiple comparison test reveals significance with three means for groups in homogenous subsets of mean total comorbidities of 1.25, 2.19, and 3.28, respectively.

We identified primary and secondary causes of death. Septic shock was the primary COD in 667 patients (39.4%), followed by cardiopulmonary arrest 304 (18.0%), respiratory failure 235 (13.9%), and cardiac arrest 180 (10.6%). The percentages of cases with (septic shock & MOF), MOF, and renal failure were 135 (8.0%), 125 (7.4), and 44 (2.6%), respectively. Hepatic failure occurred in only one case and thus ignored from further analysis. On the other hand, ARDS was one of the main reasons for ICU admissions and was reported in all deaths. Numerous cases were reported in which a combination of intermediate death complications occurred. These cases were thoroughly examined by our physicians to determine which terminal complication is more likely to be classified as the intermediate COD. It was found that around 75% of these decedents had ARDS as an intermediate COD, while the remaining 25% had intermediate COD other than ARDS. Among the other causes are AKI, AR, BL, DI, HF, MI, PE, PN, ST, and UT. The frequency distribution of intermediate combined complications along with the frequency distribution of the terminal complication leading to intermediate COD are shown in Figure 4. Table 6 shows the count and percentage of counts for primary and intermediate causes, as well as the column percentages for primary causes. While ARDS is the most prevalent intermediate COD regardless of primary cause, AR and MI disorders were significantly (7.2–14.4%) linked with cardiac arrest and MOF.

Age distribution appears to be similar by primary COD, with a median age at death of 64.5 years and an interquartile range (IQR = 20.7). However, a few young patients, approximately the age of 20, died because of MOF or renal failure (Figure 5). The median length of stay in the ICU prior to death was approximately 16 days overall but was significantly longer (~ 20 days) for those who died of septic shock or (septic shock + MOF). Patients who died because of MOF had an average of three or more comorbidities. Those who died of renal failure and (septic shock + MOF) died in a manner like that described above.

Those who died because of AR, DI, or MI had the highest average age (70 years) and total comorbidities (3 or more), as well as the shortest average stay in the ICU. Patients who died of HF were younger (average age 50 years) and had more than two comorbidities, with an ICU stay of < 20 days. We also noticed the sequences of (MI → septic shock) and (PE → respiratory failure) were associate with 4 or above comorbidities on the average.

The following is a list of the comorbidities of dead patients in this study. Hypertension (57%) is the most common condition, followed by diabetes (52%), coronary artery disease (23%), and chronic renal disease (14%). 12% for each arrhythmia and dyslipidemia, cancer (11%), Rheumatoid arthritis (10%), obesity (9%), thyroid disease (8%), stroke (7%), pulmonary embolism (5%), asthma (4%), valvular heart disease (4%), bleeding disorders (4%), and 3% for each COPD and dementia. The remaining comorbidities with < 3% reported incidence include Anemia, heart failure, prostate hyperplasia, liver disease, epilepsy, cardiomyopathy, peripheral vascular disease, lung disease, psychiatric disorders, osteoporosis, multiple sclerosis, down syndrome, Parkinson's disease, inflammatory bowel disease, gastroesophageal reflux disease, Addison disease, and HIV infection.

Next, we present the results of the Total Unduplicated Reach and Frequency (TURF) method. TURF is a popular statistical technique in market research that ranks product combinations according to the number of customers who favor them (12). In this study, we applied the method in a clinical setting, treating comorbidities and patients as products and people. The goal is to determine the most likely disease combinations that these patients share. The analysis traverses all possible combinations of comorbidities and records two statistics for each: reach and frequency. The reach is the percentage of individuals who exhibit at least one comorbidity in a given combination, and the frequency is the total number of times comorbidities are exhibited in a given combination. We tested the method for all comorbidities listed in Table 3 and a range of reach values. Table 7 provides a summary of the ideal choices according to the number of diseases (Size). For instance, the optimal combination of four comorbidities has a 73 percent success rate with RA, cancer, DM, and HTN. This indicates that seventy-three percent of the patients had at least one of the conditions (rheumatological disorders, cancer, DM, HTN). If Diabetes and High Blood Pressure were eliminated from the analysis due to their high prevalence and we wanted to evaluate other possible combinations of diseases, the one with the highest prevalence was (obesity, CAD, Cancer, RA) with 43.6%.

When we looked at the general disease classification frequencies, we found that over 60% of the patients had circulatory (DCS) and endocrine (ENMS) disorders, one-third had cardiovascular diseases (CVD), and the remaining categories (RDs, CRC, DNS, ADs, DGS) varied from 8 to 15%. In compared to patients who died of ARDS/PE/Other, approximately 65 percent of patients who died of MI or AR had cardiovascular illnesses (Figure 6). Those who die from ARDS, on the other hand, usually have endocrine or circulatory system problems. Nervous system diseases were the least common among the PE dead. With chi-square test findings of (175.5, p-value 0.001) and (12.2, p-value = 0.016), the circulatory and nervous systems had the most significant association with intermediate COD.

The total comorbidities distribution by age group of COVID-19 deaths due to ARDS or other cause is displayed in Figure 7. Patients under the age of 50 have a similar comorbidity distribution, with an average of one disease. Two comorbidities were found on average per age group (50–64) with more variation among those who died from causes other than ARDS. In contrast, older patients (age >65) who died from causes other than ARDS have an average of four comorbidities, compared to three for the other group who died mainly from ARDS.

In this section, we used decision tree (DT) to determine the most parsimonious predictors of intermediate COD among COVID-19 patients in intensive care units. Decision trees learn to divide data into smaller and smaller categories to forecast the goal. The test is represented by a node, while the numerous outcomes are represented by edges. The dividing process is repeated until no further gains can be obtained or a preset rule is reached. Three common decision tree techniques include classification and regression tree (CART), chi-squared automatic interaction detection (CHAID), and quick unbiased efficient statistical tree (QUIEST). For mathematical explanations and performance comparisons of these DT approaches, see Lin et al. (13). Figure 8 illustrates the results of the QUEST model, which demonstrate that the existence of an arrhythmia (AF) was the best indicator of the intermediate cause (ARDS/Other). Patients with AF are more likely to have a cause other than ARDS (54.9%). Node 1 is considered a terminal node for predicting a cause of death other than ARDS since no child nodes was found below it. In patients without AF, on the other hand, CAD was the second-best predictor of (ARDS/Other). In patients without AF but with CAD, the terminal Node 3 predicted 66.9 ARDS vs. 33.1% for other causes. PE is an additional predictor in the model for patients who do not have AF or CAD. ARDS is the main intermediate COD in this group, accounting for over 83% of patients without PE and 58% of patients with PE who died from ARDS. The risk and classification tables allow for a quick evaluation of the model's performance. The risk of misclassifying the cause of death is estimated to be 0.272 (or 27.2%), which is consistent with the results of the classification table, which show that 76% of causes of death are correctly classified.

The dynamic nature of the method, its usability, and its potential to maintain self-control all contribute to its strength. In addition, the sampled data span both significant pandemic waves and death peaks, accounting for 70% of the total reported COVID-19 fatality cases in Kuwait. The death rate drastically decreased after then. Therefore, our sample represents the population under consideration to a high degree of accuracy. Nevertheless, our study has several limitations. First, there is a chance of selection or referral bias as the research was conducted at a single location, i.e., Jaber Hospital. Second, the lack of information extracted from the inadequate documentation of the patient records. The absence of a symptom (such as obesity, smoking, etc.) does not necessarily suggest that a patient is symptom-free. Thirdly, patients were typically transferred late in the course of their disease, and their medical records lacked vital medical history information. Such discrepancies in clinical data may result in information bias that contributes to a decrease in model precision.

Future studies could potentially investigate the impact of vaccines on the time to death, provide survival time estimates by cause of death, and perform spatiotemporal analyses of transferable patients. Knowing the COVID-19 death rate and patient survival rate can help risk management experts. COVID-19 or its evolving variants can be avoided, and strategies can be used to slow their spread.