The predictive power of data: machine learning analysis for Covid-19 mortality based on personal, clinical, preclinical, and laboratory variables in a case–control study

Using data from the COVID-19 hospital-based registry database, a retrospective study was conducted from April 2020 to December 2022 at Ayatollah Talleghani Hospital (a COVID‑19 referral center) in Abadan City, Iran.

A total of 14,938 patients were initially screened for eligibility for the study. Of these, 9509 patients were excluded because their transcriptase polymerase chain reaction (RT-PCR) test results were negative or unspecified. The exclusion of patients due to incomplete or missing data is a common issue in medical research, particularly in the use of electronic medical records (EMRs) [27]. In addition, 1623 patients were excluded because their medical records contained more than 70% incomplete or missing data. In addition, patients younger than 18 years were not included in the study. The criterion for excluding 1623 patients due to "70% incomplete or missing data" means that the medical records of these patients did not contain at least 30% of the data required for a meaningful analysis. This threshold was set to ensure that the dataset used for the study contained a sufficient amount of complete and reliable information to draw accurate conclusions. Incomplete or missing data in a medical record may relate to key variables such as patient demographics, symptoms, lab results, treatment information, outcomes, or other data points important to the research. Insufficient data can affect the validity and reliability of study results and lead to potential bias or inaccuracies in the findings. It is important to exclude such incomplete records to maintain the quality and integrity of the research findings and to ensure that the conclusions drawn are based on robust and reliable data. After these exclusions, 3806 patients remained. Of these patients, 474 died due to COVID -19, while the remaining 3332 patients recovered and were included in the control group. To obtain a balanced sample, the control group was selected with a propensity score matching (PSM). The PSM refers to a statistical technique used to create a balanced comparison group by matching individuals in the control group (in this case, the survived group) with individuals in the case group (in this case, the deceased group) based on their propensity scores. In this study, the propensity scores for each person represented the probability of death (coded as a binary outcome; survived = 0, deceased = 1) calculated from a set of covariates (demographic factors) using the matchit function from the MatchIt library. Two individuals, one from the deceased group and one from the survived group, are considered matched if the difference between their propensity scores is small. Non-matching participants are discarded. The matching aims to reduce bias by making the distribution of observed characteristics similar between groups, which ultimately improves the comparability of groups in observational studies [28]. In total, the study included 1063 COVID-19 patients who belonged to either the deceased group (case = 474) or the survived group (control = 589) (Fig. 1).

In the COVID‑19 hospital‑based registry database, one hundred forty primary features in eight main classes including patient’s demographics (eight features), clinical and conditions features (16 features), comorbidities (18 features), treatment (17 features), initial vital sign (14 features), symptoms during hospitalization (31 features), laboratory results (35 features), and an output (0 for survived and 1 for deceased) was recorded for COVID-19 patients. The main features included in the hospital-based COVID-19 registry database are provided in Appendix Table 1.

To ensure the accuracy of the recorded information, discharged patients or their relatives were called and asked to review some of the recorded information (demographic information, symptoms, and medical history). Clinical symptoms and vital signs were referenced to the first day of hospitalization (at admission). Laboratory test results were also referenced to the patient’s first blood sample at the time of hospitalization.

The study analyzed 140 variables in patients' records, normalizing continuous variables and creating a binary feature to categorize patients based on outcomes. To address the issue of an imbalanced dataset, the Synthetic Minority Over-sampling Technique (SMOTE) was utilized. Some classes were combined to simplify variables. For missing data, an imputation technique was applied, assuming a random distribution [29]. Little's MCAR test was performed with the naniar package to assess whether missing data in a dataset is missing completely at random (MCAR) [30]. The null hypothesis in this test is that the data are MCAR, and the test statistic is a chi-square value.

The Ethics Committee of Abadan University of Medical Science approved the research protocol (No. IR.ABADANUMS.REC.1401.095).

All data were collected in eight categories, including demographic, clinical and conditions, comorbidities, treatment, initial vital signs, symptoms, and laboratory tests in medical records, for a total of 140 variables.

The "Demographics" category encompasses eight features, three of which are binary variables and five of which are categorical. The "Clinical Conditions" category includes 16 features, comprising one quantitative variable, 12 binary variables, and five categorical features. "Comorbidities", "Treatment", and "Symptoms" each have 18, 17, and 30 binary features, respectively. Also, there is one quantitative variable in symptoms category. The "Initial Vital Signs" category features 11 quantitative variables, two binary variables, and one categorical variable. Finally, the "Laboratory Tests" category comprises 35 features, with 33 being quantitative, one categorical, and one binary (Appendix Table 1).

In case–control studies, it is common to have unequal size groups since cases are typically fewer than controls [31]. However, in case–control studies with equal sizes, data balancing may not be necessary for ML algorithms [32]. When using ML algorithms, data balancing is generally important when there is an imbalance between classes, i.e., when one class has significantly fewer observations than the other [33]. In such cases, balancing can improve the performance of the algorithm by reducing the bias in favor of the majority class [34]. For case–control studies of the same size, the balance of the classes has already been reached and balancing may not be necessary. However, it is always recommended to evaluate the performance of the ML algorithm with the given data set to determine the need for data balancing. This is because unbalanced case–control ratios can cause inflated type I error rates and deflated type I error rates in balanced studies [35].

Feature selection is about selecting important variables from a large dataset to be used in a ML model to achieve better performance and efficiency. Another goal of feature selection is to reduce computational effort by eliminating irrelevant or redundant features [36, 37]. Before generating predictions, it is important to perform feature selection to improve the accuracy of clinical decisions and reduce errors [37]. To identify the best predictors, researchers often compare the effectiveness of different feature selection methods. In this study, we used five common methods, including Decision Tree (DT), eXtreme Gradient Boosting (XGBoost), Support Vector Machine (SVM), Naïve Bayes (NB), and Random Forest (RF), to select relevant features for predicting mortality of COVID -19 patients. To avoid overfitting, we performed ten-fold cross-validation when training our dataset. This approach may help ensure that our model is optimized for accurate predictions of health status in COVID -19 patients.

In this study, the predictive models were developed with five ML algorithms, including DT, XGBoost, SVM, NB, and RF, using the R programming language (v4.3.1) and its packages [38]. We used cross-validation (CV) to tune the hyperparameters of our models based on the training subset of the dataset. For training and evaluating our ML models, we used a common technique called tenfold cross validation [39]. The primary training dataset was divided into ten folding, each containing 10% of the total data, using a technique called stratified random sampling. For each of the 30% of the data, a ML model was built and trained on the remaining 70% of the data. The performance of the model was then evaluated on the 30%-fold sample. This process was repeated 100 times with different training and test combinations, and the average performance was reported.

Performance measures include sensitivity (recall), specificity, accuracy, F1-score, and the area under the receiver operating characteristics curve (AUC ROC). Sensitivity is defined as TP / (TP + FN), whereas specificity is TN / (TN + FP). F1-score is defined as the harmonic mean of Precision and Recall with equal weight, where Precision equals TP + TN / total. Also, AUC refers to the area under the ROC curve. In the evaluation of ML techniques, values were classified as poor if below 50%, ok if between 50 and 80%, good if between 80 and 90%, and very good if greater than 90%. These criteria are commonly used in reporting model evaluations [40, 41].

Finally, the shapely additive explanation (SHAP) method was used to provide clarity and understanding of the models. SHAP uses cooperative game theory to determine how each feature contributes to the prediction of ML models. This approach allows the computation of the contribution of each feature to model performance [42, 43]. For this purpose, the package shapr was used, which includes a modified iteration of the kernel SHAP approach that takes into account the interdependence of the features when computing the Shapley values [44].

Table 1 shows the baseline characteristics of patients infected with COVID-19, including demographic data such as age and sex and other factors such as occupation, place of residence, marital status, education level, BMI, and season of admission. A total of 1063 adult patients (≥ 18 years) were enrolled in the study, of whom 589 (55.41%) survived and 474 (44.59%) died. Analysis showed that age was significantly different between the two groups, with a mean age of 54.70 ± 15.60 in the survivor group versus 65.53 ± 15.18 in the deceased group (P < 0.001). There was also a significant association between age and survival, with a higher proportion of patients aged < 40 years in the survivor group (77.0%) than in the deceased group (23.0%) (P < 0.001). No significant differences were found between the two groups in terms of sex, occupation, place of residence, marital status, and time of admission. However, there was a significant association between educational level and survival, with a lower proportion of patients with a college degree in the deceased group (37.2%) than in the survivor group (62.8%) (P = 0.017). BMI also differed significantly between the two groups, with the proportion of patients with a BMI > 30 (kg/cm²) being higher in the deceased group (56.5%) than in the survivor group (43.5%) (P < 0.001).

Important insights into the various clinical and condition characteristics associated with COVID-19 infection outcomes provides in Table 2. The results show that patients who survived the infection had a significantly shorter hospitalization time (2.20 ± 1.63 days) compared to those who died (4.05 ± 3.10 days) (P < 0.001). Patients who were admitted as elective cases had a higher survival rate (84.6%) compared to those who were admitted as urgent (61.3%) or emergency (47.4%) cases. There were no significant differences with regard to the number of infections or family infection history. However, patients who had a history of travel had a lower decease rate (40.1%).

A significantly higher proportion of deceased patients had cases requiring CPR (54.7% vs. 45.3%). Patients who had underlying medical conditions had a significantly lower survival rate (38.3%), with hyperlipidemia being the most prevalent condition (18.7%). Patients who had a history of alcohol consumption (12.5%), transplantation (30.0%), chemotropic (21.4%) or special drug use (0.0%), and immunosuppressive drug use (30.0%) also had a lower survival rate. Pregnant patients (44.4%) had similar survival outcomes compared to non-pregnant patients (55.6%). Patients who were recent or current smokers (36.4%) also had a significantly lower survival rate.

Table 3 summarizes the comorbidity characteristics of COVID-19 infected patients. Out of 1063 patients, 54.84% had comorbidities. Chi-Square tests for individual comorbidities showed that most of them had a significant association with COVID-19 outcomes, with P-values less than 0.05. Among the various comorbidities, hypertension (HTN) and diabetes mellitus (DM) were the most prevalent, with 12% and 11.5% of patients having these conditions, respectively. The highest fatality rates were observed among patients with cardiovascular disease (95.5%), chronic kidney disease (62.5%), gastrointestinal (GI) (93.3%), and liver diseases (73.3%). Conversely, patients with neurology comorbidities had the lowest fatality rate (0%). These results highlight the significant role of comorbidities in COVID-19 outcomes and emphasize the need for special attention to be paid to patients with pre-existing health conditions.

The treatment characteristics of the COVID-19 patients and the resulting outcomes are shown in Table 4. The table shows the frequency of patients who received different types of medications or therapies during their treatment. According to the results, the use of antibiotics (35.1%), remdesivir (29.6%), favipiravir (36.0%), and Vitamin zinc (33.5%) was significantly associated with a lower mortality rate (P < 0.001), suggesting that these medications may have a positive impact on patient outcomes. On the other hand, the use of Heparin (66.1%), Insulin (82.6%), Antifungal (89.6%), ACE inhibitors (78.1%), and Angiotensin II Receptor Blockers (ARB) (83.8%) was significantly associated with increased mortality (P < 0.001), suggesting that these medications may have a negative effect on the patient's outcome. Also, It seems that taking hydroxychloroquine (51.0%) is associated with a worse outcome at lower significance (P = 0.022). The use of Atrovent, Corticosteroids and Non-Steroidal Anti-Inflammatory Drugs (NSAIDs) did not show a significant association with survival or mortality rates. Similarly, the use of Intravenous Immunoglobulin (IVIg), Vitamin C, Vitamin D, and Diuretic did not show a significant association with the patient’s outcome.

Table 5 provides initial vital sign characteristics of COVID-19 patients, including heart rate, respiratory rate, temperature, blood pressure, oxygen therapy, and radiography test result. The findings shows that deceased patients had higher HR (83.03 bpm vs. 76.14 bpm, P < 0.001), lower RR (11.40 bpm vs. 16.25 bpm, P < 0.001), higher temperature (37.43 °C vs. 36.91 °C, P < 0.001), higher SBP (128.16 mmHg vs. 123.33 mmHg, P < 0.001), and higher O₂ requirements (invasive: 75.0% vs. 25.0%, P < 0.001) compared to the survived patients. Additionally, deceased patients had higher MAP (99.35 mmHg vs. 96.08 mmHg, P = 0.005), and lower SPO₂ percentage (81.29% vs. 91.95%, P < 0.001) compared to the survived patients. Furthermore, deceased patients had higher PEEP levels (5.83 cmH2O vs. 0.69 cmH2O, P < 0.001), higher FiO2 levels (51.43% vs. 8.97%, P < 0.001), and more frequent bilateral pneumonia (63.0% vs. 37.0%, P < 0.001) compared to the survived patients. There appears to be no relationship between diastolic blood pressure and treatment outcome (83.44 mmHg vs. 85.61 mmHg).

Table 6 provides information on the symptoms of patients infected with COVID-19 by survival outcome. The table also shows the frequency of symptoms among patients. The most common symptom reported by patients was fever, which occurred in 67.0% of surviving and deceased patients. Dyspnea and nonproductive cough were the second and third most common symptoms, reported by 40.4% and 29.3% of the total sample, respectively. Other common symptoms listed in the Table were malodor (28.7%), dyspepsia (28.4%), and myalgia (25.6%).

Table 7 shows the laboratory values of COVID-19 patients with the average values of the different laboratory results. The results show that the deceased patients had significantly lower levels of red blood cells (3.78 × 106/µL vs. 5.01 × 106/µL), hemoglobin (11.22 g/dL vs. 14.10 g/dL), and hematocrit (34.10% vs. 42.46%), whereas basophils and white blood cells did not differ significantly between the two groups. The percentage of neutrophils (65.59% vs. 62.58%) and monocytes (4.34% vs. 3.93%) was significantly higher in deceased patients, while the percentage of lymphocytes and eosinophils did not differ significantly between the two groups. In addition, deceased patients had higher levels of certain biomarkers, including D-dimer (1.347 mgFEU/L vs. 0.155 mgFEU/L), lactate dehydrogenase (174.61 U/L vs. 128.48 U/L), aspartate aminotransferase (93.09 U/L vs. 39.63 U/L), alanine aminotransferase (74.48 U/L vs. 28.70 U/L), alkaline phosphatase (119.51 IU/L vs. 81.34 IU/L), creatine phosphokinase-MB (4.65 IU/L vs. 3.33 IU/L), and positive troponin I (56.5% vs. 43.5%). The proportion of patients with positive C-reactive protein was also higher in the deceased group.

Five ML algorithms, namely DT, XGBoost, SVM, NB, and RF, were used in this study to build mortality prediction models COVID -19. The models were based on the optimal feature set selected in a previous step and were trained on the same data set. The effectiveness of the models was evaluated by calculating sensitivity, specificity, accuracy, F1 score, and AUC metrics. Table 8 shows the results of this performance evaluation. The average values are expressed from the test set as the mean (standard deviation).

The results show that the performance of the models varies widely in the different feature categories. The Laboratory Tests category achieved the highest performance, with all models scoring 100% in all metrics. The Symptoms and initial Vital Signs categories also show high performance, with XGBoost achieving the highest accuracy of 98.03% and DT achieving the highest sensitivity of 92.79%.

The Clinical and Conditions category also showed high performance, with all models showing accuracy above 91%. XGBoost achieved the highest sensitivity and specificity of 92.74% and 92.96%, respectively. In contrast, the Demographics category showed the lowest performance, with all models achieving less than 66.5% accuracy.

In summary, the results suggest that certain feature categories may be more useful than others in predicting mortality from COVID-19 and that some ML models may perform better than others depending on the feature category used.

SHapley Additive exPlanations (SHAP) values indicate the importance or contribution of each feature in predicting model output. These values help to understand the influence and importance of each feature on the model's decision-making process.

In Fig. 2, the mean absolute SHAP values are shown to depict global feature importance. Figure 2 shows the contribution of each feature within its respective group as calculated by the XGBoost prediction model using SHAP. According to the SHAP method, the features that had the greatest impact on predicting COVID-19 mortality were, in descending order: D-dimer, CPR, PEEP, underlying disease, ESR, antifungal treatment, PaO2, age, dyspnea, and nausea.

On the other hand, Fig. 3 presents the local explanation summary that indicates the direction of the relationship between a variable and COVID-19 outcome. As shown in Fig. 3(I to VII), older age and very low BMI were the two demographic factors with the greatest impact on model outcome, followed by clinical factors such as higher CPR, hospitalization, and hyperlipidemia. Higher mortality rates were associated with patients who smoked and had traveled in the past 14 days. Patients with underlying diseases, especially HTN, died more frequently. In contrast, the use of remdesivir, Vit Zn, and favipiravir is associated with lower mortality. Initial vital signs such as high PEEP, low PaO2 and RR had the greatest impact, as did symptoms such as dyspnea, MODS, sore throat and LOC. A higher risk of mortality is observed in patients with higher D-dimer levels and ESR as the most consequential laboratory tests, followed by K, AST and CPK-MB.

Using the feature types listed in Appendix Table 1, Fig. 4 shows that the performance of ML algorithms can be improved by increasing the number of features used in training, especially in distinguishing between symptoms, comorbidities, and treatments. In addition, the amount and quality of data used for training can significantly affect algorithm performance, with laboratory tests being more informative than initial vital signs. Regarding the influence of features, quantitative features tend to have a more positive effect on performance than qualitative features; clinical conditions tend to be more informative than demographic data. Thus, both the amount of data and the type of features used have a significant impact on the performance of ML algorithms.

One of the limitations of this study is that it relies on data collected from a single hospital in Abadan, Iran. The data may not be representative of the diversity of COVID -19 cases in different regions, and there may be differences in data quality and completeness. In addition, retrospectively collected data may have biases and inaccuracies. Although the study included a substantial number of COVID -19 patients, the sample size may still limit the generalizability of the results, especially for less common subgroups or certain demographic characteristics.

Future studies could adopt a multi-center approach to improve the scope and depth of research on COVID-19 outcomes. This could include working with multiple hospitals in different regions of Iran to ensure a more diverse and representative sample. By conducting prospective studies, researchers can collect data in real time, which reduces the biases associated with retrospective data collection and increases the reliability of the results. Increasing sample size, conducting longitudinal studies to track patient progression, and implementing quality assurance measures are critical to improving generalizability, understanding long-term effects, and ensuring data accuracy in future research efforts. Collectively, these strategies aim to address the limitations of individual studies and make an important contribution to a more comprehensive understanding of COVID-19 outcomes in different populations and settings.