SARS-CoV-2 Infection Impacts Liver-Related Blood Tests Following Infection: A Health Administrative Data Study

Study population

This study was approved by the Queen’s University Health Sciences Research Ethics Board (PAED-542-22) and the Ontario Ministry of Health for access on the Ontario Health Data Platform (OHDP). Consent to participate was not applicable, as administrative databases were used for data collection. The data for this study derives from OHDP, a high-performance computing environment that linked large health datasets from Ontario, Canada, serving a population of approximately 14.8 million (Ontario Data Catalogue, 2021). The data analysis plan was not preregistered.

Datasets

The dataset for the present study was created by extracting variables from the following databases within the OHDP: Ontario Health Insurance Program (OHIP) for diagnostic codes, the Canadian Institute for Health Information (CIHI) discharge abstract data (DAD) for hospitalized patient information, the Ontario Laboratory Information System (OLIS) for liver related laboratory values, and the Ontario Laboratories Information System – COVID-19 subset (OLIS-C19) for any positive test for any SARS-CoV-2 variant. Further individual demographic information (year of birth, sex, geographic location information) was obtained through the Ministry of Health Registered Person DataBase (RPDB). All direct individual identifiers were removed from the data prior to analysis. Supplemental materials are available via Zenodo at doi.org/10.5281/zenodo.18236199.¹¹

Data extraction

The study period was from January 4, 2020 to September 2, 2022 (for COVID-19 tests) with follow up to 28 days after the last index date for labs. Inclusion were: any individuals over one year old (with no upper age limit) at the time of their COVID-19 test in the Ontario health administrative dataset (all were polymerase chain reaction tests for any variant), who have at least one positive COVID-19 test within the study period and at least one subsequent liver-related blood test in the 28 days following the positive test. Liver-related blood tests used in this study were based on clinical guideline recommendations for investigating liver disease: ALT, AST, international normalized ratio (INR), conjugated bilirubin, total bilirubin, and albumin.¹²

Data was extracted for cases at the time of their first positive COVID-19 test, regardless of if they had or had not had previous negative tests. If there were more than one of any given liver test in the 28 day period, then the most severely abnormal value was taken in the direction of the test that would indicate liver injury (that is, increased ALT/AST/INR/bilirubin levels, and decreased albumin levels). The values used to define normal ranges across sex and age groups was used for each test is found in Supplemental Table S1. Individuals excluded were those who: had any pre-existing billing diagnosis of hepatitis or cirrhosis (OHIP diagnostic codes 70 and 571, respectively) in DAD and OHIP (prior to their first COVID-19 test), incomplete demographic data, and individuals less than 1 year old.

Matched controls were individuals with their first negative COVID-19 test within the study period and at least one subsequent liver-related blood test (as defined above) in the 28 days following the negative test and no further COVID-19 test within the subsequent 28 days. These individuals otherwise met the inclusion criteria. COVID-19 positive cases were then each matched by age (in years) and sex with two control individuals, with the same exclusion criteria. Controls were always unique individuals from cases, by their unique identifier. Thus, those who were controls that at a later date were otherwise eligible to become cases (by later having a positive COVID-19 test and otherwise met the inclusion criteria) were excluded. The data extraction process is summarized in Figure 1.

Figure 1. Case and control identification strategy flowchart.

Missing demographics that led to exclusion included no or invalid information for: postal code, sex, age, or liver test results.

Software and statistical analysis

All analysis was performed within the OHDP environment, using R (version 4.1.3) and python (v 3.8.5). Comparisons of COVID-19 positive versus negative individuals used the Person’s chi-squared test with Yates’ continuity correction. Due to the non-parametric distribution of all the liver blood tests (most were heavily right skewed), we used the Mann-Whitney U test to compare each blood test for between cases and controls. Associations were performed using logistic regression function in the base R stats package. Statistical significance was considered a p-value less than 0.05.

For machine learning, the sklearn, LogisticRegression, RandomForestClassifier, XGBoost, and TensorFlow python packages were used. The dataset included all liver blood tests (except conjugated bilirubin, since it was strongly correlated with total bilirubin), age (in years), and sex. For the complete data-only models Rurality Index for Ontario measures were also included for each individual based on their associated census subdivision (CSD) number recorded in RPDB.¹³ The data was divided 80:20 into training and testing sets. Hyperparameters were tuned using a focused iterative search strategy. We implemented a logistic regression model configured with an L1 penalty, a regularization strength of 0.1, and a liblinear solver. We implemented a random forest classifier configured with 100 estimators, as well as an XGBoost classifier that was configured with a learning rate of 0.1, a max depth of 5, and a logloss evaluation metric. We also developed a fully connected neural network consisting of five hidden layers, each with thirty neurons, that uses the ReLU activation function connected to a single neuron output layer with sigmoid activation for binary classification. The networks was configured with an Adam optimizer and a binary cross-entropy loss function. The network was trained for 50 epochs with a batch size of 100. The testing set was only evaluated once, after all hyperparameters were set and the model fully developed on the training set.

Individual characteristics

101,574 cases were age and sex matched to 200,709 matched controls met the inclusion/exclusion criteria ( Figure 1). Of the cases, 49% (49,299) had at least one abnormal liver blood test in the 28 days following their positive COVID-19 test. Of the controls, 41% (82,495) had at least one abnormal liver blood test in the 28 days following their negative COVID-19 test. Individual demographics and blood test statistics are summarized in Table 1.

Liver blood tests differ in individuals with COVID-19

Multiple differences were found between individual liver test values when comparing individuals with a positive COVID-19 test (cases) to those with a negative COVID-19 test (controls). The rate of abnormal tests, when accounting for differences in normal ranges by age, was statistically significant for all six liver blood tests between cases and controls ( Figure 2A, with summary statistics in Supplemental Table S2).

Figure 2. The patterns of liver test results differ between patients with recent negative versus positive COVID-19 test results.

(A) Rate of abnormal test results differ by COVID-19 status, with overall increases with COVID-19 for ALT, AST, and albumin, and decreases with COVID-19 for INR, conjugated bilirubin, and total bilirubin. (B-G) Scatterplots demonstrate the substantial overlap between the absolute values of each liver test at a population level. Boxes represent the interquartile range (IQR) with the median value inside the box and the vertical lines represent the smallest and largest data points within 1.5 times the respective IQR. All absolute value tests, except INR, were statistically significant between cases and controls.

The degree of alteration to liver blood tests is also important, not simply the abnormal/normal dichotomy. Thus, we also compared the absolute values of the individual tests between cases and controls ( Figure 2B-G). In this comparison, five out of the six test absolute values were significantly different, with INR the exception (summary statistics in Supplemental Table S2).

Consistent timeline of liver blood test alterations

Since the study duration extended over 28 days following the COVID-19 test, we examined the frequency at which each test was performed in each group over this time. It would be important to identify if certain tests were more commonly done later than others as these would be theoretically more likely to be normal and thus could bias the machine learning models. However, the trends of dates of each test were similar with most of the testing being done within 3 days of the individual’s COVID-19 test and then rapidly trending down to a low rate of further testing after that for all tests (Supplemental Figure S1). Notably, there are peaks for the frequency of each test at 7-day intervals, likely representing practitioners monitoring these tests weekly.

Lack of correlation between individual liver tests

To identify variables with high likelihood to change together, we performed a correlation analysis of each permutation of pairs of liver tests (Supplemental Figure S2). Of the 15 pair-wise correlations, only ALT/AST and total bilirubin/conjugated bilirubin were statistically correlated (r-squared values of 0.56 and 0.88, respectively).

Machine learning models identify COVID-19 status using liver blood alterations

We initially compared four machine learning models with all matched individuals that met the inclusion criteria for the study. We chose to exclude conjugated bilirubin since it was highly correlated with total bilirubin, and total bilirubin was more commonly tested. We chose not to dimension reduce these two parameters into one since the explainability of the model could be important and so we kept the clinically relevant value. ALT and AST were not dimension reduced despite a modest correlation, since these parameters are both clinically important and their relative contributions to the models would be important information for determining the clinical value of these tests in suspected COVID-19 cases.

We compared logistic regression, random forest, extreme gradient boost (XGBoost), and multilayer perceptron neutral network models. The outcome was COVID-19 status (positive/negative). After hyperparameter tuning on the training set, each model was used to evaluate the test set, and the optimal model was XGBoost with an area under the curve (AUC) in the receiver operator characteristic curve of 0.71 ( Figure 3A). Notably, age was the feature with the highest relative contribution to the model (F score = 469), with four of the five blood tests contributing relatively highly to the model (F scores range of 366 to 425). INR contributed modestly with an F score of 36 and sex contributed minimally with an F score of 55 ( Figure 3B).

Figure 3. Performance comparison of machine learning models in the test set with all five liver tests for classifying patients by COVID-19 status.

(A) The optimal model using XGBoost had an area under the receiver operator curve of 0.71. (B) In the optimal model, age was the feature with the highest relative importance to this model, followed by the five liver tests. (C) The XGBoost model for the adult patient subcohort test set was the optimal model, having an area under the receiver operator curve of 0.74. (D) Confusion matrixes of each machine learning model for the adult patient subcohort test set showing a balance of sensitivity and specificity for the XGBoost model, compared to the other models.

Machine learning with complete data for adults only improves identification of COVID-19 status

We suspected that the sparse nature of the data (as seen in Table 1) was limiting the models’ ability to identify COVID-19 cases, so we then limited the cohort to only individuals with all five of the blood tests available (excluding conjugated bilirubin). We also added the rurality index Ontario (RIO) as a variable. Notably, with this data subset, there were no missing values and included 24,676 individuals ( Table 2).

Since the impact of liver disease can differ by age group, and the general models had age having a relatively strong impact on the models, we also performed subgroup analysis on the larger cohort limiting to only individuals less than 18 years old, only those 19-64 years old, and only those greater than 64 years old (and permutations of pairs of these three age groups). The AUC improved most with those less than 18 years old excluded ( Table 3). Again, the optimal performing model was XGBoost, with an increase in the AUC to 0.74 on the test set ( Figure 3C and D).