Informative Missingness: What can we learn from patterns in missing laboratory data in the electronic health record?
1Harvard Medical School, Cambridge, MA, USA
2University of Pennsylvania Perelman School of Medicine, Philadelphia, PA, USA
3Northwestern University, Chicago, IL, USA
4Massachusetts General Hospital, Boston, MA, USA
5University of Pavia, Pavia, Italy
6University of Pittsburgh, Pittsburgh, PA, USA
7University of Michigan, Ann Arbor, MI, USA
8Bordeaux University Hospital, Talence, France
9ASST Papa Giovanni XXIII, Bergamo, Italy
10Cincinnati Children’s Hospital Medical Center, University of Cincinnati, Cincinnati, OH, USA
11Duke University, Durham, NC, USA
12National University Health Systems, Singapore
13University Of Kansas Medical Center
14University of Pennsylvania Health System, Philadelphia, PA, USA
15Boston Children’s Hospital, Boston, MA, USA
# Corresponding author; email: jhholmes@pennmedicine.upenn.eduAbstract
Background
In electronic health records, patterns of missing laboratory test results could capture patients’ course of disease as well as reflect clinician’s concerns or worries for possible conditions. These patterns are often understudied and overlooked. This study aims to characterize the patterns of missingness among laboratory data collected across 15 healthcare system sites in three countries for COVID-19 inpatients.
Methods
We collected and analyzed demographic, diagnosis, and laboratory data for 69,939 patients with positive COVID-19 PCR tests across three countries from 1 January 2020 through 30 September 2021. We analyzed missing laboratory measurements across sites, missingness stratification by demographic variables, temporal trends of missingness, correlations between labs based on missingness indicators over time, and clustering of groups of labs based on their missingness/ordering pattern.
Results
With these analyses, we identified mapping issues faced in seven out of 15 sites. We also identified nuances in data collection and variable definition for the various sites. Temporal trend analyses may support the use of laboratory test result missingness patterns in identifying severe COVID-19 patients. Lastly, using missingness patterns, we determined relationships between various labs that reflect clinical behaviors.
Conclusion
This work elucidates how missing data patterns in EHRs can be leveraged to identify quality control issues and relationships between laboratory measurements. Missing data patterns will allow sites to attain better quality data for subsequent analyses and help researchers identify which sites are better poised to study particular questions. Our results could also provide insight into some of the biological relationships between labs in EHR data for COVID-19 patients.
Article notes
Competing Interest Statement
The authors have declared no competing interest.
Funding Statement
MM is supported by National Center for Advancing Translational Sciences (NCATS) UL1 TR001857. DAH is supported by NCATS UL1TR002240. WY is supported by National Institutes of Health (NIH) T32HD040128. BJA is supported by National Heart, Lung, and Blood Institute (NHLBI) U24 HL148865. WGL is supported by National Library of Medicine (NLM) R00LM012926. YL is supported by R01LM013337. SNM is supported by NCATS 5UL1TR001857-05 and National Human Genome Research Institute (NHGRI) 5R01HG009174-04. GSO is supported by NIH grants U24CA210967 and P30ES017885. LPP is supported by NCATS CTSA Award #UL1TR002366. SV is supported by NCATS UL1TR001857. GMW is supported by NCATS UL1TR002541, NCATS UL1TR000005, NLM R01LM013345, and NHGRI 3U01HG008685-05S2. ZX is supported by National Institute of Neurological Disorders and Stroke (NINDS) R01NS098023 and R01NS124882. QL is supported by National Institute of General Medical Sciences (NIGMS) R01GM124111 and NIH National Institute on Aging RF1AG063481. DLM is supported by NCATS UL1-TR001878. JHH is supported by NCATS UL1-TR001878.
Introduction
The increasing availability of electronic health record (EHR) data has led to the burgeoning use of these data in various domains, including the identification of disease phenotypes and the clinical course of disease. Most recently, the EHR has been used as a rich source of data for characterizing the trajectory of the Coronavirus Disease (COVID-19, or simply COVID) that is caused by the SARS-CoV-2 virus. However, it is commonly acknowledged that EHR data often require rigorous processing and cleaning before they are of usable quality, thereby presenting considerable challenges to those using these data for research, quality improvement, or disease surveillance. Issues such as data availability [1–3], data recording or format inconsistencies [4,5], temporal changes in data policies [5,6], poorly standardized free-text [1,6], lack of interoperability between EHR systems [6,7] and diagnostic coding errors [1] all impair the usability of EHR data. Moreover, the most frequently reported barrier to EHR usability is missing data or data which are expected to be in the record but are not [3,4,6,8–12].
There is an important difference in the way that clinicians and biomedical informaticians tend to view missing data. To a clinician, data is considered missing if a laboratory test was supposed to be conducted and its value recorded, but for some known or unknown reason, it is absent from the medical record. Therefore, if a test is measured and recorded once per week according to protocol, and this is carried out without issues, there is no missing data for this particular test. On the other hand, if an informatician is carrying out a time series analysis that requires a measured value each day, that laboratory test will be considered missing for the 6 days it was not measured. This study was originally motivated by the desire to carry out a time series analysis; but with laboratory tests collected at different frequencies, the issue of missing data and how to deal with it needed to be addressed. As we delved into this, we realized that missing data itself could be leveraged to learn more about EHRs, nuances across sites, ordering patterns, and relationships between the labs themselves.
Data is missing in the EHR for two principal reasons. First, a laboratory test might have been ordered, but the result is missing from the record. Although important for ascertaining the quality of reporting systems, characterizing this type of missing data is difficult without access to clinical notes and ordering systems in the EHR. The second reason missing data is that a laboratory test was never ordered, or where a test was ordered and resulted, but for some reason was not resulted for some time or ever again during a hospital stay, and thus a result would not be expected in the EHR. Such missing data should not be considered a direct measure of EHR data quality, since there are many factors, often clinical, that determine when and if a test result is absent from the record. We focus here on this type of missing data, and we propose that the absence of data can be informative, and that patterns of these missing data can be considered as informative missingness.
Patterns of Data Missingness
Missing data is typically characterized according to three commonly accepted missingness patterns. The first is missing completely at random (MCAR). In this pattern, the missingness of a variable is not associated with any observed or unobserved variables, including the variable itself. An example would be where responses to a survey question about smoking status is not present on some proportion of respondents because the question was asked (or not) in a truly random fashion; in other words there is no nonrandom pattern of missingness. In the second pattern, missing at random (MAR), the missingness of a variable is associated with the value of another observed variable. For example, responses to a question on smoking status are dependent on one’s occupation, resulting in a missing value for smoking when a respondent notes that their occupation is in health care. Finally, when data are missing in the EHR, it is often missing not at random (MNAR). A variable may be missing because of the value of the variable itself. For example, a smoker may be less likely to answer a survey question about smoking status because they are a heavy smoker. Note that it is not known whether or not the respondent is, in fact, a smoker, and that would not be known because they did not answer the question. In other words, the probability of determining if a respondent is a smoker is depending on the value of the smoking question. For this reason, such missing data are nonignorable, which implies that such data violate assumptions for imputation and need to be considered (encoded) explicitly as missing data for purposes of imputation.
However, we contend that there is a special case of MNAR, where the missing data are informative, in this case indicating a clinician’s assumption and a decision not to order a test subsequent to the previous one. This type of missingness has been referred to as structurally missing data [13], in that there is a logical, non-random reason the data are missing. However, we refer to this pattern as informative-missing not at random (I-MNAR). In this pattern, the missingness of a variable is dependent on the value of the variable, like MNAR, but may also be influenced by the value of other variables as well, whether they are observed or unobserved. It is a pattern commonly seen in the EHR, where once a normal laboratory result is obtained, no further assays of the same type are present in the record. The absent results indicate that the decision not to order the test after the normal result was likely due to the normal value itself, but it could be that the values of other variables (such as other laboratory tests or clinical assessments) are taken into account during the decision-making process. Thus, the absence of laboratory results after a given result is informative, perhaps about the severity of the disease, the availability of the test, practice guidelines, or clinician preferences. An example of this is in [14], where the recording of rheumatoid factor test results in the EHR was found to be missing when a test result was negative. In other words, a test was not ordered because it was assumed that the test would be negative based on a prior result of the test.
Informative Missing-Not-At Random in EHR Studies
The I-MNAR pattern has been investigated in the literature, albeit under a different nomenclature, typically referred to simply as “informative missingness”. For example, in [15] it was noted that missing data are often correlated with a target variable, such as outcome. Informative missing data has been identified in genotype analysis and genetic association studies [16–21], longitudinal cohort studies [22,23], meta-analyses [24–29], exposure assessment in case-control studies [30], and particularly in studies using EHR data [31].
The goal of this study was to identify patterns of missing laboratory tests that might suggest levels of disease severity or other factors, such as patient sex or hospital characteristics that could influence the availability of laboratory data in the EHR. From this study, we hope to determine if these missing data fit an I-MNAR pattern and interpret this informativeness. Accordingly, we focus here specifically on the laboratory data patterns found in our examination of EHR data in a large international federated data consortium.
Setting
For this study, we used the resources of the Consortium for the Clinical Characterization of COVID-19 by the EHR (4CE). The 4CE consortium includes 342 hospitals in eight countries with patients who have been hospitalized for COVID. The 4CE uses a federated data model and predictive analytics framework in a hub-and-spoke configuration. Specifically, all 4CE-contributing academic medical centers (spokes) query and standardize EHR data elements using a COVID ontology, and apply analysis locally to their COVID datasets, and then provide aggregate statistics to a coordinating academic medical center (hub). This agile, rapid, and privacy-preserving data sharing approach has efficiently and effectively supported several COVID studies over the last two years [32–37]. We examined EHR data from 15 4CE-participating sites representing 232 hospitals in the United States and Europe comprising 69,939 patients for this study. For sites with multiple hospitals, we assume that practice patterns are similar across the hospitals within those sites.
This manuscript is structured into two main components; First we use patterns in missingness for identification of QC issues; second, we delve deeper into the relationships between patterns of missingness between laboratory measures. After this introduction, we describe our methodology for defining the sample population, the variable set, and the analytic methods used to quantify and describe missing values. Next, we present the results as characterization of the distribution of missing data by laboratory test, stratified by sex, hospital site, and disease severity, across different time periods from admission through 60 days thereafter. We also present the results of an analysis that seeks to investigate the patterns of missingness with regard to pairs of laboratory tests. Finally, we consider a topic model analysis that clusters groups of labs together based on missingness patterns.
Methods
This retrospective observational study of EHR was reviewed and approved by the ethics and institutional review boards for all participating 4CE contributing sites.
We analyzed laboratory test ordering patterns in EHR data for 69,939 patients from 232 hospitals across three countries from the 4CE consortium for the period of 1 January 2020 through 30 September 2021. 4CE consortium contributing sites are described in Table 1.
The inclusion criteria include a positive COVID-19 polymerase chain reaction (PCR) test on or during admission to the inpatient setting. Only data from a patient’s first COVID-19 admission was considered; subsequent admissions were not included in the analysis. We collected results from 16 laboratory tests conducted over the entire admission. We selected these labs because (1) their abnormal values have been associated with worse outcomes among COVID patients in the literature and in our own 4CE mortality risk prediction models and (2) their ability to reflect acute pathophysiology of COVID-19 patients. The clinical significance of these tests and and common ordering practice is described in Supplementary Table 1 [38–40].
Results
1.Patterns of missingness each laboratory test across sites
Figure 1 shows that the missingness across labs varies widely across sites and lab measures. The larger variability in the number of missing values per patient Figure 1 (A) compared to the average proportion missing per patient Figure 1 (B) across sites is possibly due to differences in the distribution of the length of stay and sample size at particular sites. Nonetheless, the number of missing values per patient still informs the exact number of missing tests and is useful for identifying sites with more valid values which might be required for some analyses. To account for the length of admission, the proportion missing (Figure 1 (B)) is normalized against the total number of admitted days for each patient. Generally, creatinine and leukocytes show the lowest number and proportion missing compared to all other tests. This is closely followed by bilirubin, albumin, ALT, and AST, which have a considerably lower number of average missing values per patient and a lower proportion missing (throughout patient admission) across most of the sites. Lymphocytes and neutrophils show a similar level of missingness as the above labs except for a visibly higher measure of missingness in Site 5.
Figure 2(A) shows the difference in proportion missing between males and females for each lab. Generally, across all labs, the difference in missingness by sex is varied in both directions across all sites. We observe the largest difference in missingness between males and females at Site 2, with some female patients having much more missingness for D-dimer, PT, and leukocytes. We also observe some slight deviations at some of the sites with more missingness among females in neutrophils at Site 12 and more missingness among males at Site 13 in AST and bilirubin. It is important to note that Site 2 has the smallest sample size (n=162), and Sites 11-15 have very few female patients (ranging from 5.2-6.4 percent). Thus deviations are likely a result of variability in the data. Beyond these findings, we observe that most sites are well-balanced across the different sexes with regard to missingness in the data.
2.Patterns of missingness of laboratory tests by patient severity
Next, we quantified the difference in proportion missing between severe and non-severe patients across the entire cohort in Figure 2(B). We hypothesized for this experiment that there would be a higher proportion missing in non-severe patients. We observe in Figure 2(B) that most deviations occur at Site 2 and Site 4. We note more missing data in severe patients at Site 2 for fibrinogen, D-dimer, and CRP, and more missing procalcitonin results in non-severe patients. These attributes could be explained by the fact that Site 2 has a smaller sample size and is a pediatric hospital, hence the current severity definition might not be suited for its patients. At Site 4, we observe more missing data in severe patients for D-dimer, CRP, ferritin, albumin, AST, ALT, bilirubin, and creatinine while we observe more missing procalcitonin data in non-severe patients. Site 4 does not have an intensive care unit (ICU) which could explain the higher proportion of missingness in severe patients. Other than these two sites, the remaining sites seem to be well-balanced across levels of severity with regard to missingness in the data over the whole admission period.
Figure 3 shows the difference in proportion missing between severe and non-severe patients for the initial three days of admission. For the majority of the sites, we see more missingness in the non-severe group as opposed to the severe group. We see most deviations from this trend in Site 4, with more missingness in the severe group; and a variation of trends at Sites 2 and 9.
To understand the differences in proportion missing across admission days, we first modeled the proportion missing across admission days and retrieved the rates of change (beta values) for both severe (βsevere) and non-severe patients (βnon-severe) separately. Then, we took the difference in the rates of change in proportion missing for severe and non-severe patients to draw up the heatmaps across sites and laboratory tests (Figure 4). We considered temporal trends in three different time frames: ten days and 30 days for the short-stay patients and 60 days for the long-stay patients. Investigating the trends of missingness across different time frames allows for the visualization of trends that are specific to the time scales.
From 0-10 days, we observe that while most sites show a higher rate of change in proportion missing in non-severe patients (blue cells), Sites 6, 7, and 9 had multiple labs that show the opposite trend with a higher rate of change in severe patients (shown in green) (Figure 4(A)). Site 4 also stood out with a much more pronounced difference, with a much higher rate of change in proportion missing in non-severe patients compared to severe patients (Figure 4(A)).
From 0-30 days, we observe a similar trend with a higher rate of change in proportion missing in non-severe patients across the same sites with the addition of Site 4. Also, Site 8 showed a varied, but pronounced difference in the rate of proportion missing across some labs (Ferritin, LDH, Lymphocytes, Albumin, Creatinine) whilst showing a higher rate of change in proportion missing in non-severe patients whilst other labs (CRP, AST, ALT, Bilirubin, Leukocytes) showed a higher rate of change in proportion missing in severe patients (Figure 4(B)).
For the long-stay analysis, we included only sites that provided data for patients with hospital stays of at least 60 days. For this, we required sites to have at least three patients with data through 60 hospital days. This led to a shortlist of 10 sites having sufficient data (at least three patients) for this analysis. Figure 4(C) that most sites and labs show a greater rate of change in severe patients compared to non-severe patients (green cells). This is with the exception of Sites 1 and 4 with labs that are showing a greater rate of change in proportion in non-severe patients (Figure 4(C)). The following temporal line plots display the actual proportions of missingness at various time points for severe and non-severe patients for several labs and can help explain some of these findings.
For example, Troponin also shows a higher rate of change in severe patients across many sites across all time scales. One trend might be that a certain lab has a consistently higher rate of change in severe patients as opposed to non-severe patients (aka mostly green throughout the heatmap in the different time intervals) (Figure 4). We see that this is the case for Troponin, as shown in Figure 5. Troponin initially has more missingness than all other labs. Our range for possible missingness proportions is limited to [0,1] and initially we see more missingness in the non-severe group as opposed to the severe group. Thus, many sites hit a maximum proportion of missingness much faster in the non-severe group, making the rate of change smaller than it would be for severe patients. We also see more variability in the non-severe group out to later days in the hospital admission.
Another trend is that a test may initially have a higher rate of change in non-severe patients early on, but the rate of change is higher in severe patients out to 60 days. We see that this is the case for Ferritin, as shown in Figure 6. Initially, we see a stronger increase in missingness in the first 10 days for non-severe patients and this remains to be the case for many sites in 0-30 days. However, again, patients in the non-severe group reach the maximum level of missingness faster than patients in the severe group. Thus, because the severe patients take more time to reach the maximum level of missingness, many sites will have an overall greater rate of change in the severe group.
Lastly, some labs might have a consistently larger rate of change across time intervals for non-severe patients (or mostly blue in the heatmap for the different intervals). We see in Figure 7 that this is the case for Leukocytes, where this lab has a consistently lower amount of missingness compared to the other labs. Because the upper bound for missingness is further away from the initial amounts of missingness, we see that in non-severe patients it increases more quickly and is more pronounced overall.
Based on the trends we observe in the heatmaps and some of the temporal line plots, we conclude that overall, there is more missingness in non-severe patients as opposed to severe patients over time across many of the labs in our dataset. This is what we might expect to see because clinicians might not test non-severe patients as heavily over time as severe patients who could be having more issues as their stay continues.
3.Patterns of missingness shared between pairs of laboratory tests
In identifying relationships in missingness across labs, we investigated laboratory test pairs that show a change in correlation across admission days (Figure 8). We then shortlisted test pairs with either a significant positive or negative change in correlation over admission days (Figure 8). Only test pairs that show reproducible correlations across sites are shortlisted. To obtain the rate of change, we fit linear models across the Spearman correlation values over time (Figure 8 (A) & (B)). Sites were included for a pair of the slope = |0.3| and standard error <= 0.25.
In Table 2, we observe several lab pairs that become positively associated during the later parts of admission. Notable lab pairs indicate suspicion of infection, coagulopathy including cardiac involvement, liver involvement, severe COVID-19 outcomes as well as rule out of differential diagnoses such as bacterial pneumonia.
In Table 3, we observe several lab pairs that become strongly associated during early admission. Notable lab pairs indicate suspicion of infection with or without kidney or liver involvement.
4.Patterns of missingness shared between groups of laboratory tests
To identify groups of labs that share patterns in missingness, topic modeling was done to arrive at a common set of labs across sites (Figure 9(A)). From 15 sites, we end up with 91 topics derived from LDA. We include unique groups of labs with at least nine intersections across topics in an pSet plot. Based on these criteria, we obtained four unique groups that are reproducible across sites presented in Figure 9(A). The black points represent the labs that are consistent across all topics found across sites while the pink and red ones represent labs that are only found in some but not all of the topics across sites (Figure 9(A)). We find that overall, the groups of intersections represent groups of labs that measure for similar issues in a COVID-19 setting, and thus are related to one another. Group 1 consists of labs that are commonly ordered together on a daily basis, measuring for renal issues and infection. Group 2 are likely tests ordered individually rather than as a group. They all are labs that have at times been thought to help prognosticate potential severity of COVID. Group 3 represents tests that are ordered together to assess liver function, and Group 4 represents coagulation studies that may or may not be ordered as a group.
Discussion
Through characterizing and exploring laboratory test missingness patterns across a multi-national COVID-19 study, we identified several insights and opportunities [1].
1.Missingness patterns can indicate data quality issues
Initially, we observed some discrepancies in the average number of missing values in tests that are usually ordered together in the same panel, e.g., Neutrophils and Lymphocytes, in Figure 1. The sites that had this discrepancy were Sites 1, 6, and Sites 11 through 15. At Site 1, these findings led to a correction of lab mappings to LOINC codes used at some sites and a subsequent correction in the data. We also found some discrepancies in the proportions of hepatic function tests at Site 13 and Site 10, with slightly more AST measurements as opposed to ALT. Mapping issues can be due to miscoding and issues with LOINC code specificity. Deriving knowledge of laboratory ordering patterns from clinical experts can help with developing more robust local data quality improvement checks. Researchers can work locally with their academic medical centers to provide useful solutions to address these inaccuracies and operationalize these data quality processes at scale.
2.Missingness patterns can indicate responses to hospital treatment capacity
During the global pandemic, several academic medical centers were not prepared for the large influx of patients with intensive care needs. Many created makeshift ICUs. Furthermore, some facilities may have implemented new triaging procedures. For example, the higher proportion of missing data in the initial three days of admission in Site 4 across almost all the labs could be attributed to the lack of an ICU in Site 4. Because patients with serious conditions might be in the process of having a transfer being arranged, it could explain the higher proportion of missing labs in the severe patients from Site 4.
3.Missingness patterns can change over time and at different rates
Although all patient groups exhibit some proportion of non-reported labs, not all patient groups consistently demonstrate the same rate of non-reported labs over time. Patients among the non-severe group reach the maximum level of non-ordered labs faster than patients among the severe group. For short-stay patients, we observe this change for labs like troponin. For ferritin, we observe a higher rate of change in non-severe patients. Intuitively, for non-severe patients, clinicians may become less concerned about a patient’s condition, leading to reduced ordering rates. As a corollary, the absence of test ordering can indicate improvements in the patient’s health.
4.Missingness patterns could be predictive of clinical outcomes
Missingness patterns in healthcare data also carries a signal within itself. Lack of a reported lab could possibly indicate how unimportant the lab is in the progression or monitoring of disease and hence be an important predictor for clinical outcomes as well. Because the reporting patterns between labs are cross-correlated (groups of labs with shared reporting (and non-reporting) patterns as shown in topic modeling Figure 9), the removal of one test’s missing values from a model could potentially affect model performance as well. Furthermore, using two different data-driven methods -- spearman correlations over time as well as topic models -- we observed common clinical themes among lab pairs and sets including coagulation, infection with renal involvement, and liver involvement.
5.Computational approaches for addressing ordering patterns
Missingness patterns could inform which variables would benefit from imputation for future studies. Including them as variables could be problematic because it is not exactly independent of variables. A number of approaches to dealing with informative missingness have been reported. These include using a Monte Carlo Expectation-Maximization simulation series that incorporates within-subject autocorrelation with a latent autoregressive process for longitudinal binary data [43], a Bayesian shrinkage approach to improve estimation of sparse patterns [44], and the use of an informative ordering pattern odds ratio [45–47]. Continued work in temporal pattern estimation in the face of informative missing data will investigate these methods in the context of these COVID consortium data [48].
This study has several limitations.
- Order sets were not investigated. The patterns of missingness for laboratory and other tests and procedures were likely influenced by standing order sets. Furthermore, order sets are likely to change over time as knowledge about COVID and therapies for it evolve over time. We plan to correlate patterns of missingness with the content of order sets and how they change over the course of the pandemic.
- Missingness was not correlated with secular trends of the pandemic. As the pandemic has evolved, there have been a number of irregular cycles in the epidemic curve, with marked changes in disease incidence associated with the Delta and Omicron variants. These changes are likely associated with corresponding changes in order sets and clinical practice. We will investigate patterns of laboratory test missingness as they may reflect clinical practice patterns which in turn may reflect the undulation in the incidence of COVID over time.
- Type of patient care unit was not captured. It is possible that test ordering patterns could be influenced by the type of unit a patient was on at given points in time. For example, order sets may be used less frequently in intensive care units than in medical units, even those dedicated to COVID. Our future work will include an examination of “unit effect” in a temporal context, which will be important as patients move between acute and less-acute care settings in a given hospitalization.
- Focus on missing test values, as opposed to missing orders. We assumed that a missing laboratory test result indicated that the test was not ordered. While a reasonable assumption, a more accurate indicator of ordering behavior would be to capture orders in addition to test results. We will investigate the feasibility of obtaining these data from the EHR in future work.
- Severity definitions might change over time. When patients stay in the hospital for long periods of time, it is possible that their severity changes. Thus, there are limitations in looking at Figure 4 due to the fact that patients initially labeled as non-severe that remain in the hospital out to 60 days might not actually be non-severe at that point.
Conclusion
In this study, we investigated and demonstrated how characterization of missing data patterns in EHRs, particularly lab results, could support various steps in scientific study ranging from data quality to hypothesis generation. Furthermore, missing data patterns will enable consortia to identify which sites are better poised to study particular questions and potentially inform the use of imputation methods for addressing these challenges. Finally, our results may provide insights into some of the biological relationships between labs in EHRs data for COVID-19 patients.
Data Availability
All data produced in the present study are available upon reasonable request to the authors.