South Africa highlights the issue
The recent data from South Africa have highlighted an interesting problem with COVID-19 hospitalization data. Omicron in South Africa generated a strikingly large number of hospital admissions with COVID, but not for COVID. That is, patients who were admitted for some other treatment, but were found to have COVID-19 upon testing.
That change in case mix, in turn, messed up any inferences you might try to make regarding changes in hospitalizations. For example, the “average COVID-19 patient” in South Africa appeared to have vastly less severe respiratory disease, because the typical patient in that category isn’t in the hospital for a respiratory problem. Worse, news reporting that does not grasp this change in the mix of hospitalized patients will mis-represent the situation. As was done in South Africa, the popular media will incorrectly report on a surge in cases essentially unrelated to COVID-19, as if it were a surge in cases for treatment of COVID-19.
The root problem is that South Africa tracks counts of people hospitalized with any diagnosis of COVID-19. That’s a mix of people hospitalized for COVID, and people who were hospitalized for treatment of something else, who just happened to have an active COVID-19 infection.
So, what do we track, here in the U.S.? And will we see the same misinterpretation of the data, if the case mix changes here, as it did in South Africa under Omicron?
American terminology
I might as well introduce the American terminology now, because it makes distinguishing these groups a lot more precise.
The terms of art in U.S. hospital data are:
- Principal diagnosis of …
- Secondary diagnosis of …
- Any diagnosis of …
As defined in the U.S., the principal diagnosis is , “that condition established after study to be chiefly responsible for occasioning the admission of the patient to the hospital for care.” (That’s from the Uniform Hospital Discharge Data Set, which is the U.S. standard for reporting hospital discharge abstracts and claims.) Alternatively, some definitions focus on the condition that accounted for the majority of the resources used in the hospital stay. For almost all cases, it’s the reason you were hospitalized.
By contrast, a secondary diagnosis is some other condition that is present, but is not the reason for the hospitalization. These are often termed “comorbidities and complications”. Broadly speaking, a comorbidity is something you already had when you entered the hospital, even if it might only have been uncovered during the stay. E.g., long-standing diabetes present in a heart attack victim. A complication is something that you picked up in the course of the illness, generally related to what you were being treated for. E.g., atrial fibrillation as a permanent condition following recovery from a heart attack.
You lump together all cases, of either sort, under the term “any diagnosis of”.
In addition, U.S. hospital administrative data have two more relevant terms:
- Admitting diagnosis of …
- Present on admission …
The “admitting diagnosis” is what the medical staff believe to be the reason for hospital admission, at the time of admission. So, for example, “chest pain” is the most common admitting diagnosis for cases that are eventually coded with principal diagnosis of heart attack.
Finally, almost all hospitals treating U.S. Medicare patients have to flag the diagnoses that were present when the patient was admitted, versus conditions that only developed during the course of the stay. Medicare uses those flags to avoid paying for complications of care that were presumed to be caused by the hospital.
The issue raised by Omicron in South Africa
In U.S. terminology, South Africa reports a count of inpatient admissions with any diagnosis of COVID-19.
As long as the mix of cases with principal and secondary diagnosis remains constant, that’s not a problem. For example, if the mix is constant, and cases double, we know that the number of persons admitted for treatment of COVID-19 (principal diagnosis of …) has doubled.
But that apparently has not been true in South Africa, with Omicrion. We now know that under Omicron, the mix shifted enough that it drew the attention of medical personnel in that country. Under Omicron, the vast majority (76%) of COVID-19 hospitalizations are persons with secondary diagnosis of COVID. Only a small minority (24%) are persons with principal diagnosis of COVID, that is, persons actually in the hospital for treatment of COVID.
That high fraction of admissions “with” rather than “for” COVID-19 is consistent with Omicron being less severe than Delta. Less severe would mean that a higher fraction of infections were asymptomatic. And it’s those asymptomatically-infected people who are showing up to the hospital (for some other problem) and are now boosting the counts of persons hospitalized with any diagnosis of COVID-19 in South Africa.
Let me sum that up as follows: As I see it, the shift to much lower severity made the South African COVID-19 hospitalization data inconsistent over time. As Omicron grew, it created a surge in the count of cases with any diagnosis of COVID-19. That’s solely a consequence of having a lot of very mild (asymptomatic) infections in the population. Some of those will randomly show up at the hospital — think of a heart attack victim who just happens to have an asymptomatic COVID-19 infection. But this increase was picked up and reported as if it had been a surge in cases with principal diagnosis of COVID-19. And that does not appear to have been true.
As an afterthought, I’ll note that it makes sense for hospitals to track and report the “any diagnosis of” number, because that’s what they have to deal with, administratively. They have to isolate all their patients who have a COVID-19 infection, whether or not that infection is the reason for being in the hospital.
So that “any diagnosis of” statistic makes perfect sense from the standpoint of hospital administration. It’s just not the statistic you want, if you’re trying to figure out whether or not COVID-19 is putting a lot of people into the hospital.
What about the U.S. hospital data?
Near as I can tell, it boils down to this: Nobody has even considered being in a situation like South Africa, where the overwhelming majority of inpatient admissions with COVID are actually cases unrelated to the treatment of COVID.
In the Medicare program, in particular, so far, cases with any diagnosis of COVID were virtually all cases with indications of respiratory infections of some type. (That’s my reading of this this DHHS Office of Inspector General report. )
In other words, as I read it, so far, our hospitalized population with any mention of COVID-19 has, by and large, been a population that appears to have been admitted for and treated for COVID-19. Nobody in the U.S. appears to have set up their data reporting to deal with a large number of cases that are with COVID, but not for COVID.
This is not an idle observation, as we get ready to see Omicron take over as the dominant strain of COVID-19 in the U.S. If that new strain creates a far larger share of asymptomatic infections, we may or may not face the same problem as South Africa, depending on exactly what is being reported as our count of COVID-19 hospitalizations.
Turns out, this being the U.S.A., we report those cases every-which-way. State hospital associations report them as they see fit. The Federal government has two different reporting systems via CDC. Medicare does its own reporting. State health departments or hospital associations do their own reporting. And so on.
Let me now walk through that. I’m looking for two things. First, I want to know the potential for in influx of asymptomatic cases to disrupt the consistency of data reporting over time. Second, if possible, I’d like any data source that can provide the breakdown of principal versus secondary diagnosis of COVID-19.
DHHS Unified Hospital Dataset
This is what the CDC calls it, and it is arguably the data set of record for U.S. COVID-19 hospitalizations. You can find the state-level summary, and more-or-less exact descriptions of what should be reported, as well as other cuts of the file, on healthdata.gov.
This has two different measures, total patients currently in the hospital, and new patients admitted in the prior day.
For total COVID-19 patients currently in the hospital, it’s clearly “any diagnosis of COVID-19”. “31. Reported patients currently hospitalized in an adult inpatient bed who have laboratory-confirmed or suspected COVID-19. This include those in observation beds.” So our official measure of the stock of persons in the hospital is the same as South Africa: It’s anyone with any known (or suspected) diagnosis of COVID-19.
For daily new cases, it appears to be something like “present on admission”. The definition is given as “17. Number of patients who were admitted … inpatient bed on the previous calendar day who had confirmed COVID-19 at the time of admission … ” So that includes any diagnosis of COVID-19, but would appear to exclude asymptomatic individuals whose COVID-19 was only found via in-hospital testing.
CDC COVID-net
The CDC collects information on “COVID-associated” discharges from a sample of about 250 U.S. hospitals. This one counts any case admitted to the hospital within 14 days of a positive clinical lab test for COVID-19. This one will capture something similar to “any diagnosis of ” COVID-19. It won’t capture individuals whose COVID-19 status was revealed only by in-hospital testing. And it could very well capture some people who are completely over their COVID, they just happened to be hospitalized within 14 days of a positive test.
Medicare program data.
Medicare has been producing a detailed summary of their COVID-19 cases based on claims data. For hospitalizations, their methodology document makes it clear that they took any diagnosis of COVID on inpatient claims.
For Medicare, at least, this issue of principal versus secondary diagnosis appears to have been a non-issue so far in the pandemic. At least, that’s my reading of this DHHS Office of Inspector General report. Those guys are usually quite precise, and they didn’t even bother to try to distinguish cases where COVID-19 wasn’t the principal diagnosis.
State hospital systems.
I can’t check them all but let me try a few.
California: Appears to follow the US DHHS Unified Hospital Dataset format.
Florida: Weekly report does not track hospitalizations. It looks like all news reporting there relies on the DHHS data.
New York: Impossible to tell. Near as I can see, there is no documentation as to what, exactly, they are counting.
Texas: Impossible to tell, but appears to be “any laboratory-confirmed diagnosis of ” COVID-19.
Pennsylvania: Impossible to tell, but they focus on available of airborne infection isolation ward beds. So it’s a good bet that it’s “any diagnosis of” COVID-19.
Conclusion
As far as I can tell, the U.S. data reporting systems for COVID-19 hospitalizations are not robust to an influx of asymptomatic infections in the population. If we get a lot of asymptomatic COVID-19 cases under Omicron, we are going to see an increase in some aspects of our “COVID-19” hospitalization counts.
Arguably, the most important data file for tracking U.S. COVID-19 hospitalizations is the DHHS Unified Hospital Dataset. There, the count of new admissions should be robust. Hospitals only count individuals known to be COVID-19 positive at time of admission. But the count of all persons currently hospitalized would NOT be robust. Any person found to have COVID-19, after admission, would be classified as a COVID-19 patient. We’d have the same issue with our patient counts as South Africa has with its patient counts.
The magnitude of this presumably will vary by type of admission: Emergency, urgent, or elective (scheduled).
For scheduled (elective) admissions, I believe that it is now the norm for hospitals to require patients to get tested prior to admission. Thus, for elective admissions, this isn’t an issue. Those with either symptomatic or asymptomatic COVID-19 infections will not be able to be admitted to the hospital.
For emergency admissions, no such screening is possible. In the U.S., in 2018, per the AHRQ HCUP hospital database, 55% of hospital admissions occurred via the emergency room. (And so were presumed to be emergency admissions). It’s a fair bet that almost all hospitals test such patients for COVID-19 now, upon admission. Any patients admitted this way, with any type of active COVID-19 infection, will be identified in the “any diagnosis of” COVID population.
The upshot is that for the majority of U.S. hospital admissions, both symptomatic and asymptomatic COVID-19 cases will be able to enter the hospital for treatment other than for COVID-19. In other words, the data reporting situation in the U.S. is really not all that different from the situation in South Africa. Asymptomatic COVID cases can enter the inpatient side of the hospital in large numbers, if such cases are prevalent among emergency admissions. They will then be reported in the “stock” number — the count of persons currently in the hospital, with COVID-19. The sole difference is that the “new admissions” numbers should exclude such patients, because their COVID-19 status would not be known at the time of admission.
