Add ECDS data to backend

[inglesp]

Split out from https://github.com/opensafely/cohort-extractor/issues/163.

We want:

• Date of attendance
• No of attendances in X period of time
• COVID attendance (from codelist)
• Discharge destination

[inglesp]

Proposed API:

# date of attendance
patients.admitted_to_emergency_care(
    on_or_after="2020-02-01",
    returning="date_admitted"  # or "binary_flag"
)

# no of attendances in in period of time
patients.admitted_to_emergency_care(
    between=["2020-02-01", "2021-02-01"],
    returning="number_of_matches_in_period"
)

# covid attendance
patients.admitted_to_emergency_care(
    on_or_after="2020-02-01",
    with_these_codes=[...],
    returning="date_admitted"  # or "binary_flag"
)

# discharge destination
patients.admitted_to_emergency_care(
    on_or_after="2020-02-01",
    returning="discharge_destination"
)

@CarolineMorton, @alexwalkercebm, does this look like it captures everything you might want to do with ECDS data?

@evansd, @sebbacon, does this fit in with existing patterns? Does anything look unimplementable?

[CarolineMorton]

Yes this looks exactly what we need. Thanks @inglesp

[HelenCEBM]

Some notes on this here: https://github.com/opensafely/outcomes-notebook/issues/3

Some points and questions:

• I wouldn't use the phrase admitted_to_emergency_care because that usually means being admitted to a ward/bed from A&E and so may cause confusion with inpatient records - would suggest attended_accident_and_emergency or similar is more appropriate.
• Just to check, can the above items be combined? E.g. Patient had a covid attendance and was admitted from that particular attendance?
• do we want to have a separate "likely covid" category based upon "shortness of breath" and similar diagnoses? Or will that be built into the "covid attendance" item in the study def as needed?
• a data quality flag might be helpful, e.g. if a trust does not complete diagnosis codes then the flag for covid attendance cannot be completed.
• I doubt we will need all the possible outputs from discharge_destination - there are a dozen or so but this could be simplified to a flag for "admitted/transferred" (vs "not admitted")
• For aftershocks we will need to identify many possible diagnosis codes other than Covid.

[inglesp]

I wouldn't use the phrase admitted_to_emergency_care because that usually means being admitted to a ward/bed from A&E and so may cause confusion with inpatient records - would suggest attended_accident_and_emergency or similar is more appropriate.

This is helpful.

Just to check, can the above items be combined? E.g. Patient had a covid attendance and was admitted from that particular attendance?

Yes.

do we want to have a separate "likely covid" category based upon "shortness of breath" and similar diagnoses? Or will that be built into the "covid attendance" item in the study def as needed?

Will these categories come from a codelist (like patients.with_these_clinical_events) or will they be hardcoded (like patients.most_recent_bmi)? I would expect the former, in which case the API doesn't need to know about "likely covid" etc.

a data quality flag might be helpful, e.g. if a trust does not complete diagnosis codes then the flag for covid attendance cannot be completed.

What would the API for this look like?

I doubt we will need all the possible outputs from discharge_destination - there are a dozen or so but this could be simplified to a flag for "admitted/transferred" (vs "not admitted")

OK. Which field, and which values in that field, should we use to compute that flag? (see https://github.com/ebmdatalab/tpp-sql-notebook/blob/master/notebooks/tpp-schema.ipynb)

For aftershocks we will need to identify many possible diagnosis codes other than Covid.

That's what with_these_codes=[...], supports.

[HelenCEBM]

a data quality flag might be helpful, e.g. if a trust does not complete diagnosis codes then the flag for covid attendance cannot be completed.

What would the API for this look like?

Perhaps we need another category when checking against a list of diagnosis codes: yes (patient had one of the relevant diagnoses), no, or "unable to tell". However we should probably consult the ECDS data quality report to figure out what to do here. It's possible that the discharge destination is also poorly completed by some trusts.

I doubt we will need all the possible outputs from discharge_destination - there are a dozen or so but this could be simplified to a flag for "admitted/transferred" (vs "not admitted")

OK. Which field, and which values in that field, should we use to compute that flag? (see https://github.com/ebmdatalab/tpp-sql-notebook/blob/master/notebooks/tpp-schema.ipynb)

• Field Discharge_Destination_SNOMED_CT.
• Here is the list of discharge destinations. Perhaps Henry Drysdale can help here (I can't seem to tag him though...). We need to include admitted to ward, critical care, high dependency, and possibly transferred (which can indicate an admission, but may also be someone needing an A&E department with an MRI machine for example, we can't distinguish). We could also consider "death" in the same category (admitted to mortuary)

[inglesp]

@HenryDrysdale

[HenryDrysdale]

@HelenCEBM @inglesp Thanks. Helen's suggestion of "admitted" vs "not admitted" sounds sensible.

I think "not admitted" should mean the patient attended A&E, but didn't require further secondary care given their clinical status at that time. There are a few codes for discussion:

"Emergency department discharge to ambulatory emergency care service"

• This means the patient is going home, but coming back to hospital for a clinical review, some blood tests, or a scan, normally within a week or so.
• This implies the patient is well enough to go home at the time of the A&E attendence, so I think this should come under "not admitted"

"Discharge to hospital at home service"

• This means the patient is going home (most likely because they are well enough), but requires some treatment or care that can't be managed by primary care (e.g. IV antibiotics)
• Again, I think this indicates low disease severity, and should come under "not admitted"

"Patient transfer, to another health care facility"

• This is likely to mean the patient was really unwell and needed to go to a hospital with an ICU or a cardiac cath lab, i.e. they needed lots more secondary care because they had severe disease. It's unlikely to mean they were transfered somewhere for a routine scan and then went home.
• So I think this comes under "admitted"

"Emergency department discharge to emergency department short stay ward"

• This quite often happens to patients who are well and don't need admission, but are "admitted" so they don't breach the 4-hour A&E waiting target.
• However, this would be impossible to ascertain, so I think we need to assume this means the patient needs further secondary care, and put this under "admitted"

"Not admitted" would then include:

• Emergency department discharge to ambulatory emergency care service
• Patient discharge, to legal custody
• Discharge to nursing home
• Discharge to residential home
• Discharge to home
• Discharge to police custody
• Discharge to hospital at home service

All other codes indicate severe disease and would come under "admitted". As Helen mentioned, "Admission to mortuary" could be considered separately.

[sebbacon]

Just to note we should capture this documentation-type info in docstrings or elsewhere!

[HelenCEBM]

Thanks @HenryDrysdale

"Patient transfer, to another health care facility"

This is likely to mean the patient was really unwell and needed to go to a hospital with an ICU or a cardiac cath lab, i.e. they needed lots more secondary care because they had severe disease. It's unlikely to mean they were transfered somewhere for a routine scan and then went home. So I think this comes under "admitted"

I would agree provided we limit attendances to those in major A&E units (type 1 & 2). I'm aware of people being transferred from MIUs to major hospitals because they have a suspected thing that needs a particular scanner not available in the MIU - but they may remain in A&E in the second hospital before being sent home. However, in such cases we should (in theory!) have a record of their attendance the second hospital, which we should be able to select instead to find the patient's ultimate destination (admitted/not).

[HenryDrysdale]

@HelenCEBM Yes good point - I hadn't thought of MIUs.

A common example of this is probably CTPA scans to rule out PE: I think all A&E departments will have CT scanners, but MIUs probably don't. So patients who go to an MIU with ?PE would need to be transfered for a CTPA, but it's very often negative and the patient then goes home.

The trouble with limiting to type 1 and 2 is patients are often referred straight from MIUs (and similar) to the hospital medical teams for further assessment or admission, bypassing A&E. So we might miss these.

[HelenCEBM]

I suppose we should consider what it means if someone with covid (or suspected) is in an MIU and they get transferred. Are they likely be just having more tests and likely sent home? Or more likely in need of ICU?

[HenryDrysdale]

Yes I agree. I'm really not sure what's more likely, and it's not clear to me from anecdotal experience. I think ICU is unlikely from an MIU, but an assessment at hospital and short admission for monitoring / fluids is very common. (More serious admissions also common).

So in summary, we're unsure what the code "Patient transfer, to another health care facility" says about the severity of disease for COVID patients in the ECDS dataset, because:

1. Lots of these patients may be very unwell, and transfered to hospitals for escalation of care, such as HDU or ICU.

2. On the other hand, lots of these patients may be relatively well, and transfered from an MIU to hospital for a precautionary scan before being sent home.

On balance, I think of all uses of this code in ECDS, more are likely to represent unwell patients than well patients.

[inglesp]

This is all useful, thanks @HelenCEBM and @HenryDrysdale.

In the first instance, I don't think we should hardcode the codes that we're interested in, either for diagnoses or for discharge destinations. These might be different for different studies, and may change over time. We can always add them later if the same patterns keep coming up. Instead, we can have eg:

was_discharged_to_care_home_or_nursing_home=patients.attended_emergency_care(
    on_or_after="2020-02-01",
    returning="binary_flag",
    discharged_to=[
        "306706006",  # care home
        "306689006",  # nursing home
    ],
),

And for larger lists of codes for diagnoses, we can use existing codelist functionality.

[HelenCEBM]

Just to note, I wouldn't rely on those particular discharge codes to be accurate - largely staff will use the code for "discharged home" which may or may not be a care home. No one would newly enter a care home directly from A&E, and we have better ways of determining care home residency. So I think admitted vs not admitted is the best use of these codes.

[HelenCEBM]

Diagnosis codes for people attending A&E with a positive SGSS test result within a few weeks of the attendance, up to 18 April, with approximate counts. cc @HenryDrysdale

Notes

• several codes may appear in each attendance.
• SGSS tests are typically in hospital (inpatients) and healthcare workers, tests will rarely be performed in A&E.
• Other variants of the COVID code may also be used (perhaps after our current date cut-off)

DiagnosisCode	diagnosis_code_count	SNOMED_Description	Possible Covid?
50417007	3000	Lower respiratory tract infection (disorder)	yes
1240751000000100	1500	Coronavirus disease 19 caused by severe acute respiratory syndrome coronavirus 2 (disorder)	yes
91302008	700	Sepsis (disorder)	yes
233604007	650	Pneumonia (disorder)	yes
278516003	350	Lobar pneumonia (disorder)	yes (less likely)
306206005	300	Referral to service (procedure)	no
54150009	300	Upper respiratory infection (disorder)	yes
68566005	300	Urinary tract infectious disease (disorder)	no
13645005	200	Chronic obstructive lung disease (disorder)	yes (with exacerbation)
281900007	200	No abnormality detected (finding)	no
398447004	200	Severe acute respiratory syndrome (disorder)	yes
14669001	150	Acute renal failure syndrome (disorder)	no
2776000	150	Delirium (disorder)	no
6142004	150	Influenza (disorder)	no
195967001	100	Asthma (disorder)	yes (with exacerbation)
42343007	100	Congestive heart failure (disorder)	no
49436004	100	Atrial fibrillation (disorder)	no
394659003		Acute coronary syndrome (disorder)	no
422588002		Aspiration pneumonia (disorder)	no (but possible confusion with covid)
12463005		Infectious gastroenteritis (disorder)	no

[HelenCEBM]

@HenryDrysdale @rozeggo @wjchulme @amirmehrkar cc @CarolineMorton @inglesp

We discussed how to select "Covid A&E attendances" (ie. people with severe enough disease to warrant urgent medical care), using diagnosis codes, starting from the list above.

• None of the diagnosis codes alone are definitively Covid - especially in early weeks when testing was not widely available and was likely performed in hospital after admission. Coding behaviour will vary between trusts.
• We have categorised the above into which ones can be a likely sign of Covid, potentially in combination with other information (this is intended as an inclusion list, though it could be possible to use some codes in an exclusion list, e.g. Influenza).

Other useful information to consider:

• Being admitted to a ward is a useful indicator for severity (but if we are just interested in admissions, this dataset may be best superseded by inpatient data in due course, which should be more complete for people who bypassed A&E, and more detailed)
• Referral source (e.g. 111/GP) or arriving by ambulance could be useful to indicate severe disease. E.g. It is likely that children and younger people who present with Covid-like symptoms independently were often sent home to recover, whereas being referred in indicates an medical assessment has already indicated that treatment is required. However, this is unlikely to be useful for distinguishing Covid from similar conditions.
• We did not think of any useful distinguishing investigations or treatments that might be recorded in A&E (e.g. X-rays, blood tests, oxygen) as many could be carried out equally in Covid and non-Covid patients (but again may indicate severity).
• Chief complaint may complement the diagnosis codes to help distinguish likely Covid from other conditions. Helen to investigate.
• The best method will probably be to assert that the patient had a positive test result around the time of the attendance, so we may need to build in SGSS results into the backend (though later in the pandemic this may require primary care record input to capture tests carried out in the community prior to attendance at A&E)

[amirmehrkar]

The SUS ECDS dataset processed by NHS England data store does not include the diagnosis qualifier code (table 16; row 5 and on SNOMED browser refset.

It is unlikely to be accurate anyway, as there are no validated point of care tests. Unless the patient was a readmission with know previous C19 result; however, it is possible clinically the qualifier confirmed may have also been used.

[HelenCEBM]

Top 20 chief complaints for everyone with a positive test near the time of their attendance are now within here and pasted below.

I can't find a comprehensive dictionary of all possible codes - @HenryDrysdale could you have a look?? The first is shortness of breath.

EC_Chief_Complaint_SNOMED_CT	Count	SNOMED_Description
267036007	5000	NaN
386661006	1000	NaN
230145002	850	NaN
13791008	550	NaN
394616008	300	NaN
29857009	300	NaN
21522001	250	NaN
49727002	200	NaN
40917007	200	NaN
427461000	150	NaN
3006004	100	NaN
127279002	100	NaN
422400008	100	NaN
62315008	100	NaN
82271004	100	Injury of head (disorder)
271594007	50	NaN
91175000	50	NaN
25064002	50	NaN
80313002	50	NaN
80394007	50	Hyperglycemia (disorder)

[HenryDrysdale]

@HelenCEBM great work.

Yes I tried before and couldn't find one - you can get the codes from the data quality dashboard but this isn't up to date. Can we ask NHS D for a list of currently used codes? Who could we speak to?

[HelenCEBM]

@inglesp any tips?

[HelenCEBM]

EC_Chief_Complaint_SNOMED_CT	SNOMED_Description
267036007	Dyspnea (finding)
386661006	Fever (finding)
230145002	Difficulty breathing (finding)
13791008	Asthenia (finding)
394616008	Unsteady gait (finding)
29857009	Chest pain (finding)
21522001	Abdominal pain (finding)
49727002	Cough (finding)
40917007	Clouded consciousness (finding)
427461000	Near syncope (disorder)
3006004	Disturbance of consciousness (finding)
127279002	Injury of lower extremity (disorder)
422400008	Vomiting (disorder)
62315008	Diarrhea (finding)
82271004	Injury of head (disorder)
271594007	Syncope (disorder)
91175000	Seizure (finding)
25064002	Headache (finding)
80313002	Palpitations (finding)
80394007	Hyperglycemia (disorder)

[HelenCEBM]

Summary of chat today with @wjchulme @HenryDrysdale and an A&E registrar:

• A&E coding is generally unreliable and certainly less so than inpatient records.
• Having any information on how patients were triaged into Red/Green covid zones is unlikely to be helpful as they could be triaged to covid-zone due to recent exposure to someone with symptoms without having any symptoms themselves.
• It might be common for people to be admitted for further investigation, discharged, and be re-admitted later with deterioration, so we might need to take this into account when selecting attendances/admissions.
• It is not possible to definitively identify covid-related attendances because there will be so many people presenting with similar symptoms (or people with dementia with poorly described symptoms) that cannot be distinguished, so we should take a conservative approach and rely on some external information such as primary care records and/or testing positive.
• Deterioration of symptoms leading to hospitalisation likely to be within 7-10 days of onset (to bear in mind when linking to other data).

Summary of conservative approach to identifying possible covid attendances:

• Patient has relevant diagnoses OR chief complaint (@HenryDrysdale )
• Patient did not have an injury (hopefully covered by the above criteria but we can additionally filter on injury_date if needed).
• Patient was admitted
• Patient had a positive test within 2 weeks before/ up to 1 week after their attendance OR had a record in primary care of suspected/confirmed Covid up to 2 weeks prior to their attendance (but not after, as these might represent people catching it in hospital and developing symptoms later)
  • (@inglesp these will need to be linked in the backend if possible)

Questions remaining:

• codelists to finalise
• Do we select include MIUs?
  • sometimes people will attend an MIU because it is the nearest to them, not based on severity of symptoms
  • There may have been some changes in how these were operated during covid, especially where a Trust has more than one (eg. use one as a covid triage centre and keep the other covid free)
  • People cannot be admitted if there is no suitable ward on site but they could be transferred directly to a ward in another hospital, bypassing A&E.

[wjchulme]

Looks great. Just to be explicit that this conservative approach is identifying attendances where the patient's condition is severe enough for admission, but will still miss those without a positive test (due to no test or poor test sensitivity).

Tom rang back just now and is increasingly sceptical about the utility of ECDS for identifying covid related attendance / admission due to: inconsistent/unreliable coding practices, especially for diagnoses (chief complaint slightly better as done by nurses, but there's only one code per patient); limited clinical info on investigations, bloods, etc; limited discharge destination info; general lack of specificity for all the available info even if it was complete and reliable.

So it's useful to see what we get from this, but if we want an intermediate clinical state that's worse than suspected covid at home but not as severe as an ICU admission we should think about other ways too.

@HenryDrysdale are you happy for me to make a start on the codelists and then you can update as necessary?

[HenryDrysdale]

@wjchulme Yes sounds great, thanks.

I agree with Tom's concerns, particularly around using investigations, bloods, discharge destinations etc. But I think our sensitivity for detecting COVID is better than it might be given that patients in our cohort: were unwell and admitted without an injury; had a positive covid swab close to admission; were admitted during the COVID-19 pandemic. It'll be interesting to see what we get, and we'll be open about limitations.

[wjchulme]

Formalising our approach in pseudocode:

IF diagnosis == "covid" THEN covid_admission = 1
ELSE IF (
    (test_pos_date BETWEEN(attendance_date-14, attendance_date+7) OR suspected_covid BETWEEN(attendance_date-14, attendance_date)) AND 
    injury_date == NULL AND
    discharge_destination == [admitted_for_further_care_codes] AND
    (diagnosis IN [covid_related_diagnosis_codes] OR complaint IN [covid_related_complaint_codes]) AND
    NOT (diagnosis IN [covid_ruleout_diagnosis_codes] OR complaint IN [covid_ruleout_complaint_codes])
) THEN covid_admission = 1
ELSE covid_admission = 0

I assume we need the ruleout codes as for instance complaint="fever+cough" and diagnosis="aspiration pneumonia" would be picked up as COVID. Or even diagnosis1="aspiration pneumonia" and diagnosis2="upper resp. infection". If so, we need a list for those too -- I wonder if for the ruleout list we should search amongst more than just top diagnoses as done before?

Alternatively, we're more strict with our choice of covid diagnosis codes and we ignore complaint:

IF diagnosis == "covid" THEN covid_admission = 1
ELSE IF (
    (test_pos_date BETWEEN(attendance_date-14, attendance_date+7) OR suspected_covid BETWEEN(attendance_date-14, attendance_date)) AND 
    injury_date == NULL AND
    discharge_destination == [admitted_for_further_care_codes] AND
    diagnosis IN [strictly_covid_related_diagnosis_codes]
) THEN covid_admission = 1
ELSE covid_admission = 0

But we risk missing a lot of cases then.

Either way, we need codelists for all the items in square brackets.

[Bengoldacre]

epic and magnificent.

re implementing the logic, just to reiterate a general principle: where there are no architectural blockages, it's always best if we can implement logic like this in the study definition, rather than the backend, as that leaves a brighter legacy of shared code for others.

[HelenCEBM]

Note this now applies to the new EC table which should replace ECDS

[HelenCEBM]

@inglesp @evansd this needs switching over to use EC instead of ECDS. I'm not sure where to find the fields we're currently using but if you point me in the direction I'm happy to help identify which fields need change of name

[inglesp]

@HelenCEBM the SQL gets built in this function.

Specifically, we're using:

• ECSD
  • EC_Ident
  • Patient_ID
  • Arrival_Date
  • Discharge_Destination_SNOMED_CT
• ECDS_EC_Diagnoses
  • EC_Ident
  • DiagnosisCode

[HelenCEBM]

Thanks @inglesp. Using tpp schema for reference:

1. ECDS can simply change to EC. The fields have the same names.

2. There are a couple of options for ECDS_EC_Diagnoses

   • Use table EC_Diagnosis - but this is one-line-per-attendance with fields EC_Dianosis_01, EC_Dianosis_02 etc unlike the previous table we were using (which had one row per diagnosis, i.e. potentially many rows per attendance).
   • Use the field Der_EC_Diagnosis_All in the main table (EC) which has a comma-separated list of diagnosis codes.

[rohinimathur]

Hi @inglesp I am using ECDS data currently for the ethnicity study. Could you please advise on how to update study definition to point to EC tables instead?

[inglesp]

Could you please advise on how to update study definition to point to EC tables instead?

You shouldn't have to make changes to your code. I'll let you know once the backend has been updated.

[inglesp]

There are a couple of options for ECDS_EC_Diagnoses

Grr, they're both less good than what we currently have. Will fix up now.

[HelenCEBM]

Grr, they're both less good than what we currently have. Will fix up now.

That's because TPP made the previous table! We would have had to work around for new vendors at some point anyway.

[inglesp]

I'll let you know once the backend has been updated.

@rohinimathur the backend has been updated, but you shouldn't need to make any changes to your code at all.

[rohinimathur]

@inglesp wow thanks. will just update the database end date in my analysis code. Much appreciated!