Phenotype submission - Guillian Barre Syndrome

Gowtham_Rao · November 29, 2022, 7:26pm

Clinical Description

Guillain-Barre syndrome (GBS)

Authoritative source:

Abstracted from authoritative source:

Overview: Guillain-Barre syndrome (GBS) is a rare but serious post-infectious immune-mediated neuropathy. It results from the autoimmune destruction of nerves in the peripheral nervous system causing symptoms such as numbness, tingling, and weakness that can progress to paralysis. Guillain-Barre syndrome (GBS) is the most common cause of acute, flaccid, neuromuscular paralysis in the United States. Many infections have been linked with GBS. The most common are gastrointestinal or respiratory illnesses. Up to 70% of patients have reported an antecedent illness in the 1 to 6 weeks before the presentation of GBS. rare, with an incidence of 0.4 to 2 per 100,000. Miller Fisher syndrome (MFS) is a considered a variant of GBS, characterized by ophthalmoplegia, ataxia, and areflexia.

Presentation: Guillain-Barre syndrome (GBS) patients describe a fulminant course of symptoms that usually include ascending weakness and non-length dependent sensory symptoms. Required features - progressive weakness of legs and arms, trunk, bulbar and facial muscles, and external ophthalmoplegia; areflexia or decreased reflexes in weak limbs. By definition, the nadir is usually reached within 4 weeks. Symmetric involvement is a key feature of GBS. GBS is usually considered monophasic; therefore, a relapsing or remitting course at presentation would be considered atypical. Additionally, a prior GBS event (recurrent GBS) is also unusual, occurring in < 10% of all patients.

Diagnostics Evaluation: Guillain-Barre syndrome (GBS) is considered a clinical diagnosis; therefore, a diagnosis can be made with confidence at the bedside in most cases. History of physical, CSF analysis, electrodiagnostic studies, MRI spine MRI brain. Nerve conduction studies abnormal (normal in Miller Fisher syndrome).

Therapy Plan: intravenous immunoglobulin (IVIG) or plasma exchange.

Prognosis: Overall, most patients with GBS do well, with up to 85% of patients achieving independent ambulation with recovery around 6 months; however, there is a significant proportion of patients (20%) with morbidity. Rarely recurrent.

Differential Diagnosis: Tick paralysis.

Regular Expression: weakness

Phenotype Development:

Logic Description: First occurrence of Guillain-Barre in an inpatient setting expected to last atleast 180 days.

Literature review—several literature sources cited use of codes mapping to acute infective polyneuritis as complete capture of ICD9 codes

Source of errors in real world and impact on algorithm:

Miss rate/False negative rate - we hypothesize that it is less likely that this condition would be missed when present because of its characteristic presentation and relatively rapid onset.
Index date misclassification: we hypothesize that because presents itself relatively rapidly and it is not an indolent disease - and so the probability of index date misclassification is lower.
Specificity - since GB is the most common cause of acute, flaccid, neuromuscular paralysis it is unlikely to be mistaken for rare forms of acute flaccid paralysis. However, it is possible that it may mistaken for the UMN/stroke related plegia.

Cohort Submission:

This cohort definition has cohort id # 235 in OHDSI Phenotype library (pending peer review).

Phenotype Evaluation:

Insights from Cohort Diagnostics:

since this cohort definition requires inpatient or inpatient/er, data sources that do not have good capture of such visits are not good candidates (sensitivity error).
Incidence rate: Overall annual Incidence rate: ~0.1 per 1,000 person-years, across age/sex/years is probably higher than expected. We observe that incidence increases with age decile - however we did not find that reported anywhere else. The rates also appeared to be higher in males.
Index event breakdown: Guillain-Barré syndrome and Acute infective polyneuritis are the two most common concepts at inex. We observe Acute infective polyneuritis mostly prior to 2015 when ICD9CM vocabulary was used. In ICD10CM G61.0 concept is considered equivalent to ICD9CM 357.0. The inclusion of Fisher’s syndrome codes represents <5% patients in Optum Panther.
Characterization:
On Day 0: The top strengtheners in condition domain were malaise, asthenia, muscle weakness. Interestingly we also observed electrolyte imbalance. Corticosteroids utilization was high up to 80% on index date. Immunoglobulins utilization was high in JMDC and Optum EHR (80%) suggesting high specificity. Immunoglobulin utilization was not observe in other US data sources (conceptId 38003174 was present in > 30% in optum dod). There appears to be an increased utilization of gabapentin in the period immediately after diagnosis - the significance of this is unknown.
On Day -30 to -1: we observe the diagnosis of GBS in about 10% of persons. This suggest that some person were being evaluated in outpatient basis prior to inpatient stay. This is evidence of index date misclassification. Top concept in this period are GB syndrome, polyneuropathy, acute bronchitits * upper respiratory infections (common among persons with GBS syndrome, asthenia, malaise, paresethsia, hypokalemia

source of errors:

Sensitivity errors: we require persons to have inpatient or ER visit on index. We do not know if we are missing persons who truly have guillian barre syndrome, but are not being seen in the visit classified by Inpatient or Inpatient & ER. In general we would not expect sensitivity error as this being a relatively serious condition, patients are more likely to seek care at higher acuity setting. We have not tested this assumption.
Index date misclassification: this definition has index date misclassification because we are indexing on inpatient stay. We observe evidence that persons are being managed in non inpatient setting prior to being in an inpatient setting. So- a better cohort definition may be requiring inpatient stay at a future date.

Overall: this cohort definition appear to identify persons with the phenotype of interest. Should not be used in data sources without good capture of inpatient visit. A future work to correct index date misclassification may be needed.

Performance characteristics

Pending - PheValuator

Evan_Minty · December 2, 2022, 10:54pm

Review initiated

Evan_Minty · January 24, 2023, 1:02am

Thanks Gowtham. I learned a lot from your submission.

To review this, I reviewed 4 databases in cohort diagnostics output.
I also reviewed 20 random cases (in sample) in CCAE using cohort explorer.

I typically reviewed the timelines briefly, then used the following regular expression to filter data in those timelines

(In practice, I generally combined that to get all such features at once).
I tracked the following attributes in a sheet:

Presence of risk factors, Presence of symptoms, Dx count >= 3, EMG or NCS, Alternative Dx, Treatment, Prognosis.

I also tracked my level of confidence (high, moderate, low); adjudication was made on the basis of the preponderance of evidence in a case.

From that review, I found the following:

GB:	high	5
	mod	2
	low	4
TOTAL		11
not GB:	high	6
	mod	2
	low	1
TOTAL		9

PPV	0.55

A few comments:

Incidence rate: Overall annual Incidence rate: ~0.1 per 1,000 person-years, across age/sex/years is probably higher than expected.

Agreed - it looks like about 10x what I see reported (e.g. nice article showing global incidence rates in the Lancet.
Based on my review, I suspect some increase is driven picking up prevalent disease in those that have been chronically disabled, and some indiscriminent use of the Acute infective polyneuritis (129131007) concept, (although that inference comes from CCAE). The denominators are very different however (as of time of writing I couldn’t get the incidence rates to compute on data.ohdsi.org’s version of this cohort).

Index event breakdown: Guillain-Barré syndrome and Acute infective polyneuritis are the two most common concepts at index. We observe Acute infective polyneuritis mostly prior to 2015 when ICD9CM vocabulary was used.

Of relevance here is that my profile review corresponds to the DB with the highest useage (>60%) of the AIP code. False positives seemed to be loaded on that code, so it may not be getting applied correctly in the coding process.

Corticosteroids utilization was high up to 80% on index date. Immunoglobulins utilization was high in JMDC and Optum EHR (80%)

Speaks to your later point about the datasource. I’d agree the capture of inpatient stays and the pharmacy record contributes to the trust I’d invest in this phenotype.

There appears to be an increased utilization of gabapentin in the period immediately after diagnosis - the significance of this is unknown.

Neuropathic pain is common both at presentation, and during the recovery course, Gabapentin is likely being used in that context. So I would consider this a strengthener

RE: your consideration of error sources:

Sensitivity Errors … …We do not know if we are missing persons who truly have guillian barre syndrome, but are not being seen in the visit classified by Inpatient or Inpatient & ER. In general we would not expect sensitivity error as this being a relatively serious condition

I agree that for incident disease this would require a hospitalization. Feared complications include respiratory failure (in up to 20%; less in CohortDx but not an easy feature to capture in observational data, unfortunately).

Index Date misclassification: … …We observe evidence that persons are being managed in non inpatient setting prior to being in an inpatient setting. So- a better cohort definition may be requiring inpatient stay at a future date.

Because of the comments above, I do not think the outpatient management reflects incident disease, rather prevalent disease which enters the record with a pre-established diagnosis. This was seen to generate false positives at an admission level as (I suspect) a chronic issue that is recapitulated at time of admission for other reasons.

Instead, I would suggest that:

If looking for prevalent disease, remove the inpatient stay requirement, and increase the condition count.
If looking for incident disease, add a 365 day clean window prior to index, keep the hospitalization requirement. (h/t to @Patrick_Ryan has noted in last year’s phentoype phebruary that this can be a fairly modular recommendation; speaks to @Christian_Reich 's aspiration with regards to cohort logic building blocks).

Recommendation

I’ve been thinking more about our peer review process. I wonder if we aspire to transparently assemble ‘what’s been done’ with regards to examining a phenotype, and a structured assessment as to where we see consistency or deviations from the clinical idea.

e.g. for the ‘what’s been done’

	Clinical Description	Literature Review	PHOEBE / orphan concepts	CohortDx	PheVal	Profile Review (in sample)	Discordant case Review
Submission (GR)	Completed (GR)	Completed (GR)	Reviewed	Reviewed in 10 DB	pending
Review (EM)	Reviewed (EM)			Reviewed in 4 DB		20 pts in CCAE (PPV 55%)	Pending

This is not to say any all items here are essential in every phenotype submission, but they each might contribute to trust.

I agree with @Gowhtam’s recommendation of caution in data sources with incomplete records of hospitalizations. I’d like to (eventually) review disjoint cases given that I found that ‘AIP’ seemed to get applied to cases where the evidence in the timeline eventually pointed to other causes (e.g. spinal stenosis, myositis, others).

We perform assessments across clinical domains in our verbal description of the results, (e.g. epidemiology, signs / symptoms, diagnostic assessment, therapy, prognosis), but could perhaps that would also be be amenable to a structured format.

Based on the CohortDx results, I’d trust it’s application in Optum and JDMC; I’d consider modification based on observations in CCAE (especially if I was interested in incident disease).

Patrick_Ryan · January 24, 2023, 1:45pm

Thanks for continuing this discussion @Evan_Minty .

Just one note about ‘AIP’ vs. ‘GBS’ codes. I think the context of the source vocabulary matters here. In a US claims database, pre-2015, conditions were coded in ICD9CM, which only had the code '357.0, with the label of ‘Acute infective polyneuritis’". there was no ICD9CM code for the label of ‘Guillian Barre Syndrome’, though GBS was an acknowledged synonym for AIP in the ICD9 codebook. Prior literature with validation studies of GBS in US data from the ICD9CM era found lower PPV and suggested to impose additional requirements, such as inpatient visit, primary position, and neurology visit to increase specificity (example article here). I’m NOT advocating for these suggestions, just pointing out the challenges in the ICD9 times, which I don’t think are attributed purely to the label of the ICD9CM code. When US claims data transitioned to ICD10CM post-2015, there was a code introduced with the label ‘Guillain Barre Syndrome’, ‘G61.0’. I don’t believe we should be overly reliant on the string term of the code label to do our review; in this case, it could be more of a reflection of the difference in coding practices for all conditions pre-2015 vs. post-2015, not just about credibility of ‘AIP’ label vs ‘GBS’ label.

Evan_Minty · January 25, 2023, 6:22am

Thanks @Patrick_Ryan.

RE: your point about coding systems, thanks for pointing out that paper. @Gowtham_Rao noted the ICD9 to 10 transition as well.

A priori, I would not have thought the AIP code would have been as poorly predictive as that article found. For single occurrence of the ICDC9 AIP code, PPV ~30% ; interestingly, in implementing a definition requiring hospitalization, they got 50%(any position) - 55% (primary position), similar to what we see here with a definition requiring hospitalization and a (rough) estimate of PPV.

It may well be that this has improved with the shift to ICD10 GBS (and whatever else may have changed in the coding recommendations; although I can’t help but wonder that semantics in the string itself led to confusion and misapplication).

Those coding shifts may be a reasonable component to consider in our assessments. They may not be easily handled in many cases, where there’s a common standard concept (without going to a source code definition, which I know we try to avoid).

In this case, there are separate standard concepts that essentially map to ICD9 (AIP) and ICD10 (GBS) . So we could implement a definition that requires more evidence (e.g. ≥ 2 or 3 condition count ) in AIP; and use cohort diagnostics to confirm the populations of the 2 branches in our definition are similar. At the very least would be heartening to know that GBS is performing better than AIP across databases.