Missed diagnosis of stroke in the emergency department: a cross-sectional analysis of a large population-based sample

Study design and sample

Outcomes

Analysis

We examined adult (≥18 years old at the admission date) stroke admissions via the ED in calendar year 2009. We analyzed patients <18 years separately. We assessed the 20 most common diagnostic groups from ED visits with potential misdiagnoses (see Supplemental Data, Appendix 1 for a detailed description of diagnosis groupings). To test our hypothesis regarding the association between symptoms and missed stroke, we performed a visit-level analysis, calculating the frequency and distribution of diagnoses (observed) relative to a reference ED population (expected). To examine the association of a probable misdiagnosis with patient and hospital characteristics, we performed a person-level analysis, allowing only a single admission and single ED visit per patient. For person-level multivariate analyses, we used generalized estimation equation (GEE) models [27] to analyze the odds of a probable stroke misdiagnosis among adult patients admitted for stroke (with or without prior treat-and-release ED visits), controlling for patient and facility characteristics.

We conducted several validation checks to ensure coherence between our misdiagnosis construct and the results we report. First, we compared the distribution of ED diagnoses in those with potential misdiagnoses to a general ED sample of treat-and-release visits, drawn from the same national sample as our cases and controls (see Supplemental Data, Appendix 1 for details). We anticipated that the relative distribution of diagnoses would be skewed toward dizziness and headaches (known to be associated with stroke) and away from back or abdominal pain (known to not be associated with stroke). Second, we assessed the temporal profile of probable missed diagnoses within the 30-day window prior to index stroke admission. We hypothesized readmissions would be disproportionate in the first few days after the initial ED treat-and-release and diminish over time, as shown previously for major stroke after TIA and minor stroke [16]. Third, we analyzed the association between stroke type (TIA, ischemic stroke, intracerebral hemorrhage, subarachnoid hemorrhage) and initial presenting symptom. We hypothesized that dizziness/vestibular diagnoses would be associated with ischemic events (TIA or stroke) and headache/migraine diagnoses would be linked to hemorrhages (often misdiagnosed as benign headaches) and TIAs (often misdiagnosed as migraine-related transient neurologic symptoms).

Frequencies and distributions

Figure 1 shows data flow. Of 198,819 stroke admissions in 2009, 187,188 (94%) had complete and administratively correct records for analysis. Of these, 28,248 (15.1%) were potentially missed or ‘bouncebacks’ and 23,809 (12.7%) were potentially missed strokes (i.e., discharged from the initial ED visit with a non-cerebrovascular diagnosis).

Figure 1

Flow diagram of study population derivation.

The 20 most common reasons for the initial ED visit among those potentially missed or ‘bouncebacks’ are shown in Table 1. Most potential misdiagnoses involved none of the targeted ‘case’ or ‘control’ diagnoses, but these ‘other’ diagnoses were, in aggregate, under-represented relative to an average ED treat-and-release population (Table 2). Patients with an initial cerebrovascular diagnosis were the most likely to return for a stroke admission (Table 2). As hypothesized, those released with dizziness or headache (probable misdiagnoses) were over-represented, while back pain or abdominal pain diagnoses (control conditions) were under-represented (Table 2; Supplemental Data, Appendix 2 contains data for patients <18 years). Among potential misdiagnoses, 10.4% involved a dizziness or headache diagnosis (i.e., probable misdiagnoses), representing 1.2% of all stroke admissions. Dizziness was associated predominantly with missed ischemic strokes, whereas headache was associated with both ischemic and hemorrhagic forms of stroke (Table 3). Temporal profile analysis (Figure 2; Supplemental Data, Appendix 3) indicated that treat-and-release ED visits for probable misdiagnoses were clustered in the few days prior to the hospital admission for stroke, compatible with the known biology of TIA and minor stroke preceding major, disabling stroke [16]. This is in contrast to controls with back problems or abdominal pain whose visits were evenly distributed throughout the 30-day window (Figure 2). Note also that among other diagnoses (non-stroke, non-dizziness/headache, non-back/abdominal pain), a similar clustering was observed (Supplemental Data, Appendix 3), arguing that some of these potential misdiagnoses, which outnumbered the dizziness/headache diagnoses, were also likely true misdiagnoses.

Figure 2

Temporal profile analysis of initial ED treat-and-release visits for probable misdiagnoses (dizziness/headaches) versus controls (back problems/abdominal pain) in the 30 days prior to an index stroke admission.

ED, emergency department.

Multivariate analysis

Multivariate analysis of patient and hospital characteristics associated with probable missed stroke diagnoses is shown in Table 4.

Overall missed stroke estimates

To our knowledge, this is the first large-scale, multi-regional study to examine clinical, hospital, and visit-related predictors of missed diagnosis of stroke in US EDs. Using 30-day readmission for stroke, we found that 12.7% of stroke admissions were potentially misdiagnosed. The top two potential misdiagnoses preceding a stroke readmission were headache and dizziness. This finding is consonant with previous studies indicating these particular symptoms are often associated with missed strokes [10, 14, 18, 20, 23, 28, 29]. These probable misdiagnoses accounted for 1.2% of all strokes admissions and 10.4% of potential misdiagnoses. Thus, we estimate that the overall prevalence of misdiagnosis followed by delayed hospital admission after ED discharge ranges between 1.2% and 12.7% of all stroke admissions. With 1.3 million new or recurrent cerebrovascular events per year in the US [6], this corresponds to approximately 15,000 to 165,000 missed strokes and TIAs annually. Most missed strokes were ischemic rather than hemorrhagic (ratio 5:1), roughly in proportion to their population prevalence [6]. The impact and preventability of these misdiagnoses remains unknown.

Previous estimates of ED stroke misdiagnosis rates range from as low as 1.2% (considering admitted patients at an academic teaching hospital but not those discharged from the ED [30]) to as high as 50% (for predominantly Hispanic patients presenting with isolated dizziness or vertigo to non-teaching hospital EDs in Nueces County, Texas [18]). Our findings accord with the results of a systematic review (9.2% missed among more than 5000 patients with stroke from seven studies [10]) and the only population-based study of all-stroke misdiagnosis identified in that review (8.5% missed among 1800 patients with stroke in rural Texas [31]).

The current study’s 1.2% estimate for probable misdiagnosis is conservative. It does not count those who suffered no major adverse event (i.e., readmission); those with an erroneous diagnosis other than dizziness or headache (e.g., “gastroenteritis” diagnosed at the treat-and-release visit, with cerebellar infarction identified during the subsequent visit and hospitalization [11]); those with stroke at the initial ED visit admitted to a non-stroke clinical service with the wrong tentative diagnosis; those whose causally-related stroke occurred more than 30 days after an initial ED visit (e.g., carotid artery dissection presenting with headache, followed by a stroke at day 31 [32]); or those who suffered fatal stroke after ED treat-and-release. Although some initial ED visits we counted as probable misdiagnoses might have been coincidental, the fact that most were clustered in the few days before the stroke admission suggests otherwise. This temporal pattern with the highest risk in the first 48 h after discharge matches the time window of greatest risk for major stroke after TIA and minor stroke, where about three-fourths of the 30-day risk occurs in the first week [16]. It is further corroborated by the fact that this pattern was not observed for control subjects (Figure 2).

Symptom characteristics

As we hypothesized, dizziness and headache visits were over-represented while back and abdominal pain visits were under-represented among subsequent stroke admissions. The odds of a dizziness discharge preceding a stroke admission were 2.4-times expected, and headache discharges were 1.9-times expected. This is not surprising, given prior smaller-scale studies suggesting that presentations with isolated dizziness and headaches caused by stroke are more often misdiagnosed. For example, population-based studies in Texas found ischemic stroke misdiagnosis rates of 4% for patients with motor symptoms, 35% for those with dizziness, and 50% for those with isolated dizziness (i.e., with no other neurological symptoms) [18, 31]. Aneurysmal subarachnoid hemorrhages are more likely to be misdiagnosed when they present with isolated headache (19%) versus with accompanying mental status changes (5%) [13]. Strokes presenting with only “non-traditional” symptoms are misdiagnosed frequently compared to those with “traditional” symptoms (64% vs. 4%, OR 43.4; 95% CI 15.0–125.4) [19]. Considered with our findings, these results suggest that clinical presentations with “atypical” or “mild” stroke symptoms likely have relevance at a national level and should be targeted for misdiagnosis-reduction interventions.

Patient characteristics

Age was strongly (and inversely) associated with misdiagnosis. Those older than 75 years were 80% less likely to be misdiagnosed than those 18–44, corroborating data from small studies showing mostly young patients among those misdiagnosed [11, 20]. It is probably not surprising that stroke, generally considered a disease of the elderly, would be misdiagnosed in the young. Although missing strokes in younger patients may result from appropriate diagnostic reasoning that considers age-specific prevalence, clinical decision rules to detect stroke in dizziness [33] and headache [34] are not age dependent and may enable efficient bedside stroke detection using history and physical examination, even in patients with low baseline prevalence.

Women were at greater odds of misdiagnosis, and a greater frequency of non-classic stroke presentations could be to blame [35], as with acute coronary syndromes [36]. Women are also more likely to experience dizziness and headaches of benign cause [37–39], increasing the challenge of differentiating strokes from stroke mimics. Hispanic and non-White patients were more likely to be misdiagnosed, in keeping with prior studies of stroke [31] and myocardial infarction [40]. The causes of this racial disparity are not clear, but some evidence suggests that diagnostic stroke workups are less thorough for Hispanics [41]. Greater vigilance for stroke in women, minorities, and younger patients appears warranted.

Comorbid conditions may increase the suspicion of stroke or be associated with greater overall likelihood of admission, so it seems reasonable that lesser comorbidity was associated with greater likelihood of missed stroke. Private insurance was a risk factor for misdiagnosis relative to both Medicare and Medicaid but income was not, perhaps suggesting that private insurance was a partial surrogate for lower medical comorbidity or complexity, or that private insurance reduced the likelihood of initial admission in a nonspecific way.

Hospital characteristics, workflow, and ED visit characteristics

Hospitals with high annual ED volume and teaching hospitals had lower odds of misdiagnosis. Whether these differences reflect greater diagnostic acumen, a lower threshold for additional diagnostic testing, or a general tendency toward admission driven by profit motive or legal conservatism remains uncertain. There was no geographic variation by hospital region but, somewhat surprisingly, hospitals located in large metropolitan areas were more likely to miss stroke than those in small metropolitan areas. The source of this variation is unclear, but it matches that found in a similarly conducted analysis of HCUP data for misdiagnosis of acute myocardial infarction [40], so probably merits further investigation.

One of the largest factors identified in this study was progressively greater odds of misdiagnosis on days with lower ED admission rates – more than six-fold higher on days in the lowest ED admission rate quintile. It seems sensible that discharging a greater fraction of ED patients would be associated with a greater likelihood of misdiagnosis. Although there are many plausible explanations, one possibility could be that different ED physicians have different risk thresholds for obtaining diagnostic workups and admitting patients [42], resulting in substantial day-to-day variations in admission rates that correlate with the likelihood of misdiagnosis.

Although ED crowding as measured here was not associated with missed stroke, leaving against medical advice was associated with a 2.9-fold increase in odds of misdiagnosis. Since leaving against medical advice is associated with ED crowding [43], the impact of crowding on misdiagnosis should be explored further in future studies. The finding of greater odds of misdiagnosis on weekdays runs counter to findings from prior studies and should be interpreted with caution. Further study of hospital and physician determinants of misdiagnosis is warranted.

Implications for diagnostic error research

The analytic methods described in this manuscript hold promise as a possible trigger tool [44] or metric for identifying and assessing the overall burden of dangerous diagnostic errors in the ED using administrative data. Although generic ED revisit analyses do not predict errors well [45], the specific symptom-disease pair approach has been used successfully to identify diagnostic errors based on ED revisit analyses [23]. Our novel approach of referencing observed (‘cases’) versus expected (‘control’) distributions of symptoms at ED visits offers two important advantages over prior methods. First, it serves as an important clinical validity check on the accuracy of results derived from billing data. Second, it offers insight into the relative risks of misdiagnosis across a range of symptoms, allowing prioritization across symptoms-disease pairs. From an epidemiologic perspective, our ‘look-back’ approach (from revisit/harm back to initial presentation) could be used to identify symptom-disease pairs that are prevalent and high-risk for misdiagnosis-related harm. These highest-impact symptom-disease pairs could then be monitored using ‘look-forward’ approaches (from initial presentation to revisit/harm) [28].

Implications for clinical care

Reducing ED stroke misdiagnosis may not be easy. Despite the rapid increase in use of advanced neuroimaging and other expensive laboratory diagnostic tests [46], there is little evidence that this “technocentric” approach correctly diagnoses more strokes [47, 48]. Routine access to neurologists, advocated by some [49], is probably impractical. Telestroke consultation could be cost-effective for accessing expertise remotely [50]. Unfortunately, tele-consults do not directly address atypical cases most likely to be missed [19], since consultation may not be sought. Computer-based diagnostic decision support may someday help, but no systems have yet demonstrated real-world benefit in stroke diagnosis [51]. Bedside methods for identifying high-risk patients [33, 34, 52–54] offer evidence-based [17], cost-effective [55] alternatives. Efforts to disseminate these history and physical examination-based approaches outside of specialty circles [56–58], possibly assisted by emerging device-based technologies [59], could be a promising strategy to reduce misdiagnosis. Implementations should be coupled to misdiagnosis metrics similar to those used here to objectively demonstrate error reduction.

Limitations

Limitations relate primarily to use of administrative data (without granular clinical details) and cross-sectional study design. All misdiagnoses reported here are inferred, since charts could not be directly reviewed; thus we could not control or adjust for potential confounding related to comorbidities. Administrative errors in coding may have influenced the results, and it is possible that some inpatient stroke diagnoses were incorrect. Misdiagnosed strokes not resulting in readmission were not captured. Some probable misdiagnoses may have been coincidental, and many of the ‘bounceback’ cases (stroke discharge followed by readmission) may have been unavoidable. Our cross-sectional results can only determine associations, not causation. Our categories of ‘probable’ and ‘possible’ misdiagnoses have not been previously validated with ‘gold standard’ cases where misdiagnosis status was known definitively. Unmeasured factors that could influence odds of misdiagnosis (e.g., physician training, experience, or risk aversion; ready availability of neuroimaging or neurologic consultation in the ED; stroke center certification [60]) might confound the association with some of our measured predictors (e.g., hospital type, ED admission fraction). Some known predictors of misdiagnosis (e.g., ED triage severity level [23]) could not be measured. Multivariate associations may partly reflect specific attributes of misdiagnosis related to headache and dizziness, so may not generalize completely to other clinical presentations of stroke. Harm and preventability could not be directly assessed; it is unknown how many of those misdiagnosed underwent appropriate neuroimaging. Although the sample is large (187,188 patients and 1016 hospitals), this work is based on data derived from only nine US states, does not consider non-ED (e.g., primary care) missed stroke, and might not be generalizable to other settings nationally or internationally.