Diagnosis of Thyroid Disease in Hospitalized Patients:  A Systematic Review FREE

HAVING A sensitivity and specificity of more than 99%,¹ the sensitive thyrotropin (thyroid-stimulating hormone) (sTSH) assay currently is the best initial test for the diagnosis of thyroid disease in outpatients.² Nonthyroidal illness (NTI) has a negative impact on both the sensitivity and specificity of sTSH. Recommendations on the use of the sTSH assay in patients with NTI range from limiting use of the test to patients with high clinical suspicion^3- 6 to generalized screening.^7- 9

Determining the optimal strategy for the use of any diagnostic test requires elucidating 3 variables: (1) the underlying risk of the target disease in the population of interest, (2) clinical variables that may raise or lower this risk, and (3) characteristics of the test in the population of interest. The first 2 variables lead to a pretest likelihood of disease and the third leads to a posttest likelihood of disease.

We applied this 3-fold framework to the diagnosis of thyroid disease in acutely ill, hospitalized patients with NTI. We conducted a systematic review that addressed the following 3 questions: (1) What is the prevalence of unrecognized thyroid disease? (2) What is the usefulness of clinical signs and symptoms in predicting thyroid disease? and (3) What is the sensitivity and specificity of the sTSH assay in the diagnosis of thyroid disease?

In acute illness, all thyroid indices have been shown to yield false-positive results.¹⁰ Because of this, we have decided that the gold standard for diagnosing thyroid disease in this population is follow-up after resolution rather than the actual thyroid index chosen in the particular study. We recognize that the new third-generation sTSH assay has greatly increased the sensitivity of diagnosis, mostly of hyperthyroidism, in an outpatient setting.

Throughout, we sought studies reported in the English language that enrolled more than 50 patients. To address the prevalence of unrecognized thyroid disease in acutely ill, hospitalized patients, we used the following MEDLINE search strategy: (explode) thyroid diseases and prevalence (text word) and 1 of the following: hospitalized or medical or inpatient (all text words) or atrial fibrillation or (explode) mental disorder or (explode) dementia (medical subject heading [MeSH] terms) or geriatrics (both MeSH and text word). We retrieved and included in the analysis all articles in which the study population was acutely ill or hospitalized with an NTI and those in which the reference standard included biochemical markers (with or without clinical features) and follow-up after resolution of the NTI. For this section, therefore, our main reference standard was long-term follow-up, and we did not limit ourselves to the sTSH assay as a diagnostic test.

To address the usefulness of clinical signs and symptoms in predicting thyroid disease in hospitalized patients, we used the following search strategy: (explode) thyroid diseases and (explode) cohort studies and either signs or symptoms (text words). We retrieved the articles in which the authors described clinical signs and symptoms in sufficient detail to calculate likelihood ratios (LRs). Because we found no articles addressing the usefulness of clinical signs and symptoms in the diagnosis of thyroid disease among inpatients, we included studies of outpatients only in this section.

To address the sensitivity and specificity of the sTSH assay in the diagnosis of thyroid disease in hospitalized patients, we combined thyrotropin (MeSH) and 1 of the following: hospitalized or medical or inpatient (text words) and either sensitivity or specificity (text words). We retrieved articles in which the study population was acutely ill or hospitalized with an NTI and the reference standard included second- or third-generation sTSH assays (ie, functional lower limit of 0.1 and 0.01, respectively) and follow-up after resolution of the NTI. We also searched the bibliographies of relevant articles and review articles for additional studies.

Two of us (J.A. and P.M.) carried out the searches and data extraction independently. We resolved discrepancies by consensus. In abstracting the data, we excluded postpartum patients and those with previously diagnosed thyroid disease, as well as subclinical hypothyroidism or hyperthyroidism. Subclinical thyroid disease was defined as abnormalities in 1 thyroid index (ie, sTSH), while other indices (ie, free thyroxine) were normal in a patient without overt symptoms of thyroid disease.

We graded the quality of the studies selected using the following 3 criteria: (1) follow-up of normal results included vs follow-up only of abnormal results; (2) greater than 60% of abnormal results followed up after resolution of NTI vs less than 60%; and (3) criteria for diagnosis of thyroid disease clearly and explicitly stated vs implicit or clinical diagnosis.

We pooled the results of individual studies using a weighted average (according to sample size) only if a formal test of heterogeneity using a χ² test was not significant at P<.05.

The MEDLINE search identified 110 titles. Thirty-one articles addressed the inpatient population, 16 of which had the required follow-up after resolution of the NTI, allowing us to calculate the prevalence of unrecognized thyroid disease. We grouped these 16 articles according to whether they studied general medical inpatients, geriatric inpatients, inpatients with atrial fibrillation, or psychiatric inpatients.

Two studies addressing the prevalence of thyroid disease in general medical inpatients met our inclusion criteria (Table 1). The first study¹¹ is the most methodologically sound of all those included in this review. The authors screened 1580 general medical inpatients on admission with a sTSH level and found 17.2% to be outside the normal range. The investigators followed up 63% of the patients with abnormal results until resolution of the acute illness and found that, on retesting, 85% were euthyroid by criteria that combined thyroid indices and clinical examination. False-positive results were attributed to acute illness and drug interactions, in particular glucocorticoids. Investigators also followed up an arbitrary sample of 48% of the initially euthyroid group and found no new cases of thyroid disease.

The second study¹² tested 363 female medical inpatients on admission using free thyroxine and found 9.3% to be outside the euthyroid range. The investigators followed up 76% of these patients, 92% of whom were euthyroid or had previously diagnosed thyroid disease. Combining the 2 studies, we found a prevalence of thyroid disease of approximately 1.5%.

Six studies addressing the prevalence of unrecognized thyroid disease in the geriatric population met the inclusion criteria (Table 2). Three studies are prospective,^3- 4,13 and 3 are retrospective.^14- 16 Despite describing varying populations (eg, those drawn from a psychogeriatric unit¹⁴ and those referred for assessment of dementia¹⁶), all studies indicated a high rate of false-positive initial thyroid indices and an overall estimate of true thyroid disease of approximately 2%.

Two studies describing the prevalence of thyroid disease in patients with atrial fibrillation met our criteria (Table 3). The first, a longitudinal cohort study,¹⁷ included both inpatients and outpatients. The diagnosis of thyroid disease was left to routine follow-up by the physician. The second study¹⁸ documented a much higher rate of hyperthyroidism and did not examine hypothyroidism. The authors used a shorter follow-up (6 weeks) and an older population and relied on TSH response to thyrotropin-releasing hormone. We believed that their population, methods, and results were too disparate to combine, and this was confirmed by a significant χ² test for heterogeneity.

Six studies^19- 24 addressing the prevalence of thyroid disease in admissions for acute psychiatric illness met our criteria (Table 4). In total, 1479 patients were studied, 181 (12.2%) of whom had initial abnormal thyroid indices. With a follow-up rate of about 90%, approximately 97% of these patients were subsequently found to be euthyroid.

In summary, the overall incidence of thyroid disease among inpatients was between 1.0% and 2.5%, with psychiatric inpatients at the lower end of this range and geriatric inpatients at the upper end. The highest incidence of hyperthyroidism was 0.6%, while hypothyroidism was up to 3 times more common, ie, approximately 1.8%.

No studies addressed the question of the usefulness of clinical signs and symptoms of thyroid disease in inpatients. We reviewed several studies that addressed this issue in outpatients, keeping the remaining inclusion criteria fixed. With the relaxed criteria, our MEDLINE search identified 147 titles, 8 of which met our inclusion criteria; 6 of these addressed individual signs and symptoms, and 2 addressed the total number of signs and symptoms in the diagnosis of thyroid disease.

Table 5 presents the 6 studies that provide information on individual signs and symptoms. The first study²⁵ is the most comprehensive, listing 38 signs and symptoms of thyroid disease in 99 euthyroid, 131 hypothyroid, and 235 hyperthyroid patients from 6 specialty centers in the United States and Australia. We calculated LRs from these data and included clinical features with the highest predictive power. These LRs are similar to those calculated from results in 2 other studies,^26- 27 but they are probably inflated due to the fact that the clinicians were not blinded to the thyroid function test results before performing their physical examination.

Other studies found that clinical features were disappointingly poor in predicting thyroid disease. In a retrospective review of 982 charts, Schectman et al²⁸ found a poor correlation between clinical features and thyroid disease. These investigators depended on the primary care physicians' records of the initial history and physical examination, and it was unclear whether specific signs and symptoms of thyroid disease were sought and not found or were not sought at all; the latter would underestimate test sensitivity. Two other studies^29- 30 also gave disappointing results, perhaps because they depended on patients' self-report in the form of a questionnaire.

Another approach is to examine the total number of signs and symptoms of thyroid disease, rather than the individual signs and symptoms themselves. Two studies^31- 32 allow us to calculate the relationship between the number of signs and symptoms and the probability of thyroid disease (Table 6). In a review³¹ of 135 family practice charts, there was a good correlation between the number of clinical features and the presence of thyroid disease. This was confirmed in a second study³² in a retrospective chart review of 500 patients seen in a thyroid clinic. In both studies, the absence of signs and symptoms significantly decreased the chance of thyroid disease.

In a study of the diagnostic properties of a clinical examination for thyroid disease, we would ideally want a prospective, blinded trial that included physician-generated information (as opposed to a self-administered questionnaire) that would best replicate the usual diagnostic process. Unfortunately, none of these studies combined all such elements.

Of 30 titles in the initial search, 7 studies and 1 abstract focused on inpatients. Of these, only 2^11,33 described the results in sufficient detail to allow us to calculate the LRs of sTSH in hospitalized patients (Table 7). The first study by Spencer et al¹¹ used a second-generation TSH assay with a functional lower limit of 0.1. Table 7 shows the positive LRs associated with various ranges. In a follow-up study,³³ some of the same serum samples plus additional inpatient serum samples were reassayed using a third-generation TSH test with a functional lower limit of 0.01. Of patients who truly had hyperthyroidism (as judged by follow-up) and an original TSH level of less than 0.1 µIU/mL, 22 of 23 had a TSH level of less than 0.005 µIU/mL using the third-generation assay. Of those who had NTI and an original TSH level of less than 0.1 µIU/mL, only 5 of 37 had a TSH level of less than 0.005 µIU/mL, indicating that the third-generation assay had additional power in detecting inpatients who were truly hyperthyroid. The LRs much greater than 1 are associated only with the extremes of the TSH range. Normal and mildly abnormal values are very useful in ruling out disease.

We applied the available evidence to the problem of diagnosing thyroid disease in hospitalized patients using a 3-part diagnostic strategy.

We estimated the pretest probability of thyroid disease from the prevalence of unrecognized thyroid disease among inpatients, and found a prevalence of 1.0% to 2.5% (Tables 1-4). These data have a number of limitations, including varying reference standards in the diagnosis of thyroid disease, varying thyroid assays, and inconsistent follow-up that focused mainly on selected patients with abnormal initial biochemical test results. The degree of NTI of the patients varied between studies and was often not explicitly stated. The diagnosis of thyroid disease had to be inferred from the data presented.

Despite these caveats, estimates of thyroid disease were similar, and formal statistical testing did not show heterogeneity of estimates across studies. In addition, these estimates of prevalence among inpatients agree with those of recent reviews¹⁰ but are lower than those of population studies.^34- 36 This discrepancy is probably due to the more stringent criterion of thyroid disease in our search, ie, clinical and biochemical follow-up, as well as the inclusion of subclinical and biochemical thyroid disease in the population studies. We chose to include older studies, including those that relied on free thyroxine and early-generation TSH assays. As mentioned, we believe that follow up is the appropriate gold standard in diagnosis of thyroid disease in hospitalized patients after resolution of the acute illness.

Clinicians modify pretest probability by noting the clinical signs and symptoms. Because there were no studies of the impact of clinical signs and symptoms on probabilities in inpatients, our options were to either neglect this area altogether or broaden our focus to include outpatients. Because we believe the biological relationship between underlying thyroid disease and signs and symptoms is similar in inpatients and outpatients, we chose the latter course. However, the specificity of symptoms and signs may decrease among inpatients, and the values of LRs we found should be treated as upper-limit values. Some findings, such as goiter, may retain their specificity in the hospitalized population.

Another reason the LRs in Tables 5 and 6 may be spuriously high is that examination of a large number of signs and symptoms in a relatively small population increases the likelihood of chance associations. Even more important, the lack of blinding to test results may cause a clinician to overinterpret physical signs that he or she expects to see. Studies that depend on physical signs and symptoms elicited before diagnostic test results found much lower LRs, in the range of 2 to 3. These studies do not, however, clarify whether clinical signs and symptoms were specifically sought.

The total number of signs and symptoms appears to be a more robust indicator of thyroid disease, with LRs between 7 and 19, although these estimates may also be inflated by the lack of blinding. Some descriptive studies suggest that elderly patients with known thyroid disease have fewer signs and symptoms than younger patients,^37- 38 and one might expect this effect in hospitalized patients as well.

Acute illness influences thyroid indices such that while the sensitivity of the test is maintained (a normal value carries an LR of 0), the specificity of the sTSH assay is reduced. The majority of abnormal sTSH test results obtained during acute illness return to normal once the illness has resolved. The LR for very abnormal results (<0.1 or >20) is in the range of 7 to 11. In contrast, the LR of an abnormal TSH test result in outpatients is close to 100, emphasizing its superior test characteristics in that population.

Clinical examples help illustrate the application of this information. Suppose a general medical patient (prevalence of hyperthyroidism, 0.6%) has no clinical features of thyroid disease (LR in the range of 0.2) and thus has a pretest likelihood of 0.12%. A second-generation sTSH assay yields a value of less than 0.1 (LR = 7.7). The resulting posttest likelihood of hyperthyroidism is now about 1%, sufficiently low that specific action is likely unwarranted.

Further conider a patient admitted with congestive heart failure who is noted to have a goiter. Using the prevalence of hypothyroidism calculated for general medical inpatients, ie, 0.6%, and modifying this by the LR of 5 associated with a goiter, we arrive at an estimated probability of hypothyroidism of 5%. If the TSH test result is greater than 20 (LR = 11.1), posttest probability is around 40%.

These examples stress the limitations of the sTSH test in the face of very low prior probability and suggest the following approach. First, in hospitalized patients in whom pretest probability is very low, thyroid testing should be deferred until patients have recovered from their acute illness.^5- 6 Second, among inpatients with numerous signs and symptoms of thyroid disease (data would support a value of at least 5, perhaps less in elderly patients) or inpatients with signs particular to thyroid disease, eg, goiter and eye signs, sTSH testing is appropriate. Third, clinicians should use the pretest-posttest framework to interpret sTSH test results.

We reviewed the prevalence of undiagnosed thyroid disease among inpatients and the LRs associated with various signs and symptoms of thyroid disease and clarified characteristics of the sTSH test in hospitalized patients. These data should help clinicians better interpret the physical examination and laboratory results that lead to the diagnosis of thyroid disease in patients whose clinical picture is confounded by acute illness.

Out thanks to John Booth, MD, FRCPC, for comments on an earlier version of the manuscript, and to Ms Deborah Maddock for assistance with manuscript preparation.

Accepted for publication July 14, 1998.

Reprints: Gordon Guyatt, MD, MSc, FRCP, Clinical Epidemiology & Biostatistics, McMaster University Health Sciences Centre, 1200 Main St W, Room 2C12, Hamilton, Ontario, Canada L8N 3Z5 (e-mail: guyatt@fhs.csu.mcmaster.ca).