J Endocrinol Invest DOI 10.1007/s40618-013-0031-z
ORIGINAL ARTICLE
Establish and verify TSH reference intervals using optimized statistical method by analyzing laboratory-stored data Y. Feng • W. Bian • C. Mu • Y. Xu F. Wang • W. Qiao • Y. Huang
•
Received: 30 July 2013 / Accepted: 16 November 2013 Ó Italian Society of Endocrinology (SIE) 2013
Abstract Objective To establish reference intervals using an optimized statistical method by collecting available laboratory data of thyroid stimulating hormone (TSH), and then to verify with the laboratory-present reference intervals. Methods TSH RIs of the total population and different races, genders, age, source of sample are established through improved Hoffmann and Katayev’s method with TSH test results data from Jan 2010 to April 2012 were collected, and finally conduct comparative verification with the laboratory present RIs. Results According to the improved method, we get various RIs of different sample populations. On comparing with the laboratory current RI (0.270–4.200 mIU/L) most reference change values (RCV) were within acceptable limits. Only lower limit of Han male, Uygur male and outpatient male populations outwith acceptable limits. On excluding the different values, finally, the new RI by the optimized statistical method is 0.233–4.979 mIU/L. Because the new RI expanded the current RI and was not different from the current RI, it was indicated that new RI could be used to verify the laboratory current RIs and seen as the current RI’s confidence interval (CI). Inference TSH RIs established by optimized Hoffmann’s and Katayev’s methods is viable and can be used to verify RIs provided by manufacturers or other laboratories. Keywords Reference intervals Thyroid stimulating hormone Verification
Y. Feng W. Bian C. Mu Y. Xu F. Wang W. Qiao Y. Huang (&) The Clinical Laboratory Center, The Tumor Hospital Affiliated to Xin Jiang Medical University, Urumqi, China e-mail: [email protected]
Introduction To establish suitable thyroid stimulating hormone (TSH) reference intervals (RI) which is regarded as an evidence for the diagnosis of sub-clinical hypothyroidism is important in clinical laboratory [1]. A special concern is that the TSH upper reference limit which defines subclinical hypothyroidism is in violent controversy [2–4]. So Finnish report redefined the TSH reference interval in the adult Finnish population [5]. Current TSH RI has also been criticized because it was established from unselected background populations which are provided by internationalized device or reagent manufacturers in China. Since all the reagents or devices are almost intended to be marketed in a nationwide setting, the reference intervals provided by the manufacturers are impossible to meet the demands of Chinese patients, including 1.3 billion people and 56 different ethnic groups. It is very necessary for establishing various TSH RIs for different ethnic groups to diagnose sub-clinical hypothyroidism accurately. How to define a RI in various laboratories? The American Clinical and Laboratory Standards Institute (CLSI) has provided authoritative standards [6], but it is difficult to meet those standards in practice. As required by CLSI, all clinical laboratories which adopt reference intervals provided by other laboratories or manufacturers shall verify the reference intervals. However, statistically, serious sampling bias could be inevitable when 20 samples are collected from local so-called healthy people and used to verify laboratory results obtained from large-scale tests. Therefore, Katayev and his colleagues published an article about establishment of reference intervals using available laboratory data in American Journal of Clinical Pathology, 2010, which is based on Application of Statistical Methods in Medicine [7] published by Hoffmann in 1963. A new
123
J Endocrinol Invest
statistical method is introduced in detail about establishment of reference intervals of hemoglobin, creatinine, urea nitrogen and TSH [8]. Horowitz et al. [9] questioned the method provided in this article and pointed out some flaws in Katayev’s method such as reference intervals that failed to be established when considering factors including age, gender and races, and they believed that it was different from the method preliminarily reported by Hoffman. Dorizzi et al. [10] repeated the establishment of TSH reference intervals by adopting the same method, their results indicated concordance with Katayev’s results. In consideration that preliminary discussion about defining TSH RI by Statistical Method have been done in America and Italy, so our article is intended to establish TSH RIs by Statistical Method, the only difference is that here are some improvements when using Statistical Method, and then to compare the authors’ results with current RI in the laboratory
Materials and methods Source of sample materials All sample materials in this research were from the Clinical Laboratory of the Tumor Hospital Affiliated to Xinjiang Medical University. The samples were collected from various ethnic groups in Xinjiang Uygur Autonomous Region (including Han, Uygur). First, samples with abnormal thyroid function were ruled out including abnormal level of triiodothyronine (T3), thyroxine (T4), serum free triiodothyronine (FT3), serum free thyroxine (FT4), thyroglobulin (TG), thyroid peroxidase antibody (TPOAb) and thyrotropin receptor antibody (TRAb). The Reference values for thyroid hormones were T3 (1.3–3.1 nmol/L), T4 (66–181 nmol/L), FT3 (3.1–6.8 pmol/L), FT4 (12–22 pmol/L), TG (1.4–78 ng/ml), TPOAb (0–34 IU/mL), TRAb (0–1.75 IU/L), any value beyond its RI has been seen as abnormal. Second, samples collected from pregnant women also were ruled out. Lastly, if a sample has several TSH test results, the earliest result was retained, the others were ruled out. The first test always happened before using thyroid medication in most thyroid illness patients, to eliminate the influence of thyroid medication, As far as possible to collect the data in TSH natural state (excluding the medicine or treatment state). There were 10,870 samples in total from remaining people aged from 25 to 85 years, and median age 47 years. The out-patients, in-patients and health-check population were 40.9 % (4,446 cases), 38.3 % (4,163 cases) and 20.8 % (2,261 cases), respectively. The current TSH reference interval is 0.270–4.200 mIU/L, has provided by the manufacturer for the entire Chinese population and also
123
verified in our laboratories. According to the upper limit (4.200 mIU/L), there were 22.85 % (2,484 cases) ‘‘subclinical hypothyroidism’’ in the remaining 10,870 patients. All the TSH test results were obtained after the laboratory LIS system was installed in The Tumor Hospital Affiliated to Xinjiang Medical University, Urumqi, China. Test period of the collected samples were from Jan 2010 to April 2012, all the TSH test samples collected in the hospital laboratory were tested using original devices and supportive reagents from Roche Diagnostics Company, Switzerland. Roche Cobas e601 electrochemical analyser was used. Standard Operating Procedure of the device test items was strictly followed every day. All the tests were completed in clinical laboratory center of The Tumor Hospital Affiliated to Xinjiang Medical University, Urumqi, China. Research methods The research methods were based on the statistical method reported by Hoffman in 1963, and with reference to the method reported by Katayev et al. Then according to our statistical thinking, TSH data were classified by races, genders, age and source of sample to establish TSH reference intervals, and then to compare the results with reference intervals currently applied in the laboratory for verification. Statistical description was made on all statistical data with software including EXCEL and SPSS17.0 to establish the frequency table and work out the distribution curve. Origin8.0 software was used to establish the regression straight line model, and thus to calculate reference intervals accordingly (refer to the results part for details of the research procedure).
Results Establishment of TSH reference intervals Preliminary statistical description over the 10,870 samples Work out the frequency of samples using SPASS17.0: the mean value and standard deviation of the 10,870 samples were 4.402 and 8.981 mIU/L, respectively; percentiles of 2.5 and 97.5 were 0.037 and 24.940 mIU/L; maximum value and minimum value were 99.920 and 0.005 mIU/L; it shows that TSH distribution is skewed, and several outliers could be noticed in the Fig. 1. Preliminary judgment and screening of outliers Outliers of the 10,870 samples were processed with reference to the method in the procedure for establishment of
J Endocrinol Invest
Fig. 1 Histogram of 10,870 TSH samples; horizontal ordinates represent TSH test values, and vertical ordinates represent the frequency. The curve is the TSH normal distribution curve
reference intervals set by CLSI (i.e., the 1/3 rule) [6].After processing of the 10,870 samples, 170 samples were determined to be outliers; the rest 10,700 samples were considered as non-outliers according to the preliminary judgment. There are certain exceptions: as to this test R = 99.920–0.005 = 99.915. For example, there are three continuous test values 61.13, 3.94, 55.43, according to the rule A1 = (3.94–61.13)/99.915; A2 = (55.43–3.94)/ 99.915, absolute values of A1 and A2 are 0.572 and 0.515, then 3.94 becomes an outlier which obviously is a false one. In addition, there are two cases such as 61.13, 73.94, 55.43 and 3.3, 53.94, 75.43 in which true outliers such as 73.94 and 53.94 will not be ruled out. However, these false outliers are very rare and negligible in a test including 10,870 samples. TSH frequency distribution after preliminary exclusion of outliers is shown in Fig. 2. Estimation of reference intervals of the 10,700 processed samples At first, statistical description of the 10,700 samples was made to get the table of frequency distribution and cumulative frequency at the same time. Regression analysis on TSH values and cumulative frequency was made in Origin8.0. With the cumulative frequency as the independent variable (X) and TSH values as the dependent variable (Y), the linear regression equation of the two variables was obtained. The determination coefficient R of the regression equation was taken as the standard to judge efficacy of the equation, the closer the absolute value of R to 1 the better; if used for prediction, the R value shall be[0.99 in general.
Fig. 2 TSH mean value and standard deviation of the 10,700 samples after preliminary exclusion of outliers are 3.45 and 4.322 mIU/L; percentiles of 2.5 and 97.5 are 0.036 and 14.930 mIU/L; maximum value and minimum value are 86.430 and 0.005 mIU/L. The curve is the TSH normal distribution. Rmax Rmin = R (Rmax and Rmin, respectively, refer to maximum value and minimum value of the 10,870 samples). D1 = X - Xi (X is an arbitrary value, Xi is a consecutive value of X); D2 = X – Xo (X is an arbitrary value, Xo is another consecutive value of X). A1 = D1/R; A2 = D2/R (if absolute values of A1 and A2 are both greater than 1/3, then the value is determined to be an outlier.)
Therefore, it is assumed in our article that only the regression line whose R value is [0.99 is the desired model. Regression line established directly based on the 10,700 samples is probably unable to meet the requirement of R value. Therefore, visual analysis (i.e., fuzzy logic theory originated from human beings [8]) is required to find out linear regression intervals suitable for the independent variables as well as the dependent variables, and then to establish a suitable linear regression model and to find out suitable reference intervals (Fig. 3). According to the definition of reference interval, the model established based on the regression equation takes the TSH values (Y) corresponding to cumulative frequency (X) 2.5 and 97.5 % as the upper and lower limits of reference intervals: RImin ¼ 0:04659 2:5 þ 0:24003 ¼ 0:366 ðmIU=LÞ RImax ¼ 0:04659 97:5 þ 0:24003 ¼ 4:782 ðmIU=LÞ Establishment of TSH RIs for groups of various races, genders, age and sample population Establishment of TSH RIs for groups of various races, genders, age and source of sample populations according to the method as result 1.3, the detail information of these RIs was shown in Table 1, 2 and 3; Fig. 4. It is shown in Fig. 4 that the female’s RI is higher than the male’s in same factor, the broadest RI range is the female in Han race population (0.373–4.979). the
123
J Endocrinol Invest
Fig. 3 Regression analysis of the 10,700 samples is shown in the diagram on the left: the regression curve is Y = 0.26065X - 3.96926, and R value is 0.31382; according to the significance of R value, it could be confirmed that there is no significance in RI when the equation is used to estimating TSH. The regression line of 10–80 %
cumulative frequency and TSH values obtained after visual analysis is shown in the diagram on the right; the regression line is Y = 0.04659X ? 0.24003; the R value is 0.99567; R [ 0.99. It is in accordance with R value setting, and the regression line model meets relevant requirements
Table 1 TSH reference intervals of various races groups in Xinjiang Uygur autonomous regions Groups
Total number
Linear range (%)
Regression equation
R value
Reference intervals (mIU/L)
The entire population
10,700
10–80
Y = 0.04659X ? 0.24003
0.99567
0.366–4.782
Han (female)
6,598
10–75
Y = 0.04849X ? 0.25142
0.99354
0.373–4.979
Han (male)
1,169
10–75
Y = 0.03708X ? 0.34394
0.99418
0.437–3.959
Uygur (female)
1,259
10–80
Y = 0.04074X ? 0.13108
0.99194
0.233–4.103
166
10–60
Y = 0.03142X ? 0.02113
0.99223
0.099–3.085
Uygur (male)
Table 2 TSH reference intervals of various age groups in Xinjiang Uygur autonomous regions Groups
Total number
Linear range (%)
Regression equation
R value
Reference intervals (mIU/L)
\45 years-old Female Male
4,114
5–80
Y = 0.04762X ? 0.21630
0.99116
0.335–4.859
609
15–80
Y = 0.03781X ? 0.20998
0.99318
0.304–3.890
45–65 years-old Female Male
3,936
5–80
Y = 0.04310X ? 0.24501
0.99165
0.353-4.447
612
15–80
Y = 0.03652X ? 0.27543
0.99295
0.367–3.836
[65 years-old Female
507
5–75
Y = 0.04594X ? 0.13257
0.99130
0.247–4.612
Male
301
5–75
Y = 0.04038X ? 0.18983
0.99010
0.291–4.127
123
J Endocrinol Invest Table 3 TSH reference intervals for different source of sample populations Groups
Total number
Linear range (%)
Regression equation
R value
Reference intervals (mIU/L)
3,315
5–80
Y = 0.04418X ? 0.22352
0.99143
0.334–4.531
523
5–75
Y = 0.03999X ? 0.29464
0.99079
0.395–4.194
3,595
5–80
Y = 0.04741X ? 0.16162
0.99175
0.280–4.784
604
5–80
Y = 0.03923X ? 0.17013
0.99070
0.268–3.996
1,861
5–80
Y = 0.04672X ? 0.13055
0.99096
0.247–4.686
342
5–80
Y = 0.04292X ? 0.16371
0.99051
0.271–4.348
Out-patients Female Male In-patients Female Male Health-check Female Male
Fig. 4 The summary of different groups populations’ RIs from Tables 1, 2 and 3. In the figure, the manufacturer’s RI is 0.270–4.200 mIU/L, which is also the laboratory current RI
narrowest RI range is the male in Uygur race population (0.099–3.085). Conclusion of results of reference intervals for different groups and their comparison with reference intervals provided by manufacturers Reference change value (RCV) is used to evaluate the reference interval (RI) obtained based on laboratory data and RI currently used in the laboratory, and to determine whether there are statistical differences between different RIs. 1=2 RCV ¼ 21=2 Z CV2a þ CV2W Z is 95 % estimation interval, according to statistical regulations, Z value is 1.96; CVa is the analysis variability of this laboratory analysis program; with reference to the result of internal data control in the laboratory since the laboratory began to collect data, the internal quality control CV% of the laboratory in this period was 3.8 %; internal
quality control CV was considered as the actual analysis CV of the program, i.e., CVa, CVW is the inter-individual variability, and CVW was confirmed to be 19.3 % with reference to relevant data [11]. Calculated RCV ¼ 21=2 1:96 3:8%2 þ 19:3%2 ¼ 54:5%
1=2
Reference interval provided by the reagent manufacturer is 0.270–4.200 mIU/L. Differences between the reference intervals are calculated (including RCVmin and RCVmax). RCVmin ¼ RImin RI0min =RI0min RI0min is the smaller of two: RCVmax ¼ RImax RI0max =RI0max RI0max is the larger of two: RCVmin = (0.366–0.270)/0.270 9 100 % = 35.6 %; RCVmax = (4.782–4.200)/4.200 9 100 % = 13.9 %.
123
J Endocrinol Invest Table 4 Comparison of reference change values between reference intervals of different races gender groups and reference intervals provided by manufacturer (RCVmin and RCVmax) Entire population
Han (female)
Han (male)
Uygur (female)
Uygur (male)
Manufacturer (%) RCVmin
35.6
38.1
61.9a
15.9
172.7a
RCVmax
13.9
18.5
6.1
2.4
36.1
Table 7 Change frequencies of abnormal TSH with CI of manufacturer RI and the manufacturer RI TSH decreased (%)
TSH elevated (%)#
Manufacturer RI
622 (5.72)
2,484 (22.85)
CI of manufacturer RI
579 (5.33)
1,886 (17.35)
#
Compare the abnormal frequencies with SPASS17.0, P \ 0.001, has statistical differences
a
Either RCVmin or RCVmax calculated between two reference intervals is larger than the allowable RCV
Compare the calculated RCVmin and RCVmax with RCV (54.5 %); if RCVmin and RCVmax are smaller than RCV, there is no difference between two RIs; if RCVmin and RCVmax are larger than RCV, there is difference between two RIs. It could be seen from Tables 4, 5 and 6 that statistical differences are found between the lower limits of TSH references intervals and those provided by the manufacturer only for Han male group, Uygur male group and outpatient male group. It is thus indicated that there is almost no difference between RIs currently used in the laboratory and that obtained based on laboratory data. Attention shall only be paid to the lower limits of RIs of some male groups. So establish the new TSH RI (0.233–4.979 mIU/ L), the lower limit is 0.233 mIU/L, which is the lowest level exclude the statistical differences number (including Han male group, Uygur male group and out-patient male group); the high limit is 4.979 mIU/L, which is the highest level in various RIs. The new RI expanded the manufacturer RI (0.270–4.200 mIU/L), but has no statistical differences with manufacturer RI, it can be seen as the confidence interval (CI) of manufacturer RI.
Comparison of the change of the abnormal TSH percent with the new RI (CI of manufacturer RI) and the manufacturer RI The results of calculation of the frequencies of abnormal TSH in 10,870 case samples are shown in Table 7:
Discussion For the establishment of laboratory reference intervals, the National Clinical Test Operation Regulation issued by China’s Ministry of Health has provided reference intervals under certain confirmed test system conditions. With persistent development of laboratory medicine, new test items and test methods are created continuously, authorities will be unable to provide reference intervals meeting clinical requirements of different groups in different regions. Therefore, at present, it is necessary for clinical laboratories to establish reference intervals fit for the laboratories according to requirements of their objective groups, and it also is essential requirement for a laboratory to conduct medical laboratory quality control and laboratory accreditation. However, according to requirements of CLSI document, it takes enormous time and efforts to establish
Table 5 Comparison of reference change values between reference intervals of different sample populations groups and reference intervals provided by manufacturers (RCVmin and RCVmax) Out-patient (female)
Out-patient (male)
In-patient (female)
In-patient (male)
Health-check (female)
Health-check (male)
Manufacturer (%)
a
RCVmin
27.4
64.8a
3.7
0.7
9.3
0.4
RCVmax
7.9
0.1
13.9
5.1
11.6
3.5
Either RCVmin or RCVmax calculated between two reference intervals is larger than the allowable RCV
Table 6 Comparison of reference change values between reference intervals of different age groups and reference intervals provided by manufacturer (RCVmin and RCVmax) \45 year-old (female)
\45 year-old (male)
45-65 year-old (female)
45-65 year-old (male)
[65 year-old (female)
[65 year-old (male)
24.1 15.7
12.6 8.0
30.7 5.9
35.9 9.5
9.3 9.8
7.8 2.0
Manufacturer (%) RCVmin RCVmax
123
J Endocrinol Invest
reference intervals for a laboratory. There are also many requirements and limitations even for translation and establishment of reference intervals based on those provided by the manufacturer according to relevant requirements, and it is difficult for most laboratories to meet these requirements. Therefore, it is necessary to seek for other methods to establish and verify laboratory RIs. Hoffmann in 1960s had been awared that there seemed to be no much connection among clinicians, laboratory personnel and statisticians before, but the truth was that there should be more connection among them. Laboratory personnel could obtain reference intervals with better clinical significance only by adopting relevant statistical knowledge in combination with clinical significance of test indexes. It is considered that an excellent laboratory person should not only understand clinician knowledge but also have certain degree of statistical knowledge to conduct statistical analysis on laboratory data. Hoffmann has already preliminarily discussed about the feasibility of establishing appropriate reference intervals based on available test data in the laboratory adopting statistical method when electronic computer technology was still limited. Ilcol and Aslan [12] also set the RIs for forty test items in his laboratory by analyzing the stored data, and thought that the new RIs are more suitable than the RIs provided by the International Federation of Clinical Chemistry and Laboratory Medicine or manufacturers. With continuous development of electronic computer technology, new computer software could be used to realize statistical analysis of laboratory data simply and conveniently. Katayev et al. [8] made improvements to Hoffmann’s method by brand new computer technology, got reference intervals comparatively consistent with those reported in literatures, preliminarily verified the theory that reference intervals more concordant with requirements could be obtained based on laboratory data with the help of statistical software. Dorizzi et al. [10] in Italy verified the feasibility of Katayev’s method again. Although Horowitz [9] pointed out differences between Katayev’s method and Hoffmann’s, one point was neglected that greater development had been made on computer technology compared with that in Hoffmann’s time. Restricted by application of computer technology, Hoffmann in those days could only discuss about the estimation model using 60 data to establish reference intervals. Now with the help of software, data of 10,000 of samples could be processed fast and easily. Statistically speaking, estimation made using large volume of samples could definitely produce much fewer errors than that made using small volume of samples, and also the results are more convinced. Therefore, Katayev’s method is certainly an improvement of Hoffmann’s method, and the so-called differences are actually a kind of improvement to the previous method.
Comparing the method of RI establishment in this article with Katayev’s method, further improvements are made. First, according to Katayev’s method, only visual observation was adopted to define the linear interval of the regression straight line, and the significance of R value which was the important determination coefficient of the regression straight line was not mentioned; in this article, the determination coefficient is required to meet R [ 0.99 when visual observation is adopted to determine the linear interval and to establish the model of regression straight line. Therefore, the advantage of the fuzzy logic theory of human being is taken and strict mathematical logic is added to make the results more convincing. Second, in Katayev’s method, the data from the abnormal thyroid function individuals cannot be ruled out, but in the optimized method, the step was done as first step. In Katayev’s method, Chauvenet’s criterion has been used, but in the optimized method, the 1/3 rule according to CLSI has been used. In Chauvenet’s criterion, when the total data are more than 1,000, ratio of maximum acceptable deviation to standard deviation (dmax/SD) is 3.48. In the study, SD is 8.98 mIU/L, mean value is 4.40 mIU/L. So the minimum eliminated data is 35.65 mIU/L, a lot of abnormal data cannot be ruled out such as 35.64 mIU/L and so on. In the large sample study, Standard Deviation of the data is always small, and many abnormal data cannot be extremely ruled out. In other side, the Chauvenet’s criterion cannot rule out the low value, for example, 0.01 mIU/L also has no chance to be ruled out. Compared to Chauvenet’s criterion, the 1/3 rule is recommended by CLSI, which can better rule out the abnormal data. In addition, TSH RIs are established for groups of different races, genders, age and source of sample according to relevant requirements to remedy the defect of previously reported methods. Of course, the optimized method also has some defects, to establish suitable linear regression model is difficult, especially though the easy computer software such as Origin8.0 software only set linear regression model mechanically. We must eliminate the unlinear range data step by step through subjective observation to get the R [ 0.99 linear model. So we can see the linear range is decreased every 5 %. The linear model which we get may be not the best model, but only the better model. There is no obviously difference between various TSH RIs based on laboratory data analysis and RIs (0.270–4.200 mIU/L) currently used in the laboratory. According to results of test, certain differences were found between the RI lower limits of Han males, Uygur males, out-patient males and used in the laboratory currently. If we consider only the attention the upper TSH reference limit, no significant difference between TSH RIs of various races, genders, age, sample populations and that currently
123
J Endocrinol Invest
used in the laboratory can be seen, which means that RIs provided by the manufacturer can be used for the entire population. So it also indicates that the improved method could be used for verification of TSH RIs currently used in the laboratory. Then it get a new RI expanded the current RI, which can be seen as the confidence interval (CI) of manufacturer RI. Why lower limit of some male RIs have statistical differences? First, the differences from the statistical error and the character of special male populations such as smoking status and body mass index (BMI) [13] and so on. Second, data collected in the laboratory, and samples of some unhealthy groups are included (only samples with TSH index-related abnormities are ruled out). The objective is to get the distribution curve of TSH test values of the entire population in natural state, and thus to establish relevant model according to the distribution curve. Compared with the large sample investigation, a narrow TSH RI rage will be offen defined based on strictly criteria selection small sample investigation [14]. The CLSI method is a strict criteria selection small sample investigation, and using the stored data is a large sample investigation. Therefore, RIs obtained by our method has a wider range than RIs obtained by adopting CLSI reference method. This should be the reason why there are differences between the lower limits of RIs of male groups with that of current RI in the laboratory. From the result of comparing the change in the abnormal TSH percent according to the new RI (CI of manufacturer RI) and the manufacturer RI, it shows that the change in RI upper limit, the abnormal TSH percent, easily changes. Because the upper limit is related to the subclinical hypothyroidism, most sub-clinical hypothyroidism occurs in female population. The female number is large in the entire TSH test population, so the upper limit RI is very important. If the upper limit is uncomfortable, the subclinical hypothyroidism will be omission diagnosis and excessive diagnosis. The manufacturer provides the RI to laboratory, but does not give comfortable confidence interval. So here, we using the Optimized method firstly verified the current RI, then see the new RI as the CI of current RI, if an individual test result is 4.200–4.979 mIU/ L, we can see it as ‘‘high risk sub-clinical hypothyroidism’’, then ask for patient follow-up. The benefit of doing these is to decrease the omission diagnosis and excessive diagnosis, and to reduce the patient’s psychological pressure.
Inference Regression straight line model and appropriate TSH reference intervals could be established using test data
123
collected from the laboratory as well as statistical analysis method. This method is also adopted to verify TSH RIs currently used in the laboratory. At present, result verification among laboratories is a trend of laboratory medicine development. TSH RIs should be set using the optimized statistical method by analyzing laboratory-stored data from multicenter laboratories with the same test conditions. It can provide help for result verification among local laboratories. Conflict of interest The authors Feng Yang-chun, Bian Wen-an, Mu Chao-dong, Xu Yi, Wang Fei, Qiao Wen-bin, and Huang Yanchun declare that they have no conflict of interest.
References 1. Surks MI, Ortiz E, Daniels GH et al (2004) Subclinical thyroid disease: scientific review and guidelines for diagnosis and management. J Am Med Assoc 291(2):228–238 2. Gietka-Czernel M, Jastrzebska H (2007) Thyrotropin reference range–should it be changed? Endokrynol Pol 58(5):454–560 3. Surks MI, Goswami G, Daniels GH (2005) The thyrotropin reference range should remain unchanged. J Clin Endocrinol Metab 90(9):5489–5496 4. Laurberg P, Andersen S, Carle´ A et al (2011) The TSH upper reference limit: where are we at? Nat Rev Endocrinol 7(4): 232–239 5. Schalin-Ja¨ntti C, Tanner P, Va¨lima¨ki MJ et al (2011) Serum TSH reference interval in healthy Finnish adults using the Abbott Architect 2000i Analyzer. Scand J Clin Lab Invest 71(4):344–349 6. Clinical and Laboratory Standards Institute (2008) Defining, establishing, and verifying reference intervals in the clinical laboratory; approved guideline, 3rd edn. Clinical and Laboratory Standards Institute, Wayne (CLSI document C28-A3) 7. Hoffmann RG (1963) Statistics in the practice of medicine. JAMA 185:864–873 8. Katayev A, Balciza C, David W et al (2010) Establishing reference intervals for clinical laboratory test results: is there a better way? Am J Clin Pathol 133(2):180–186 9. Horowitz GL (2010) Estimating reference intervals. Am J Clin Pathol 133(2):175–177 10. Dorizzi RM, Giannone G et al (2011) Indirect methods for TSH reference interval: at last fit for purpose? Am J Clin Pathol 135(1):167–174 11. Ricos C, Alvarez V, Cava F, et al. Desirable specifications for total error, imprecision, and bias, derived from biologic variation. Westgard QC Web page. http://www.westgard.com/biodatabase1. htm. Accessed 5 March 2008 12. Ilcol YO, Aslan D (2006) Use of total patient data for indirect estimation of reference intervals for 40 clinical chemical analytes in Turkey. Clin Chem Lab Med 44(7):867–876 13. Lucas A, Granada ML, Olaizola I et al (2013) Leptin and thyrotropin relationship is modulated by smoking status in euthyroid subjects [J]. Thyroid 23(8):964–970 14. Spencer CA, Hollowell JG, Kazarosyan M et al (2007) National health and nutrition examination survey III thyroid-stimulating hormone (TSH)-thyroperoxidase antibody relationships demonstrate that TSH upper reference limits may be skewed by occult thyroid dysfunction. J Clin Endocrinol Metab 92(11):4236–4240
ORIGINAL ARTICLE
Establish and verify TSH reference intervals using optimized statistical method by analyzing laboratory-stored data Y. Feng • W. Bian • C. Mu • Y. Xu F. Wang • W. Qiao • Y. Huang
•
Received: 30 July 2013 / Accepted: 16 November 2013 Ó Italian Society of Endocrinology (SIE) 2013
Abstract Objective To establish reference intervals using an optimized statistical method by collecting available laboratory data of thyroid stimulating hormone (TSH), and then to verify with the laboratory-present reference intervals. Methods TSH RIs of the total population and different races, genders, age, source of sample are established through improved Hoffmann and Katayev’s method with TSH test results data from Jan 2010 to April 2012 were collected, and finally conduct comparative verification with the laboratory present RIs. Results According to the improved method, we get various RIs of different sample populations. On comparing with the laboratory current RI (0.270–4.200 mIU/L) most reference change values (RCV) were within acceptable limits. Only lower limit of Han male, Uygur male and outpatient male populations outwith acceptable limits. On excluding the different values, finally, the new RI by the optimized statistical method is 0.233–4.979 mIU/L. Because the new RI expanded the current RI and was not different from the current RI, it was indicated that new RI could be used to verify the laboratory current RIs and seen as the current RI’s confidence interval (CI). Inference TSH RIs established by optimized Hoffmann’s and Katayev’s methods is viable and can be used to verify RIs provided by manufacturers or other laboratories. Keywords Reference intervals Thyroid stimulating hormone Verification
Y. Feng W. Bian C. Mu Y. Xu F. Wang W. Qiao Y. Huang (&) The Clinical Laboratory Center, The Tumor Hospital Affiliated to Xin Jiang Medical University, Urumqi, China e-mail: [email protected]
Introduction To establish suitable thyroid stimulating hormone (TSH) reference intervals (RI) which is regarded as an evidence for the diagnosis of sub-clinical hypothyroidism is important in clinical laboratory [1]. A special concern is that the TSH upper reference limit which defines subclinical hypothyroidism is in violent controversy [2–4]. So Finnish report redefined the TSH reference interval in the adult Finnish population [5]. Current TSH RI has also been criticized because it was established from unselected background populations which are provided by internationalized device or reagent manufacturers in China. Since all the reagents or devices are almost intended to be marketed in a nationwide setting, the reference intervals provided by the manufacturers are impossible to meet the demands of Chinese patients, including 1.3 billion people and 56 different ethnic groups. It is very necessary for establishing various TSH RIs for different ethnic groups to diagnose sub-clinical hypothyroidism accurately. How to define a RI in various laboratories? The American Clinical and Laboratory Standards Institute (CLSI) has provided authoritative standards [6], but it is difficult to meet those standards in practice. As required by CLSI, all clinical laboratories which adopt reference intervals provided by other laboratories or manufacturers shall verify the reference intervals. However, statistically, serious sampling bias could be inevitable when 20 samples are collected from local so-called healthy people and used to verify laboratory results obtained from large-scale tests. Therefore, Katayev and his colleagues published an article about establishment of reference intervals using available laboratory data in American Journal of Clinical Pathology, 2010, which is based on Application of Statistical Methods in Medicine [7] published by Hoffmann in 1963. A new
123
J Endocrinol Invest
statistical method is introduced in detail about establishment of reference intervals of hemoglobin, creatinine, urea nitrogen and TSH [8]. Horowitz et al. [9] questioned the method provided in this article and pointed out some flaws in Katayev’s method such as reference intervals that failed to be established when considering factors including age, gender and races, and they believed that it was different from the method preliminarily reported by Hoffman. Dorizzi et al. [10] repeated the establishment of TSH reference intervals by adopting the same method, their results indicated concordance with Katayev’s results. In consideration that preliminary discussion about defining TSH RI by Statistical Method have been done in America and Italy, so our article is intended to establish TSH RIs by Statistical Method, the only difference is that here are some improvements when using Statistical Method, and then to compare the authors’ results with current RI in the laboratory
Materials and methods Source of sample materials All sample materials in this research were from the Clinical Laboratory of the Tumor Hospital Affiliated to Xinjiang Medical University. The samples were collected from various ethnic groups in Xinjiang Uygur Autonomous Region (including Han, Uygur). First, samples with abnormal thyroid function were ruled out including abnormal level of triiodothyronine (T3), thyroxine (T4), serum free triiodothyronine (FT3), serum free thyroxine (FT4), thyroglobulin (TG), thyroid peroxidase antibody (TPOAb) and thyrotropin receptor antibody (TRAb). The Reference values for thyroid hormones were T3 (1.3–3.1 nmol/L), T4 (66–181 nmol/L), FT3 (3.1–6.8 pmol/L), FT4 (12–22 pmol/L), TG (1.4–78 ng/ml), TPOAb (0–34 IU/mL), TRAb (0–1.75 IU/L), any value beyond its RI has been seen as abnormal. Second, samples collected from pregnant women also were ruled out. Lastly, if a sample has several TSH test results, the earliest result was retained, the others were ruled out. The first test always happened before using thyroid medication in most thyroid illness patients, to eliminate the influence of thyroid medication, As far as possible to collect the data in TSH natural state (excluding the medicine or treatment state). There were 10,870 samples in total from remaining people aged from 25 to 85 years, and median age 47 years. The out-patients, in-patients and health-check population were 40.9 % (4,446 cases), 38.3 % (4,163 cases) and 20.8 % (2,261 cases), respectively. The current TSH reference interval is 0.270–4.200 mIU/L, has provided by the manufacturer for the entire Chinese population and also
123
verified in our laboratories. According to the upper limit (4.200 mIU/L), there were 22.85 % (2,484 cases) ‘‘subclinical hypothyroidism’’ in the remaining 10,870 patients. All the TSH test results were obtained after the laboratory LIS system was installed in The Tumor Hospital Affiliated to Xinjiang Medical University, Urumqi, China. Test period of the collected samples were from Jan 2010 to April 2012, all the TSH test samples collected in the hospital laboratory were tested using original devices and supportive reagents from Roche Diagnostics Company, Switzerland. Roche Cobas e601 electrochemical analyser was used. Standard Operating Procedure of the device test items was strictly followed every day. All the tests were completed in clinical laboratory center of The Tumor Hospital Affiliated to Xinjiang Medical University, Urumqi, China. Research methods The research methods were based on the statistical method reported by Hoffman in 1963, and with reference to the method reported by Katayev et al. Then according to our statistical thinking, TSH data were classified by races, genders, age and source of sample to establish TSH reference intervals, and then to compare the results with reference intervals currently applied in the laboratory for verification. Statistical description was made on all statistical data with software including EXCEL and SPSS17.0 to establish the frequency table and work out the distribution curve. Origin8.0 software was used to establish the regression straight line model, and thus to calculate reference intervals accordingly (refer to the results part for details of the research procedure).
Results Establishment of TSH reference intervals Preliminary statistical description over the 10,870 samples Work out the frequency of samples using SPASS17.0: the mean value and standard deviation of the 10,870 samples were 4.402 and 8.981 mIU/L, respectively; percentiles of 2.5 and 97.5 were 0.037 and 24.940 mIU/L; maximum value and minimum value were 99.920 and 0.005 mIU/L; it shows that TSH distribution is skewed, and several outliers could be noticed in the Fig. 1. Preliminary judgment and screening of outliers Outliers of the 10,870 samples were processed with reference to the method in the procedure for establishment of
J Endocrinol Invest
Fig. 1 Histogram of 10,870 TSH samples; horizontal ordinates represent TSH test values, and vertical ordinates represent the frequency. The curve is the TSH normal distribution curve
reference intervals set by CLSI (i.e., the 1/3 rule) [6].After processing of the 10,870 samples, 170 samples were determined to be outliers; the rest 10,700 samples were considered as non-outliers according to the preliminary judgment. There are certain exceptions: as to this test R = 99.920–0.005 = 99.915. For example, there are three continuous test values 61.13, 3.94, 55.43, according to the rule A1 = (3.94–61.13)/99.915; A2 = (55.43–3.94)/ 99.915, absolute values of A1 and A2 are 0.572 and 0.515, then 3.94 becomes an outlier which obviously is a false one. In addition, there are two cases such as 61.13, 73.94, 55.43 and 3.3, 53.94, 75.43 in which true outliers such as 73.94 and 53.94 will not be ruled out. However, these false outliers are very rare and negligible in a test including 10,870 samples. TSH frequency distribution after preliminary exclusion of outliers is shown in Fig. 2. Estimation of reference intervals of the 10,700 processed samples At first, statistical description of the 10,700 samples was made to get the table of frequency distribution and cumulative frequency at the same time. Regression analysis on TSH values and cumulative frequency was made in Origin8.0. With the cumulative frequency as the independent variable (X) and TSH values as the dependent variable (Y), the linear regression equation of the two variables was obtained. The determination coefficient R of the regression equation was taken as the standard to judge efficacy of the equation, the closer the absolute value of R to 1 the better; if used for prediction, the R value shall be[0.99 in general.
Fig. 2 TSH mean value and standard deviation of the 10,700 samples after preliminary exclusion of outliers are 3.45 and 4.322 mIU/L; percentiles of 2.5 and 97.5 are 0.036 and 14.930 mIU/L; maximum value and minimum value are 86.430 and 0.005 mIU/L. The curve is the TSH normal distribution. Rmax Rmin = R (Rmax and Rmin, respectively, refer to maximum value and minimum value of the 10,870 samples). D1 = X - Xi (X is an arbitrary value, Xi is a consecutive value of X); D2 = X – Xo (X is an arbitrary value, Xo is another consecutive value of X). A1 = D1/R; A2 = D2/R (if absolute values of A1 and A2 are both greater than 1/3, then the value is determined to be an outlier.)
Therefore, it is assumed in our article that only the regression line whose R value is [0.99 is the desired model. Regression line established directly based on the 10,700 samples is probably unable to meet the requirement of R value. Therefore, visual analysis (i.e., fuzzy logic theory originated from human beings [8]) is required to find out linear regression intervals suitable for the independent variables as well as the dependent variables, and then to establish a suitable linear regression model and to find out suitable reference intervals (Fig. 3). According to the definition of reference interval, the model established based on the regression equation takes the TSH values (Y) corresponding to cumulative frequency (X) 2.5 and 97.5 % as the upper and lower limits of reference intervals: RImin ¼ 0:04659 2:5 þ 0:24003 ¼ 0:366 ðmIU=LÞ RImax ¼ 0:04659 97:5 þ 0:24003 ¼ 4:782 ðmIU=LÞ Establishment of TSH RIs for groups of various races, genders, age and sample population Establishment of TSH RIs for groups of various races, genders, age and source of sample populations according to the method as result 1.3, the detail information of these RIs was shown in Table 1, 2 and 3; Fig. 4. It is shown in Fig. 4 that the female’s RI is higher than the male’s in same factor, the broadest RI range is the female in Han race population (0.373–4.979). the
123
J Endocrinol Invest
Fig. 3 Regression analysis of the 10,700 samples is shown in the diagram on the left: the regression curve is Y = 0.26065X - 3.96926, and R value is 0.31382; according to the significance of R value, it could be confirmed that there is no significance in RI when the equation is used to estimating TSH. The regression line of 10–80 %
cumulative frequency and TSH values obtained after visual analysis is shown in the diagram on the right; the regression line is Y = 0.04659X ? 0.24003; the R value is 0.99567; R [ 0.99. It is in accordance with R value setting, and the regression line model meets relevant requirements
Table 1 TSH reference intervals of various races groups in Xinjiang Uygur autonomous regions Groups
Total number
Linear range (%)
Regression equation
R value
Reference intervals (mIU/L)
The entire population
10,700
10–80
Y = 0.04659X ? 0.24003
0.99567
0.366–4.782
Han (female)
6,598
10–75
Y = 0.04849X ? 0.25142
0.99354
0.373–4.979
Han (male)
1,169
10–75
Y = 0.03708X ? 0.34394
0.99418
0.437–3.959
Uygur (female)
1,259
10–80
Y = 0.04074X ? 0.13108
0.99194
0.233–4.103
166
10–60
Y = 0.03142X ? 0.02113
0.99223
0.099–3.085
Uygur (male)
Table 2 TSH reference intervals of various age groups in Xinjiang Uygur autonomous regions Groups
Total number
Linear range (%)
Regression equation
R value
Reference intervals (mIU/L)
\45 years-old Female Male
4,114
5–80
Y = 0.04762X ? 0.21630
0.99116
0.335–4.859
609
15–80
Y = 0.03781X ? 0.20998
0.99318
0.304–3.890
45–65 years-old Female Male
3,936
5–80
Y = 0.04310X ? 0.24501
0.99165
0.353-4.447
612
15–80
Y = 0.03652X ? 0.27543
0.99295
0.367–3.836
[65 years-old Female
507
5–75
Y = 0.04594X ? 0.13257
0.99130
0.247–4.612
Male
301
5–75
Y = 0.04038X ? 0.18983
0.99010
0.291–4.127
123
J Endocrinol Invest Table 3 TSH reference intervals for different source of sample populations Groups
Total number
Linear range (%)
Regression equation
R value
Reference intervals (mIU/L)
3,315
5–80
Y = 0.04418X ? 0.22352
0.99143
0.334–4.531
523
5–75
Y = 0.03999X ? 0.29464
0.99079
0.395–4.194
3,595
5–80
Y = 0.04741X ? 0.16162
0.99175
0.280–4.784
604
5–80
Y = 0.03923X ? 0.17013
0.99070
0.268–3.996
1,861
5–80
Y = 0.04672X ? 0.13055
0.99096
0.247–4.686
342
5–80
Y = 0.04292X ? 0.16371
0.99051
0.271–4.348
Out-patients Female Male In-patients Female Male Health-check Female Male
Fig. 4 The summary of different groups populations’ RIs from Tables 1, 2 and 3. In the figure, the manufacturer’s RI is 0.270–4.200 mIU/L, which is also the laboratory current RI
narrowest RI range is the male in Uygur race population (0.099–3.085). Conclusion of results of reference intervals for different groups and their comparison with reference intervals provided by manufacturers Reference change value (RCV) is used to evaluate the reference interval (RI) obtained based on laboratory data and RI currently used in the laboratory, and to determine whether there are statistical differences between different RIs. 1=2 RCV ¼ 21=2 Z CV2a þ CV2W Z is 95 % estimation interval, according to statistical regulations, Z value is 1.96; CVa is the analysis variability of this laboratory analysis program; with reference to the result of internal data control in the laboratory since the laboratory began to collect data, the internal quality control CV% of the laboratory in this period was 3.8 %; internal
quality control CV was considered as the actual analysis CV of the program, i.e., CVa, CVW is the inter-individual variability, and CVW was confirmed to be 19.3 % with reference to relevant data [11]. Calculated RCV ¼ 21=2 1:96 3:8%2 þ 19:3%2 ¼ 54:5%
1=2
Reference interval provided by the reagent manufacturer is 0.270–4.200 mIU/L. Differences between the reference intervals are calculated (including RCVmin and RCVmax). RCVmin ¼ RImin RI0min =RI0min RI0min is the smaller of two: RCVmax ¼ RImax RI0max =RI0max RI0max is the larger of two: RCVmin = (0.366–0.270)/0.270 9 100 % = 35.6 %; RCVmax = (4.782–4.200)/4.200 9 100 % = 13.9 %.
123
J Endocrinol Invest Table 4 Comparison of reference change values between reference intervals of different races gender groups and reference intervals provided by manufacturer (RCVmin and RCVmax) Entire population
Han (female)
Han (male)
Uygur (female)
Uygur (male)
Manufacturer (%) RCVmin
35.6
38.1
61.9a
15.9
172.7a
RCVmax
13.9
18.5
6.1
2.4
36.1
Table 7 Change frequencies of abnormal TSH with CI of manufacturer RI and the manufacturer RI TSH decreased (%)
TSH elevated (%)#
Manufacturer RI
622 (5.72)
2,484 (22.85)
CI of manufacturer RI
579 (5.33)
1,886 (17.35)
#
Compare the abnormal frequencies with SPASS17.0, P \ 0.001, has statistical differences
a
Either RCVmin or RCVmax calculated between two reference intervals is larger than the allowable RCV
Compare the calculated RCVmin and RCVmax with RCV (54.5 %); if RCVmin and RCVmax are smaller than RCV, there is no difference between two RIs; if RCVmin and RCVmax are larger than RCV, there is difference between two RIs. It could be seen from Tables 4, 5 and 6 that statistical differences are found between the lower limits of TSH references intervals and those provided by the manufacturer only for Han male group, Uygur male group and outpatient male group. It is thus indicated that there is almost no difference between RIs currently used in the laboratory and that obtained based on laboratory data. Attention shall only be paid to the lower limits of RIs of some male groups. So establish the new TSH RI (0.233–4.979 mIU/ L), the lower limit is 0.233 mIU/L, which is the lowest level exclude the statistical differences number (including Han male group, Uygur male group and out-patient male group); the high limit is 4.979 mIU/L, which is the highest level in various RIs. The new RI expanded the manufacturer RI (0.270–4.200 mIU/L), but has no statistical differences with manufacturer RI, it can be seen as the confidence interval (CI) of manufacturer RI.
Comparison of the change of the abnormal TSH percent with the new RI (CI of manufacturer RI) and the manufacturer RI The results of calculation of the frequencies of abnormal TSH in 10,870 case samples are shown in Table 7:
Discussion For the establishment of laboratory reference intervals, the National Clinical Test Operation Regulation issued by China’s Ministry of Health has provided reference intervals under certain confirmed test system conditions. With persistent development of laboratory medicine, new test items and test methods are created continuously, authorities will be unable to provide reference intervals meeting clinical requirements of different groups in different regions. Therefore, at present, it is necessary for clinical laboratories to establish reference intervals fit for the laboratories according to requirements of their objective groups, and it also is essential requirement for a laboratory to conduct medical laboratory quality control and laboratory accreditation. However, according to requirements of CLSI document, it takes enormous time and efforts to establish
Table 5 Comparison of reference change values between reference intervals of different sample populations groups and reference intervals provided by manufacturers (RCVmin and RCVmax) Out-patient (female)
Out-patient (male)
In-patient (female)
In-patient (male)
Health-check (female)
Health-check (male)
Manufacturer (%)
a
RCVmin
27.4
64.8a
3.7
0.7
9.3
0.4
RCVmax
7.9
0.1
13.9
5.1
11.6
3.5
Either RCVmin or RCVmax calculated between two reference intervals is larger than the allowable RCV
Table 6 Comparison of reference change values between reference intervals of different age groups and reference intervals provided by manufacturer (RCVmin and RCVmax) \45 year-old (female)
\45 year-old (male)
45-65 year-old (female)
45-65 year-old (male)
[65 year-old (female)
[65 year-old (male)
24.1 15.7
12.6 8.0
30.7 5.9
35.9 9.5
9.3 9.8
7.8 2.0
Manufacturer (%) RCVmin RCVmax
123
J Endocrinol Invest
reference intervals for a laboratory. There are also many requirements and limitations even for translation and establishment of reference intervals based on those provided by the manufacturer according to relevant requirements, and it is difficult for most laboratories to meet these requirements. Therefore, it is necessary to seek for other methods to establish and verify laboratory RIs. Hoffmann in 1960s had been awared that there seemed to be no much connection among clinicians, laboratory personnel and statisticians before, but the truth was that there should be more connection among them. Laboratory personnel could obtain reference intervals with better clinical significance only by adopting relevant statistical knowledge in combination with clinical significance of test indexes. It is considered that an excellent laboratory person should not only understand clinician knowledge but also have certain degree of statistical knowledge to conduct statistical analysis on laboratory data. Hoffmann has already preliminarily discussed about the feasibility of establishing appropriate reference intervals based on available test data in the laboratory adopting statistical method when electronic computer technology was still limited. Ilcol and Aslan [12] also set the RIs for forty test items in his laboratory by analyzing the stored data, and thought that the new RIs are more suitable than the RIs provided by the International Federation of Clinical Chemistry and Laboratory Medicine or manufacturers. With continuous development of electronic computer technology, new computer software could be used to realize statistical analysis of laboratory data simply and conveniently. Katayev et al. [8] made improvements to Hoffmann’s method by brand new computer technology, got reference intervals comparatively consistent with those reported in literatures, preliminarily verified the theory that reference intervals more concordant with requirements could be obtained based on laboratory data with the help of statistical software. Dorizzi et al. [10] in Italy verified the feasibility of Katayev’s method again. Although Horowitz [9] pointed out differences between Katayev’s method and Hoffmann’s, one point was neglected that greater development had been made on computer technology compared with that in Hoffmann’s time. Restricted by application of computer technology, Hoffmann in those days could only discuss about the estimation model using 60 data to establish reference intervals. Now with the help of software, data of 10,000 of samples could be processed fast and easily. Statistically speaking, estimation made using large volume of samples could definitely produce much fewer errors than that made using small volume of samples, and also the results are more convinced. Therefore, Katayev’s method is certainly an improvement of Hoffmann’s method, and the so-called differences are actually a kind of improvement to the previous method.
Comparing the method of RI establishment in this article with Katayev’s method, further improvements are made. First, according to Katayev’s method, only visual observation was adopted to define the linear interval of the regression straight line, and the significance of R value which was the important determination coefficient of the regression straight line was not mentioned; in this article, the determination coefficient is required to meet R [ 0.99 when visual observation is adopted to determine the linear interval and to establish the model of regression straight line. Therefore, the advantage of the fuzzy logic theory of human being is taken and strict mathematical logic is added to make the results more convincing. Second, in Katayev’s method, the data from the abnormal thyroid function individuals cannot be ruled out, but in the optimized method, the step was done as first step. In Katayev’s method, Chauvenet’s criterion has been used, but in the optimized method, the 1/3 rule according to CLSI has been used. In Chauvenet’s criterion, when the total data are more than 1,000, ratio of maximum acceptable deviation to standard deviation (dmax/SD) is 3.48. In the study, SD is 8.98 mIU/L, mean value is 4.40 mIU/L. So the minimum eliminated data is 35.65 mIU/L, a lot of abnormal data cannot be ruled out such as 35.64 mIU/L and so on. In the large sample study, Standard Deviation of the data is always small, and many abnormal data cannot be extremely ruled out. In other side, the Chauvenet’s criterion cannot rule out the low value, for example, 0.01 mIU/L also has no chance to be ruled out. Compared to Chauvenet’s criterion, the 1/3 rule is recommended by CLSI, which can better rule out the abnormal data. In addition, TSH RIs are established for groups of different races, genders, age and source of sample according to relevant requirements to remedy the defect of previously reported methods. Of course, the optimized method also has some defects, to establish suitable linear regression model is difficult, especially though the easy computer software such as Origin8.0 software only set linear regression model mechanically. We must eliminate the unlinear range data step by step through subjective observation to get the R [ 0.99 linear model. So we can see the linear range is decreased every 5 %. The linear model which we get may be not the best model, but only the better model. There is no obviously difference between various TSH RIs based on laboratory data analysis and RIs (0.270–4.200 mIU/L) currently used in the laboratory. According to results of test, certain differences were found between the RI lower limits of Han males, Uygur males, out-patient males and used in the laboratory currently. If we consider only the attention the upper TSH reference limit, no significant difference between TSH RIs of various races, genders, age, sample populations and that currently
123
J Endocrinol Invest
used in the laboratory can be seen, which means that RIs provided by the manufacturer can be used for the entire population. So it also indicates that the improved method could be used for verification of TSH RIs currently used in the laboratory. Then it get a new RI expanded the current RI, which can be seen as the confidence interval (CI) of manufacturer RI. Why lower limit of some male RIs have statistical differences? First, the differences from the statistical error and the character of special male populations such as smoking status and body mass index (BMI) [13] and so on. Second, data collected in the laboratory, and samples of some unhealthy groups are included (only samples with TSH index-related abnormities are ruled out). The objective is to get the distribution curve of TSH test values of the entire population in natural state, and thus to establish relevant model according to the distribution curve. Compared with the large sample investigation, a narrow TSH RI rage will be offen defined based on strictly criteria selection small sample investigation [14]. The CLSI method is a strict criteria selection small sample investigation, and using the stored data is a large sample investigation. Therefore, RIs obtained by our method has a wider range than RIs obtained by adopting CLSI reference method. This should be the reason why there are differences between the lower limits of RIs of male groups with that of current RI in the laboratory. From the result of comparing the change in the abnormal TSH percent according to the new RI (CI of manufacturer RI) and the manufacturer RI, it shows that the change in RI upper limit, the abnormal TSH percent, easily changes. Because the upper limit is related to the subclinical hypothyroidism, most sub-clinical hypothyroidism occurs in female population. The female number is large in the entire TSH test population, so the upper limit RI is very important. If the upper limit is uncomfortable, the subclinical hypothyroidism will be omission diagnosis and excessive diagnosis. The manufacturer provides the RI to laboratory, but does not give comfortable confidence interval. So here, we using the Optimized method firstly verified the current RI, then see the new RI as the CI of current RI, if an individual test result is 4.200–4.979 mIU/ L, we can see it as ‘‘high risk sub-clinical hypothyroidism’’, then ask for patient follow-up. The benefit of doing these is to decrease the omission diagnosis and excessive diagnosis, and to reduce the patient’s psychological pressure.
Inference Regression straight line model and appropriate TSH reference intervals could be established using test data
123
collected from the laboratory as well as statistical analysis method. This method is also adopted to verify TSH RIs currently used in the laboratory. At present, result verification among laboratories is a trend of laboratory medicine development. TSH RIs should be set using the optimized statistical method by analyzing laboratory-stored data from multicenter laboratories with the same test conditions. It can provide help for result verification among local laboratories. Conflict of interest The authors Feng Yang-chun, Bian Wen-an, Mu Chao-dong, Xu Yi, Wang Fei, Qiao Wen-bin, and Huang Yanchun declare that they have no conflict of interest.
References 1. Surks MI, Ortiz E, Daniels GH et al (2004) Subclinical thyroid disease: scientific review and guidelines for diagnosis and management. J Am Med Assoc 291(2):228–238 2. Gietka-Czernel M, Jastrzebska H (2007) Thyrotropin reference range–should it be changed? Endokrynol Pol 58(5):454–560 3. Surks MI, Goswami G, Daniels GH (2005) The thyrotropin reference range should remain unchanged. J Clin Endocrinol Metab 90(9):5489–5496 4. Laurberg P, Andersen S, Carle´ A et al (2011) The TSH upper reference limit: where are we at? Nat Rev Endocrinol 7(4): 232–239 5. Schalin-Ja¨ntti C, Tanner P, Va¨lima¨ki MJ et al (2011) Serum TSH reference interval in healthy Finnish adults using the Abbott Architect 2000i Analyzer. Scand J Clin Lab Invest 71(4):344–349 6. Clinical and Laboratory Standards Institute (2008) Defining, establishing, and verifying reference intervals in the clinical laboratory; approved guideline, 3rd edn. Clinical and Laboratory Standards Institute, Wayne (CLSI document C28-A3) 7. Hoffmann RG (1963) Statistics in the practice of medicine. JAMA 185:864–873 8. Katayev A, Balciza C, David W et al (2010) Establishing reference intervals for clinical laboratory test results: is there a better way? Am J Clin Pathol 133(2):180–186 9. Horowitz GL (2010) Estimating reference intervals. Am J Clin Pathol 133(2):175–177 10. Dorizzi RM, Giannone G et al (2011) Indirect methods for TSH reference interval: at last fit for purpose? Am J Clin Pathol 135(1):167–174 11. Ricos C, Alvarez V, Cava F, et al. Desirable specifications for total error, imprecision, and bias, derived from biologic variation. Westgard QC Web page. http://www.westgard.com/biodatabase1. htm. Accessed 5 March 2008 12. Ilcol YO, Aslan D (2006) Use of total patient data for indirect estimation of reference intervals for 40 clinical chemical analytes in Turkey. Clin Chem Lab Med 44(7):867–876 13. Lucas A, Granada ML, Olaizola I et al (2013) Leptin and thyrotropin relationship is modulated by smoking status in euthyroid subjects [J]. Thyroid 23(8):964–970 14. Spencer CA, Hollowell JG, Kazarosyan M et al (2007) National health and nutrition examination survey III thyroid-stimulating hormone (TSH)-thyroperoxidase antibody relationships demonstrate that TSH upper reference limits may be skewed by occult thyroid dysfunction. J Clin Endocrinol Metab 92(11):4236–4240
