Sensors and Actuators B 173 (2012) 157–163

Contents lists available at SciVerse ScienceDirect

Sensors and Actuators B: Chemical

journal homepage: www.elsevier.com/locate/snb

Detection limits of chemical sensors: Applications and misapplications

Hans-Peter Loock a,∗ , Peter D. Wentzell b
a
Department of Chemistry, Queen’s University, Kingston, ON K7L 3N6, Canada
b
Department of Chemistry, Dalhousie University, Halifax, NS B3H 4J3, Canada

a r t i c l e i n f o a b s t r a c t

Article history: The limit of detection (LOD) and the sensitivity of a chemical sensor are defined using IUPAC guidelines.
Received 13 November 2011 The LOD from simulated and experimental data is calculated from a calibration curve using a simple
Received in revised form 21 June 2012 statistical model that was implemented into a spreadsheet program. This definition of the LOD is com-
Accepted 25 June 2012
pared with the commonly used definition of the LOD, which is based on the product of sensitivity and
Available online 4 July 2012
the theoretical instrument resolution.
Crown Copyright © 2012 Published by Elsevier B.V. All rights reserved.
Keywords:
Limit of detection
Sensitivity
Calibration curve
Confidence interval
Fiber optic sensors
Photoacoustic sensor
Absorption sensor
Refractive index sensor

1. Introduction a calculated detection limit that may be orders of magnitude below

the lowest measured value. Frequently, these detection limits are
The following instructions for “Xerox Enhanced Atomic incorrectly calculated by dividing the resolution of, typically, only
Microscopy (XEAM)”were once posted near a departmental pho- one of the instrumental components of the system with the sensi-
tocopier tivity, i.e. the slope of the calibration curve. This is akin to assuming
that all measurements fall almost exactly on the calibration curve
• Take a photocopy of a sheet of paper at 4-fold magnification. and the standard deviation of the signal from the sensor is much
• Now take another photocopy of the enlarged copy – again at 4- less than the discretization-limited instrumental resolution.
fold magnification. Some authors have justified their approach by referring to an
• Repeat the process 15 times for a total magnification of 1:415 or article on refractive index sensors by White and Fan [1], who stated
1:1.07 billion. – correctly, of course – that the detection limit can be obtained by
• At this point a 0.27 nm feature of the paper has been enlarged dividing the sensor resolution with its sensitivity
to fill the entire paper copy. The black smudge you see is an R
individual carbon atom (diameter about 0.22 nm)! xLOD = (1)
r
However, the resolution, R, is a quantity that needs to be
The resolving power in this XEAM experiment was calculated by obtained either by repeated measurements near the suspected LOD
assuming that one can obtain the ultimate spatial resolution (the or by statistical analysis of a calibration curve. One cannot simply
detection limit) by extrapolation from experiments at low mag- assume that the sensor resolution is identical to the discretization
nification assuming an error-free linear calibration with constant limit of one of the components of the system. While White and Fan
magnification (sensitivity). Many authors of articles that describe defined the “sensor resolution” of their particular refractive index
spectroscopic detection systems and chemical sensors use a similar sensor as being related to the “smallest possible spectral shift that
assumption. It is quite common to find in the literature a mea- can be accurately measured”, they also stated that this is not a quan-
sured sensitivity, which was obtained at high concentrations, and tity related solely to, e.g. the spectrometer’s spectral resolution,
but an experimentally determined quantity that also includes, e.g.
amplitude noise and temperature-induced noise. Unfortunately,
∗ Corresponding author. Tel.: +1 613 533 2621. the “resolution of the measurement” is a term that is used in widely
E-mail address: [email protected] (H.-P. Loock). different ways in optical engineering (e.g. “spectral resolution of a

0925-4005/$ – see front matter. Crown Copyright © 2012 Published by Elsevier B.V. All rights reserved.
http://dx.doi.org/10.1016/j.snb.2012.06.071
158 H.-P. Loock, P.D. Wentzell / Sensors and Actuators B 173 (2012) 157–163

spectrometer”) in physics (e.g. “vertical resolution of an oscillo- • The concentration at the detection limit can then be calculated
scope”) and in analytical chemistry, where it stands for “t·sy ”, i.e. from the sensitivity, r, i.e. the slope of the calibration curve.
the value of the Student t-function multiplied with the standard Assuming a linear calibration curve near the LOD, we calculate
error of the signal, which is obtained by averaging the signal near the minimum detectable concentration as
the LOD. In this article we provide a rigorously derived but user- tsy
friendly set of equations that permit the calculation of a detection xLOD = (3)
r
limit from a linear calibration curve. A spreadsheet program is pro-
vided in the electronic database of the journal that permits the entry Eq. (3) is identical to Eq. (1), since the sensor resolution at the LOD, is
of data pairs forming a calibration curve and that then calculates simply t˛,k−1 -fold larger than the measurement uncertainty, R = tsy .
the LOD. It is not the goal of this article to provide a survey of the For convenience the value of the student t-function is frequently
considerable research effort that went into statistical data analysis assumed to be t = 3, but this implies that a minimum of about 16
(see e.g. Refs. [2–5] for reviews), nor is it claimed that the analysis samples (8 blanks and 8 low-concentration samples) have been
below will be correct for any type of data set. Some limitations of analyzed. In the sensor literature it is frequently overlooked that
the model are given below. with a single measurement near the estimated LOD it is not possible
Two different methods will be reviewed that may be used to to determine the LOD, since one cannot determine the measure-
calculate a limit of detection. The more commonly used method ments’ standard deviation sy from such a single measurement.
involves the repeated measurement of the sensor’s response when It is important to note that the calculations above assume that
exposed to a blank solution, i.e. one without analyte, and a solu- the errors are normally (Gaussian) distributed, and that the error
tion containing the analyte at a concentration close to the LOD. The distributions of the blanks and the low-concentration measure-
second method describes how the LOD may be obtained from a lin- ments have an identical width.
ear calibration curve. This requires the determination of confidence
intervals and measurements of the sensor’s response at different 3. Method II: determination of the limit of detection using
concentrations including some near the LOD. The second method a calibration curve
is particularly useful when literature data need to be re-evaluated.
To illustrate the applications and limitations of both methods, Frequently it is necessary to compare the performance of one’s
the limits of detection are calculated using both simulated and own sensor system to a system that has been reported previously
experimentally obtained calibration curves. and may not have been characterized using the above method.
Assuming that the authors of the previous study provided a calibra-
tion curve, how can one estimate the LOD of their measurements?
2. Method I: determination of the limit of detection from If we were provided with n data pairs forming a linear calibration
standard deviations at low concentration curve {x, y}, we can calculate the sensitivity as the slope of a linear
fit
The following paragraphs provide a sketch of the accepted pro-   
cedures by analytical labs and are consistent with guidelines of the y n (xi yi ) − xi yi
r= = (4)
American Chemical Society [6]. The reader is referred to analytical x D
chemistry text books for more information on the derivations and The signal offset, i.e. the intercept of the calibration curve, is
for details on chemical quantitation. Here we follow Harris in his similarly calculated as
description of the LOD determination [7].    
n xi2 yi − (xi yi ) xi
b= (5)
D
• We first require repeated measurements of a blank sample, i.e.
one that does not contain analyte. The sensor response is repeat- Here, the determinant in the denominator is given by
 
edly measured. The American Chemical Society recommends k = 10  x2  xi    2
 i 
of such measurements [6], but many agencies only require 7 D= =n xi2 − xi (6)
measurements of the blank’s response. The mean value, ȳblank  xi n 
is obtained by averaging.
• We then prepare samples that contain analyte at a concentration The standard deviations of the sensitivity and offset and their
about 1–5 times higher than the expected LOD and again perform corresponding covariance (srb ) are calculated from

k measurements at this concentration. The mean value and the  xi2
n nx̄
standard deviation of the measurement, sy , can be determined. sr = sy ; sb = sy ; srb = −sy2 (7)
Ideally the samples should be prepared independently from each D D D
other and using different stock solutions. where the average standard deviation of the sensor response is
• The signal at the detection limit is then calculated from [8]. estimated from the standard error of the fit [10]

2
yLOD = ȳblank + t˛,k−1 sy (2) (yi − rxi − b)
sy = (8)
n−2
Here ȳblank is the average signal of the k measurements of the blank These equations are commonly found in textbooks on analytical
samples and t˛,k−1 is the ␣-quantile of Student’s t-function with quantitation and are readily incorporated into a spreadsheet pro-
k − 1 degrees of freedom where (1 − ˛) designates the required con- gram. An example is provided as an electronic supplement.
fidence level. For example, if it is required that the measurement Of importance in re-analysis of the previously published calibra-
at the LOD has a 99% probability of being larger than the blank, tion curves is the uncertainty of the concentration measurement,
then ˛ = 0.02, or (1 − ˛) = 0.98, owing to the two-sided nature of the sx . Following Harris [7] we can calculate the uncertainty, sx , at the
t-distribution. When 10 samples are analyzed (k = 10) one obtains concentration x by propagating uncertainties
t0.02,9 = 2.821, whereas t0.02,7 = 2.998 for k = 8. A critique of Eq. (2) is   
given by Mocak et al. [9]. Frequently, a less stringent threshold of sy 1 x2 n xi2 2x xi
sx = + + − (9)
(1 − ˛) = 0.95 may be appropriate [4]. |r| k D D D
 H.-P. Loock, P.D. Wentzell / Sensors and Actuators B 173 (2012) 157–163 159

120

100

80

Signal /mV
60

40
3sx

20

xLOD
xC
0
0 20 40 60 80 100 120 0 50 100 150 200 250
Concentration / ppm
Fig. 1. Two different simulated linear calibration curves showing the curves for a linear fit (red) and the 99% confidence intervals (blue) calculated as ±3sx . For each datum
the uncertainties of the signal (±sy ) and of the concentration (±sx ) were calculated using Eqs. (8) and (9) and are shown as error bars. The error bars are therefore not the
result of multiple measurements at a single concentration. The values for xLOD and yLOD are shown with black lines. (For interpretation of the references to color in this figure
legend, the reader is referred to the web version of this article.)

The last term in the root considers the covariance of the errors of the An equation similar to Eq. (11) was already presented as
slope and intercept [11]. Here, k designates the number of repeat part of the IUPAC Recommendations for the Nomenclature for the
measurements that are averaged to obtain a single data pair for the Presentation of Results of Chemical Analysis [4]. The factor of two in
calibration curve. If the literature curve does not provide error bars Eq. (4.19) of Ref. [4] and in Eqs. (10) and (11) arises from the require-
one may assume that only a single measurement was performed at ment to span the entire confidence interval of concentrations at the
a given concentration x and that k = 1. We define xC as the concen- yLOD , i.e. it takes into consideration that there are uncertainties in
tration which is larger than the uncertainty at this concentration the slope and intercept of the regression curve. We note that for
by a factor given by the student-t function, i.e. poorly correlated data sets (R2 < 0.7) the IUPAC expression gives a
   detection limit that may differ from that given by Eq. (11) and may
tsy 1 x2 n xi2 2xC xi even be negative. A comparison of Eq. (11) with Eq. (4.19) of Ref.
xC = tsx = + C + − (10) [4] is provided in Appendix.
|r| k D D D
A closely related method of obtaining the LOD by statistical anal-
After squaring both sides of Eq. (10) one finds the detection limit ysis of the calibration curve was presented by Hubaux and Vos [12],
from the root of the resulting quadratic equation as and is considered superior to the above statistical analysis if the
2tsy data set is clustered in particular ways. The authors of both articles
xLOD = 2xC = [4,12] pointed out that the concentration at the limit of detection
nt 2 sy2 − Dr 2
is no more than an estimate and may differ by more than a factor of

two – even for data sets producing very similar calibration curves
 D2 r 2  D
× tsy xi − + Dr 2 xi2 − n t 2 sy2 − Dt 2 sy2 (11) with identical strong correlation, R2 > 0.98. This may be verified
k k by generating numerous random data sets using the spreadsheet
that is available in Supplementary information. Since the limit of
Please consult Supplementary material in the journal’s data base detection is so strongly dependent on the actual distribution of the
for details on the derivation. The relation between xC , xLOD , and sampled data, there really is no point in giving the LOD with a large
sx is shown in Figs. 1 and 6. When we, again, set t = 3, the signal, number of significant figures.
yLOD = yblank + rxLOD /2 at this detection limit corresponds to concen- Finally, quoting Harris “it is not reliable to extrapolate any cal-
tration, xLOD , at which an average of k measurements has a 99% ibration curve, linear or non-linear, beyond the measured range of
chance of being larger than that of a blank sample. The practical standards.” On the other hand with the above method there are
meaning of the above calculation is apparent from a linear cali- some scenarios in which one may be able to justify an extrapola-
bration curve (Fig. 1a) which was generated using simulated and tion of the calibration curve to a detection limit that may be lower
slightly scattered data. The linear fit is quite good (R2 = 0.97), but than the lowest measured concentration. Fig. 1b shows a hypo-
the lines designating the confidence intervals nevertheless indicate thetical linear calibration curve with R2 = 0.95. Here, the LOD that
that the lowest concentrations which are outside the confidence was estimated using Eq. (11) is somewhat lower (80 ppm) than the
interval of a blank sample are higher than 40 ppm. This means that lowest measured concentration (100 ppm). In this case one may be
the lowest measured concentrations lie much below the detection excused for violating Harris’ rule and give a detection below the
limit. lowest measured concentration.
160 H.-P. Loock, P.D. Wentzell / Sensors and Actuators B 173 (2012) 157–163

0.25

0.20

Absorbance, ε C d
0.15

0.10

0.05

0.00
0 20 40 60 80 100 120
Concentration / ppm

Fig. 3. A calibration curve that was obtained by measuring the ring-down time of
a fiber loop cavity that contained a liquid absorption cell (see Andachi et al. for
details [13]). A detection limit xLOD = 42 ppm is calculated using Eq. (11). When the
measurements at 100 ppm and 80 ppm are omitted one obtains xLOD = 22 ppm. The
error bars are defined as in Fig. 1 and differ from those in Ref. [13].

the upper trace of Fig. 2 shows a signal with a peak that is 3sy
higher than the baseline noise (sy = 0.5), i.e. the signal at the LOD is
Fig. 2. Simulated data set with the signal sampled at two different resolutions. The yLOD = 1.5. For the lower curve, recorded at low resolution the sig-
upper trace shows a peak (solid line) and simulated data with a standard deviation nal LOD, yLOD = 3, is reasonably calculated from the signal increment
of 0.5 and sampled at 10 bit resolution. The data points in the lower panel assume
(y = 1) and not from the lower standard deviation of the baseline
that the same data was sampled at 4 bit resolution.
signal, sy = 0.5. While one can correctly identify discretization of the
sensor signal as providing a lower limit of detection, it is not cor-
In the above analysis it is assumed that (a) the errors are nor- rect to assume that the LOD will improve by a factor of 256 when
mally distributed, (b) the concentrations, x, of the {x, y} data pairs sampling at 10-bit resolution instead of 4-bit resolution. In fact the
are known exactly and (c) that the standard deviation of the signal LOD improves only by a factor of two. This illustrates that even in
is the same for all concentration measurements. As was pointed the case of a discretization-limited measurement the LOD is rarely
out by Lavagnini and Magno [2] the last assumption is, in fact, linearly correlated to the instrumental resolution of the acquisition
not correct for many calibration curves and their article describes system – in contrast to how Eq. (1) is frequently interpreted.
cases in which the measurement uncertainty is dependent on the
concentration.
Of course, it is also assumed that the calibration curve is linear
and, in particular, that the sensitivity near the LOD is identical to 5. Application to experimental data
that of the entire calibration curve. This condition is also not always
fulfilled, since sensor measurements may saturate at high concen- In the following we use previously published calibration curves
trations and there may be an induction range or hysteresis at low of three different chemical sensors to demonstrate practical appli-
concentrations. cations as well as the limitations of Method I and Method II in
determining the detection limit.
4. Method III: determination of the limit of detection from Fig. 3 shows a calibration curve obtained by Andachi et al. for the
the instrumental resolution limit determination of the dye methylene blue in a fiber-loop cavity ring-
down experiment [13]. Light from a gain-switched 660 nm laser
Is the ratio of instrument resolution and sensitivity then a mean- diode was coupled into a 4.69 m multimode fiber loop (125/50 ␮m)
ingless quantity? In cases of low-resolution measurements near the using a 99:1 fiber:fiber coupler. A custom-made absorption cell
LOD, the frequently quoted product of the inverse of sensitivity and with a length of 100 ␮m was inserted into the loop and con-
the smallest measurable signal increment can indeed be a quantity tained methylene blue dye in methanol at different concentrations.
that is useful for comparison. For example, if the detection limit has The ring-down times of 4096 light pulses were averaged and,
been determined using the required number of blanks and samples together with the ring-down time of the used fiber cavity contain-
one may realize that the standard deviation of the measurement, ing the blank solution, t0 , used to determine the absorption term
sy , is lower than the increment at which the signal can be sampled εCd shown in Fig. 3. Details of the experiment are given in Ref.
(y). In this case the signal at the LOD, yLOD , is, indeed, limited by [13]. Using the calibration curve and Eq. (11) a limit of detection
the resolution at which the signal can be sampled, yLOD = 3y, and xLOD = 42 ppm is obtained. When the values at the highest con-
discretization determines the sensors’ limit of detection. centrations are omitted, as was done in Ref. [13], we calculate
For example, in Fig. 2 the same simulated signal peak was sam- xLOD = 22 ppm. The latter value is slightly higher than the LOD given
pled at 4-bit and at 10-bit resolution. At high sensor resolution by Andachi et al. (xLOD = 20 ppm).
 H.-P. Loock, P.D. Wentzell / Sensors and Actuators B 173 (2012) 157–163 161

This calculation may be contrasted to one that is sometimes

found in the cavity ring-down literature. Cavity ring-down times
are related to optical absorption of a sample in the cavity through
nL 0 − 
˛() = (12)
c0 d 0 
and the minimal detectable absorption loss – or the detection limit
– is calculated as [14,15]
nL min
˛min () = (13)
c0 d  2
0

The parameters in Andachi’s experiment were

L = 4.69 m, d = 100 ␮m, n = 1.46, and the empty cavity ring-
down time t0 = 45 ns. With these parameters we obtain
˛min (v) = 1.13 × 1011 m−1 s−1 min . Generally, min is not given
by the bandwidths of the data acquisition card, the photodetector
and the oscilloscope, but has to be experimentally determined.
Fig. 4. Response of the resonant wavelength of a functionalized silicon-on-insulator
In this case the minimum measurable change of ring-down time, microring resonator to m-xylene vapor (adapted from Loock et al. [16]). The micro-
min , is determined from xLOD = 22 ppm (50 ␮M) and the absorp- ring resonator was coated with a film of a PDMS/PDPS co-polymer which increased
tion coefficient of methylene blue, ε = 1.5 × 105 L mol−1 cm−1 . Their its refractive index upon exposure to m-xylene and other aromatic volatile organic
product is ˛min (v) = 7.9 cm−1 and consequently min = 7 ns. If compounds. The refractive index change increases the effective index of the prop-
agated mode and causes a wavelength shift. The detection limit is determined as
one assumes, incorrectly, that the minimal resolvable change in
xLOD = 325 ppm. The error bars are defined as in Fig. 1.
ring-down time min is solely determined by the bandwidth of
the detection system – here about 30 GHz from the bandwidth
of the oscilloscope and the rise time of the photodetector –
the minimum detectable absorption loss would be calculated
as ˛min (v) = 0.03 cm−1 and xLOD = 90 ppb. This latter calculation
ignores fluctuations in the sample concentrations, the shape of the
ring-down transient, intensity noise arising from electronics or
optical alignment, and ultimately laser shot noise. Of course, the
minimum detectable change in ring-down time is also not simply
given by the fitting error of an experimental intensity decay curve
to an exponential function. This fitting error only indicates how
well each of the experimental decay waveforms can be described
by an exponential function.
As a second example we present a very different chemical
sensor for volatile organic compounds was recently described
by one of the authors and his collaborators [16]. A silicon-
on-insulator (SOI) micro-ring resonator was coated with poly-
dimethyl/polydiphenylsiloxane (PDMS/PDPS) co-polymer and the
response to m-xylene vapor at various concentrations was deter-
mined from the shift of the SOI ring cavity resonance wavelengths.
As m-xylene partitions into the polymer coating, the refractive
index of the polymer increases and the effective index of the propa-
gated mode consequently also increases. From the calibration curve
shown in Fig. 4 and using Eq. (11) the detection limit is determined
as xLOD = 325 ppm.
The actual detection limit may be somewhat higher since the
calibration curve does not appear to be linear at very low concen-
trations. Close inspection of Fig. 4 reveals that at low concentrations
of m-xylene the refractive index of the polymer does not appear to
increase. This may be attributed to the combined effect of swelling
of the polymer, which decreases the refractive index, and increase
of the xylene volume fraction, which increases the refractive index.
This “induction range” leads to an underestimate of the detection
limit.
As a third example Fig. 5 shows the calibration curve presented
Fig. 5. Photoacoustic signal as a function of coumarin 440 dye concentration
by Yang et al. [17] for the determination of coumarin dye concentra- recorded using a fiber Bragg grating as a transducer. The top figure shows the data
tions using photoacoustic measurements with a fiber Bragg grating displayed in Fig. 7 of Yang et al. [17] on a logarithmic scale and assumes that the
as a transducer. The data are displayed both on a regular and on a errors of the both concentration and photoacoustic signal are scaled logarithmically.
logarithmic scale. The respective uncertainties, which were calcu- While xLOD cannot be determined with this assumption, the width of the confidence
interval spans 2.5 orders of magnitude in concentration. The lower panel shows the
lated using Eqs. (8) and (9) were scaled accordingly. It is clearly
same data but assumes that the photoacoustic signal measurement errors are con-
inappropriate to assume that the standard deviation of the pho- stant over the entire measurement range. The detection limit is then determined as
toacoustic signal is scaled logarithmically throughout the entire xLOD = 1.3 mM. The error bars are defined as in Fig. 1.
concentration range as this would imply that the uncertainty is zero
for the blank solution. On the other hand the measurement error is
162 H.-P. Loock, P.D. Wentzell / Sensors and Actuators B 173 (2012) 157–163

not constant, either, since xLOD calculated from the linearly scaled Supplementary material deposited in the journal’s database
calibration curve gives xLOD = 1.3 mM, i.e. a concentration that was gives a detailed derivation of Eqs. (11) and (14). When comparing
shown by Yang et al. to be 500-fold higher than what could be deter- to the detection limit defined by Eq. (11)
mined by multiple measurements near the LOD, i.e. using Method
2tsy
I presented above. The standard deviation of a series of measure- xLOD,1,2 =
ments at 1 ␮M, 2 ␮M, 5 ␮M, and 10 ␮M showed that a 5 ␮M solution nt 2 sy2 − Dr 2
produced a signal that was 3sy higher than the background noise 

[17] (see Eq. (3)). With this example we intend to illustrate that
 D2 r 2  D
× tsy xi − + Dr 2 xi2 − n t 2 sy2 − Dt 2 sy2 (11)
the method of determining a limit of detection using a calibration k k
curve cannot be relied upon when the measurement uncertainties
are not constant over the range of measurements. A comprehen- it is found that the expressions differ by the last two terms in the
sive treatment of these cases has been presented by Lavagnini and root. Also, in the derivation of Eq. (11) k is retained as a parameter,
Magno [2]. whereas Currie and Svehla implied it to be k = 1. Eqs. (15) and (11)
give the same result when
6. Summary
Dr 2   n 
+ r2 xi2  t 2 sy2 (16)
In conclusion, detection limits of chemical sensors may be k k+1
obtained in two different ways: This is very commonly the case, but differences are found when
the data are poorly correlated, or, more specifically, when the prod-
(1) As recommended by IUPAC and most analytical labs one should uct of t with the standard deviation of the sensor response sy is
determine the average signal level when repeatedly measuring large compared to the sensitivity of the measurement, r. Currie and
a blank sample, as well as the average signal and its standard Svehla also remarked that the LOD can no longer be defined by the
deviation of samples at one, or ideally more, concentrations IUPAC definition if the uncertainty of the slope (=sensitivity) sr is
near the LOD. t-fold larger than the slope r. In any case it is important to remem-
(2) When re-analyzing previously published and incompletely ber that the LOD is “the maximum null-signal upper limit for a
characterized data, the concentration at the LOD may also be particular realization of the calibration curve” [4], and that another
estimated from the calibration curve using Eq. (11) or the equa- calibration curve of the same sensor with similarly scattered data
tions presented in Ref. [4]. This can only be accurate when can provide a very different estimate.
a large number of calibration measurements have been per- From a qualitative perspective, the differences between the
formed in the concentration regime near the LOD. Of course, two approaches are subtle, but can be traced to two sources.
Eq. (11) assumes a calibration that is linear over the entire First, Eq. (15) (Currie and Svehla) incorporates the uncertainty
measurement range and also assumes that the systematic mea- in the estimated blank into the signal detection limit, whereas
surement error is negligible or has been corrected. It also needs Eq. (11) does not. This results in different critical concentra-
to be assumed that the uncertainties of the concentrations and tion limits (xC = (yLOD − b)/r) for the two methods Both methods
of the signal are similar for all measurements, that they are use the calibration data to determine the LOD as the con-
normally (Gaussian) distributed, and that the calibration curve centration where the probability that (x < xC ) is less than ˛.
shows no hysteresis, induction range, or saturation. In any case, The second difference is that Eq. (11) assumes that the uncer-
the LOD calculated from the calibration curve cannot be more tainty in the estimated x does not change from that at xC ,
than an estimate and another calibration with similarly dis- whereas Eq. (15) does not make this assumption. In virtually
tributed data can produce an LOD that may be different by a all practical situations, these differences are inconsequential,
factor of two or more. Given the many assumptions that go into especially in view of other assumptions that are likely to be
Method II it is better to obtain the LOD by repeated measure- violated.
ments near the suspected LOD, i.e. using by Method I. [18] We emphasize that no claim is made to the superiority of Eq. (11)
over existing definitions. It is introduced only because the deriva-
Acknowledgements tion from Eq. (10) is straightforward, there is an obvious graphical
interpretation (Figs. 1 and 6), and its derivation does not require
The authors thank the referees sincerely for helpful comments the introduction of more parameters. The spreadsheet program
and advice on the manuscript. We also acknowledge financial sup- associated with this article provides Eqs. (11) and (15).
port by the Natural Sciences and Engineering Research Council
(NSERC) of Canada.
Appendix II. Differences between Methods I and II

Appendix I. Relation of Eq. (11) to Eq. (4.19) by Currie and While there are only small differences between the LOD
Svehla (Ref. [4]) calculated from Eq. (11) and Eq. (15) there is an approximately
2-fold difference to an LOD calculated using Method I, i.e. when the
In Eq. (4.19) of Ref. [4] Currie and Svehla give the detection limit standard deviation of the signal, sy , is measured near the LOD. Fig. 6
(“the estimate for the minimum detectable quantity”) as is an illustration that shows measurements of 8 replicate blanks
     
 1− xi / 2n xi sb / sb2 + sy2 t(sr /r)
(x = 0) and 8 replicate standards of the same concentration (x = “10”
2t sb2 + sy2 in arbitrary units). The simulation assumed a slope of r = 10, an
xLOD =
r 1 − t 2 (sr2 /r 2 ) intercept of 10, and a measurement error standard deviation of
(14) sy = 13. For simplicity, t was taken to be 3.
The black line shows the LOD1 as calculated from Method I,
This can be rearranged using Eqs. (6) and (7) to give i.e. using Eq. (3). It considers the regression line to be error-free
    and sets the signal limit of detection based on the estimated
2tsy
xLOD = tsy xi − r 2 D2 + Dr 2 xi2 (15) uncertainty of the signal, sy , which in this case is the same whether
nt 2 sy2 − Dr 2 one uses the pooled replicates or regression standard error.
 H.-P. Loock, P.D. Wentzell / Sensors and Actuators B 173 (2012) 157–163 163

120
Type II error

100 tsx,LOD

Intensity, y / arb. units

80
Type I
error
60

tsx,C
40

tsb
20
tsy

XC: Method II
LOD1 XC: Ref 4 LOD2 LOD3
0
tsx,C

-20
-2 0 2 4 6 8 10 12
Concentration, x / arb. units
Fig. 6. Simulated data to illustrate the difference between limits of detection (LOD) obtained using Method I (LOD1 ) and Method II (LOD2 ). The LOD calculated according to
Ref. [4] is given as LOD3 .

The LOD1 = 4.05 is then determined directly from the regression [10] If you had an independent (and better) estimate of sy , you could use also that
line. value instead of Eq. (8). We use the value from a fit only for convenience. Also,
note that the degrees of freedom for t depend on how sy was determined and
In deriving Eq. (11) it was assumed that LOD2 is the concentra- does not need to be (n − 2).
tion at 2 tsx (here, LOD2 = 8.36 and t = 3) where sx is the estimated [11] C. Salter, Error analysis using the variance–covariance matrix, Journal of Chem-
uncertainty of x at the concentration xLOD . The relation is shown ical Education 77 (2000) 1239–1243.
[12] A. Hubaux, G. Vos, Decision and detection limits for linear calibration curves,
by the solid blue line. This procedure implicitly incorporates the Analytical Chemistry 42 (1970) 849–850.
uncertainty of the linear regression parameters. Note that the crit- [13] M. Andachi, T. Nakayama, M. Kawasaki, S. Kurokawa, H.-P. Loock, Fiber-optic
ical concentration (xC ) for Method II is close to LOD1 , but it is not ring-down spectroscopy using a tunable picosecond gain-switched diode laser,
Applied Physics B: Lasers and Optics 88 (2007) 131–135.
identical since the latter uses the estimated uncertainty in y and
[14] P. Zalicki, R.N. Zare, Cavity ring-down spectroscopy for quantitative absorption-
the former uses the confidence bounds for estimated x. The value measurements, Journal of Chemical Physics 102 (1995) 2708–2717.
LOD3 = 8.53 was calculated according to Ref. [4] and is only slightly [15] G. Berden, R. Engeln (Eds.), Cavity Ring-Down Spectroscopy: Techniques and
Applications, Wiley-Blackwell, 2009.
higher than (red line) LOD2 .
[16] H.-P. Loock, J.A. Barnes, K. Bescherer, A. Brzezinski, G. Gagliardi, A. Gribble, S.
Janz, R. Ma, D. Munzke, G. Ongo, J. Saunders, H. Waechter, D.-X. Xu, Chemical
microanalysis with cavity-enhanced optical waveguide devices, in: Proceed-
Appendix C. Supplementary data ings of the SPIE, 2011.
[17] Q.X. Yang, H.-P. Loock, I. Kozina, D. Pedersen, Fiber Bragg grating photoacoustic
Supplementary data associated with this article can be found, in detector for liquid chromatography, Analyst 133 (2008) 1567–1572.
[18] L.A. Currie, The many dimensions of detection in chemical analysis, with spe-
the online version, at http://dx.doi.org/10.1016/j.snb.2012.06.071. cial emphasis on the one-dimensional calibration curve, in: D. Kurtz (Ed.),
ACS Symposium Series 284: Chemometrics in Pesticide/Environmental Residue
Analytical Determination, American Chemical Society, 1984, p. 49.
References
Biographies
[1] I.M. White, X.D. Fan, On the performance quantification of resonant refractive
index sensors, Optics Express 16 (2008) 1020–1028.
[2] I. Lavagnini, F. Magno, A statistical overview on univariate calibration, inverse
Hans-Peter Loock is a Professor of Chemistry with an interest in the development of
regression, and detection limits: application to gas chromatography/mass spec-
fiber-optic and micro-photonic sensors for absorption, refractive index, strain, and
trometry technique, Mass Spectrometry Reviews 26 (2007) 1–18.
fluorescence. His group specializes on the chemical detection in very small sample
[3] R. Ferrus, M.R. Egea, Limit of discrimination, limit of detection and sensitivity
volumes. Together with his group and with collaborators he has developed optical
in analytical systems, Analytica Chimica Acta 287 (1994) 119–145.
detection methods that allow for chemical quantitation but do not require accurate
[4] L.A. Currie, G. Svehla, Nomenclature for the presentation of results of chemical-
intensity measurements. He has obtained an Engineering Chemistry degree from
analysis, Pure and Applied Chemistry 66 (1994) 595–608.
the Technical University of Darmstadt (Germany) in 1991 and a Ph.D. in physical
[5] G.L. Long, J.D. Winefordner, Limit of detection, Analytical Chemistry 55 (1983)
chemistry from the University of Victoria, Canada. After two years of postdoctoral
A712–A720.
work at the Steacie Institute for Molecular Sciences (NRC, Ottawa) he joined Queen’s
[6] Guidelines for data acquisition and data quality evaluation in environmental
University in 1999.
chemistry, Analytical Chemistry 52 (1980) 2242–2249.
[7] D.C. Harris, Quantitative Chemical Analysis, 6th ed., W.H. Freeman, 2002. Peter Wentzell is a Professor of Chemistry at Dalhousie University, where he
[8] I. Janiga, J. Mocak, I. Garaj, Comparison of minimum detectable concentra- has been since 1989. He obtained his PhD from Michigan State University and
tion with the IUPAC detection limit, Measurement Science Review 8 (2008) held an NSERC Post-doctoral Fellowship at the University of British Columbia.
108–110. His principal research interests are in the area of chemometrics, with a focus
[9] J. Mocak, A.M. Bond, S. Mitchell, G. Scollary, A statistical overview of standard on the role of analytical measurement errors in multivariate measurements and
(IUPAC and ACS) and new procedures for determining the limits of detec- on the development of data analysis strategies for high-throughput biological
tion and quantification: application to voltammetric and stripping techniques methods.
(technical report), Pure and Applied Chemistry 69 (1997) 297–328.