ArticlePDF Available

Diagnosing Insulin Resistance in the General Population

Authors:

Abstract and Figures

Difficulties in measuring insulin sensitivity prevent the identification of insulin-resistant individuals in the general population. Therefore, we compared fasting insulin, homeostasis model assessment (HOMA), insulin-to-glucose ratio, Bennett index, and a score based on weighted combinations of fasting insulin, BMI, and fasting triglycerides with the euglycemic insulin clamp to determine the most appropriate method for assessing insulin resistance in the general population. Family history of diabetes, BMI, blood pressure, waist and hip circumference, fasting lipids, glucose, insulin, liver enzymes, and insulin sensitivity index (ISI) using the euglycemic insulin clamp were obtained for 178 normoglycemic individuals aged 25-68 years. Product-moment correlations were used to examine the association between ISI and various surrogate measurements of insulin sensitivity. Regression models were used to devise weights for each variable and to identify cutoff points for individual components of the score. A bootstrap procedure was used to identify the most useful predictors of ISI. Correlation coefficients between ISI and fasting insulin, HOMA, insulin-to-glucose ratio, and the Bennett index were similar in magnitude. The variables that best predicted insulin sensitivity were fasting insulin and fasting triglycerides. The use of a score based on Mffm/I = exp[2.63 - 0.28ln(insulin) - 0.31ln(TAG)] rather than the use of fasting insulin alone resulted in a higher sensitivity and a maintained specificity when predicting insulin sensitivity. A weighted combination of two routine laboratory measurements, i.e., fasting insulin and triglycerides, provides a simple means of screening for insulin resistance in the general population.
Content may be subject to copyright.
Diagnosing Insulin Resistance in the
General Population
KIRSTEN A. MCAULEY,
MBCHB
1
SHEILA M. WILLIAMS,
BSC(HONS)
2
JIM I. MANN,
DM
,
PHD
,
FRACP
1
ROBERT J. WALKER,
MD
,
FRACP
3
NICK J. LEWIS-BARNED,
FRACP
4
LARA A. TEMPLE,
BSC
1
ASHLEY W. DUNCAN,
BSC
,
NZCS
1
OBJECTIVE Difficulties in measuring insulin sensitivity prevent the identification of in-
sulin-resistant individuals in the general population. Therefore, we compared fasting insulin,
homeostasis model assessment (HOMA), insulin-to-glucose ratio, Bennett index, and a score
based on weighted combinations of fasting insulin, BMI, and fasting triglycerides with the
euglycemic insulin clamp to determine the most appropriate method for assessing insulin resis-
tance in the general population.
RESEARCH DESIGN AND METHODS Family history of diabetes, BMI, blood pres-
sure, waist and hip circumference, fasting lipids, glucose, insulin, liver enzymes, and insulin
sensitivity index (ISI) using the euglycemic insulin clamp were obtained for 178 normoglycemic
individuals aged 25–68 years. Product-moment correlations were used to examine the associa-
tion between ISI and various surrogate measurements of insulin sensitivity. Regression models
were used to devise weights for each variable and to identify cutoff points for individual com-
ponents of the score. A bootstrap procedure was used to identify the most useful predictors of ISI.
RESULTS — Correlation coefficients between ISI and fasting insulin, HOMA, insulin-to-
glucose ratio, and the Bennett index were similar in magnitude. The variables that best predicted
insulin sensitivity were fasting insulin and fasting triglycerides. The use of a score based on
Mffm/I exp[2.63 0.28ln(insulin) 0.31ln(TAG)]
rather than the use of fasting insulin alone resulted in a higher sensitivity and a maintained
specificity when predicting insulin sensitivity.
CONCLUSIONS A weighted combination of two routine laboratory measurements, i.e.,
fasting insulin and triglycerides, provides a simple means of screening for insulin resistance in
the general population.
Diabetes Care 24:460464, 2001
I
nsulin resistance is an important risk
factor for type 2 diabetes and cardio-
vascular disease (1). There is increasing
evidence supporting the fact that by the
time glucose tolerance or fasting glucose
levels become impaired, appreciable
-cell destruction may have already oc-
curred (2). Thus, it seems likely that at-
tempts to prevent type 2 diabetes will be
more successful if intervention is com-
menced when blood glucose levels are
still in the normal range. Therefore, a sim-
ple test for identifying insulin-resistant
individuals is important both for popula-
tion-based research and clinical practice.
The euglycemic insulin clamp and the in-
travenous glucose tolerance test (IVGTT)
are standard methods for the measure-
ment of insulin resistance in research, but
they are impractical in clinical practice
and are difficult to perform in population-
based research studies (3). Fasting insu-
lin, homeostasis model assessment
(HOMA), insulin-to-glucose ratio, and
the Bennett index are all used to predict
insulin sensitivity; and several other indi-
vidual variables, such as family history of
diabetes, BMI, blood pressure (BP), waist
and hip circumference, fasting triglycer-
ides, HDL, glucose, insulin, and hepatic
enzymes, are known to correlate with in-
sulin resistance (4–7). Combinations of
variables used to predict insulin resis-
tance have been assessed in a small num-
ber of studies, and most studies have
assessed prediction in individuals with
impaired glucose tolerance (IGT) and di-
abetes. Few studies have specifically eval-
uated the prediction of insulin resistance
in a significant number of individuals
with normal glucose tolerance (4,8). We
have compared the standard techniques,
several individual variables, and a score
based on a weighted combination of se-
lected variables with the euglycemic insu-
lin clamp to evaluate the best method of
predicting insulin resistance in normogly-
cemic individuals.
RESEARCH DESIGN AND
METHODS Participants who previ-
ously volunteered for various research
projects were recruited. A total of 178
normoglycemic men and women aged
25–68 years, similar to the general popu-
lation of New Zealand with respect to BP,
BMI, and waist-to-hip ratio (WHR) (9),
gave informed consent for a euglycemic
insulin clamp to be performed (10).
After a 10-h overnight fast, each par-
ticipant’s weight, height, and BP were
measured and recorded. Intravenous can-
nulae were inserted into the cubital vein
for the administration of insulin and glu-
cose (25% dextrose) and into the dorsal
aspect of the hand for arterialized sam-
pling. Basal samples were obtained for the
measurement of fasting insulin and lipid
●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●
From the
1
Department of Human Nutrition,
2
Department of Preventive and Social Medicine, and
3
Depart-
ment of Medicine, Otago University, Dunedin, New Zealand; and the
4
Department of Diabetes and Endo-
crinology, Gloucestershire Royal Hospital, Gloucester, U.K.
Address correspondence and reprint requests to Dr. Kirsten McAuley, Department of Human Nutrition,
Otago University, P.O. Box 56, Dunedin, New Zealand. E-mail: kirsten.mcauley@stonebow.otago.ac.nz.
Received for publication 25 July 2000 and accepted in revised form 9 November 2000.
N.J.L.-B. has received grants from the National Heart Foundation of New Zealand and the Laurensen Trust
(Otago Medical Foundation).
Abbreviations: AST, aspartate aminotransferase; BP, blood pressure; HOMA, homeostasis model assess-
ment; IGT, impaired glucose tolerance; ISI, insulin sensitivity index; IVGTT, intravenous glucose tolerance
test; M, glucose disposal rate; Mbw/I, ISI corrected for total body weight divided by average insulin; Mffm/I,
ISI corrected for fat-free mass divided by average insulin; PSEP, prognostic separation index; TAG, triglyc-
erides; WHR, waist-to-hip ratio.
A table elsewhere in this issue shows conventional and Syste`me International (SI) units and conversion
factors for many substances.
Epidemiology/Health Services/Psychosocial Research
ORIGINAL ARTICLE
460 DIABETES CARE,VOLUME 24, NUMBER 3, MARCH 2001
profile and for glucose and liver function
tests. Insulin (Actrapid) was infused at 40
mU m
–2
min
–1
to achieve hyperinsulin-
emia. Arterialized samples, achieved us-
ing the heated-hand technique, were
taken from the dorsum of the hand every
10 min for immediate glucose measure-
ments using a Yellow Springs Instruments
Sidekick Glucose Analyzer (calibrated be-
fore and during the test). A variable rate
glucose infusion was given for 115 min
and adjusted every 10 min according to a
negative feedback algorithm used in the
Otago Clamp Method (11). Blood glucose
levels were maintained as close as possible
to 4.5 mmol/l. Plasma insulin levels were
measured at 0, 60, 90, and 120 min. The
glucose disposal rate (M) (milligrams per
kilogram per minute) was calculated from
measurements taken during the final 60
min of the clamp. The ISI using total body
weight (Mbw/I) was calculated by divid-
ing the average Mby the average plasma
insulin concentration over the final 60
min (Mper milliunit per liter). The ISI
corrected for fat-free mass (Mffm/I) was
also calculated (12).
HOMA was calculated as described
by Matthews et al. (5). The Bennett index
was calculated by 1/ln(glucose 0)ln(insu-
lin 0) (13).
Plasma insulin was determined using
the Coat-A-Count
125
I radioimmunoas-
say (Diagnostics Products, Los Angeles,
CA). This is a polyclonal assay with a cross-
reactivity with proinsulin at the midcurve of
40%. The interassay coefficient of varia-
tion is 10%, and the detection limit is
1.2–1.5 mU/ml. Cholesterol concentra-
tion in plasma and lipoprotein fractions
was measured enzymatically with Boehr-
inger kits and calibrators. Triglyceride
and liver function tests were measured
enzymatically using Roche kits and re-
agents on a Cobas Fara analyzer. HDL
cholesterol was measured in the superna-
tant after precipitation of apolipoprotein
B– containing lipoproteins with phospho-
tungstate/magnesium chloride solution.
BP was measured in a sitting position
after a 10-min rest. Waist girth was mea-
sured as the minimum circumference be-
tween iliac crest and rib cage, and hip
girth was measured at the maximum
width over the greater trochanters (avail-
able for 99 participants). WHR was cal-
culated from these measurements. BMI
(kg/m
2
) was also calculated. A positive
family history of type 2 diabetes was de-
fined as having a first-degree relative with
diabetes diagnosed after 30 years of age
and not requiring insulin during the first
6 months from diagnosis (available for
101 participants).
Statistical analysis
Product-moment correlations were used
to examine the association between ISI and
fasting insulin, HOMA, insulin-to-glucose
ratio, Bennett index, and other individual
variables previously listed. A log transfor-
mation was applied to all of the variables
apart from age, family history, and BP.
An ISI of 6.3 MmU
–1
l
–1
defined
individuals with insulin resistance. This
corresponded to the lowest quartile for
the lean population (BMI 27 kg/m
2
).
Two approaches were used to select the
variables to be included in the risk score.
The first approach involved regression
analysis with backward elimination,
which estimated the association between
ISI and explanatory variables. This
method started with all explanatory vari-
ables in the model and progressively de-
leted those not meeting prespecified
statistical criteria. The second approach
used a bootstrap procedure to reveal the
variables that were most important. This
involved drawing 500 samples of the
same size from the original sample, with
subsequent replacement (14). Regression
analysis with backward elimination was
repeated for each sample. The most fre-
quently selected variables were regarded
as the most important.
Three possible scoring systems were
evaluated. In the first, the log-transformed
values of Mffm/I and the predictor vari-
ables were analyzed as continuous vari-
ables. By regarding a predicted value
6.3 MmU
–1
l
–1
, obtained from re-
gressing Mffm/I on the other variables sin-
gly or in combination, it was possible to
evaluate sensitivity and specificity of the
scoring test and, when used singly, to
identify cutoff points for the variables.
In the second scoring system, Mffm/I
was divided into two groups, with values
6.3 MmU
–1
l
–1
constituting a diag-
nosis of insulin resistance. In this case, we
used logistic regression, which estimates
the probability of a diagnosis. Prediction
probabilities 0.5 were taken as positive,
so that the sensitivity and specificity
could be calculated. Again, it was possible
to identify cutoff points when the vari-
ables were used one at a time. The third
model divided each predictor into two
groups using the cutoff points identified
in the other two models.
It is well known that scoring systems
established in one set of data may not pre-
dict as well in a new sample; therefore,
using the selected variables, another boot-
strap procedure was used to carry out in-
ternal validation (15,16). The median
values of the sensitivity and specificity,
rather than those obtained using only the
original sample, are more realistic predic-
tors of how the test will perform on a new
set of data. The 2.5 and 97.5 percentiles
were used as CIs.
Finally, values for the prognostic sep-
Table 1—Clinical and metabolic descriptors of the study population
nGeometric mean SD Range
Age (years) 178 46.8 7.8 25–68
Weight (kg) 178 76.1 17.3 48–149
BMI (kg/m
2
)178 27.5 5.3 18.4–50.6
Waist (cm)
Women 68 95.8 14.5 69.0–132.5
Men 31 104.2 12.7 72.8–130.6
WHR
Women 68 0.86 0.06 0.71–1.02
Men 31 0.97 0.05 0.84–1.06
BP 177 125/81 17/10 90/58–208/110
TAG (mmol/l) 178 1.47 0.73 0.44–5.50
HDL (mmol/l) 178 1.31 0.37 0.55–2.69
Insulin (mU/l) 178 10.0 9.5 2–73
AST (U/l) 73 14.3 6.5 9–56
Mbw/I 178 4.3 2.4 1.2–18.6
Mffm/I 178 6.7 3.4 2.3–25.9
HOMA 178 2.1 2.2 0.3–15.9
Insulin-to-glucose ratio 178 2.1 1.9 0.4–14.9
Bennett index 178 0.29 0.1 0.1–1.1
MCAuley and Associates
DIABETES CARE,VOLUME 24, NUMBER 3, MARCH 2001 461
aration index (PSEP) were provided.
PSEP is based on the difference between
the positive and negative predictive val-
ues of a test and can be derived from its
sensitivity and specificity. The positive
predictive value of a test shows the prob-
ability of someone with a positive test ac-
tually having the disease. The greater the
difference or separation between the pos-
itive and negative predictive values, the
better the PSEP and the more useful the
test or score for discriminating between
individuals with and without the disease.
RESULTS — Clinical and metabolic
descriptors of the study population are
shown in Table 1. Correlation coefficients
among Mbw/I and Mffm/I, commonly
used indexes, and individual variables
known to be associated with insulin resis-
tance are presented in Table 2. A total of
75 (42%) people in our sample met the
criteria for insulin resistance. The corre-
lation between Mbw/I and Mffm/I was
0.95. The variables most strongly corre-
lated with insulin sensitivity were fasting
insulin, fasting triglycerides, aspartate
aminotransferase (AST), waist circumfer-
ence, and BMI. To determine whether
sensitivity and specificity for predicting
insulin sensitivity could be improved, all
variables considered in a score, i.e., age,
WHR, BP, HDL cholesterol, liver en-
zymes, and those previously listed, were
considered in various combinations. The
combination of insulin and triglycerides
proved to be the best predictor and was a
better predictor than insulin alone. The
addition of other variables did not im-
prove the prediction of insulin sensitivity.
BMI was considered because it is regard-
ed as an important determinant of insu-
lin resistance, and the results for this are
shown. The scores used to predict Mffm/I
as a continuous variable are as follows:
Score 1A: Mffm/I
exp[3.29 0.25ln(insulin)
0.22ln(BMI) 0.28ln(TAG)]
Score 1B: Mffm/I
exp[2.63 0.28ln(insulin)
0.31ln(TAG)]
The scores used to predict Mffm/I as a
categorical variable were based on the fol-
lowing:
Score 2A–3.62 1.90ln(insulin)
0.43ln(BMI) 1.30ln(TAG)
Score 2B4.93 1.81ln(insulin)
1.24ln(TAG)
The first two scores give values for ISI.
The second two scores, where ISI is a cat-
egorical variable, can be translated into
probabilities. More precise estimates of
the probability of having insulin resis-
tance for scores 2Aand 2Bare obtained
from the following:
A score 0 corresponds to a 50%
chance of insulin resistance, a score 1
corresponds to a 73% chance, and a score
2 corresponds to an 88% chance.
The sensitivity and specificity derived
from regressing each of these variables
alone and in combination are shown in
Table 3. Insulin alone has a specificity of
0.82; the sensitivity, however, increased
from 0.57 to 0.64, either when all three
variables were part of the score or when
only insulin and triglycerides were used.
The bootstrap procedures used to exam-
ine the internal validity of this score showed
that there was little shrinkage and provided
95% CIs for sensitivities of 0.53–0.73 and
specificities of 0.740.88. Table 3 shows
that the sensitivity and specificity are similar
to the values obtained when Mffm/I was
used as a continuous variable. However, the
validation studies show that much more
shrinkage occurs when Mffm/I is categori-
cal. This means that the score will not be
able to divide individuals into two groups
on a new sample as accurately as it did on
the sample used to establish the score. The
PSEP values show that the greatest separa-
tion between groups is provided by insulin.
A modest but significant increase occurs
with the addition of triglycerides. Although
the score is maintained when Mffm/I is used
as a categorical variable, the validation pro-
cedure suggests that this is because of a
trade-off that occurs between sensitivity
and specificity.
Score 2A: the probability (insulin resistant)
exp[–3.62 1.90ln(insulin) 0.43ln(BMI) 1.30ln(TAG)]
1exp[–3.62 190ln(insulin) 0.43ln(BMI) 1.30ln(TAG)]
Score 2B: the probability (insulin resistant)
exp[–4.93 1.81ln(insulin) 1.24ln(TAG)]
1exp[–4.93 1.81ln(insulin) 1.24ln(TAG)]
Table 2—Product moment correlation coefficients between Mbw/I and Mffm/I using the
euglycemic insulin clamp method are commonly used indexes and risk factors for insulin
resistance
nMbw/I
Log(Mbw/I) and
log(variable) Mffm/I
Log(Mffm/I) and
log(variable)
Insulin (mU/l) 178 0.40 0.56 0.37 0.50
HOMA 178 0.42 0.53 0.39 0.51
Insulin-to-glucose ratio 178 0.41 0.47 0.34 0.47
Bennett index 178 0.42 0.45 0.33 0.48
BMI (kg/m
2
) 178 0.51 0.59 0.35 0.42
Waist (cm) 99 0.53 0.55 0.41 0.43
WHR 99 0.26 0.27 0.24 0.26
TAG (mmol/l) 178 0.40 0.48 0.37 0.45
HDL (mmol/l) 178 0.36 0.40 0.30 0.35
AST (U/l) 73 0.20* 0.44 0.21* 0.44
ALT (U/l) 73 0.22* 0.20* 0.24 0.19*
GGT (U/l) 73 0.31 0.26 0.32 0.28
Family history 101 0.14* 0.11*† 0.17* 0.16*†
BP 177 0.19 0.21† 0.11* 0.13*†
*NS; †the Mbw/I and Mffm/I have been log-transformed, but the variable has not been log-transformed.
Predicting insulin resistance
462 DIABETES CARE,VOLUME 24, NUMBER 3, MARCH 2001
CONCLUSIONS — Predicting insu-
lin sensitivity in normoglycemic individ-
uals is important, as diabetes intervention
programs are more likely to be successful
at this stage rather than after the develop-
ment of impaired glucose tolerance. Most
studies have investigated predictors of
insulin resistance; however, almost all
studies have included people with IGT
and diabetes, rather than normoglyce-
mic individuals in the general population
(5–7,13,17). In our study, fasting insulin
alone was as accurate at predicting insulin
resistance in the normoglycemic popula-
tion as HOMA, insulin-to-glucose ratio,
and the Bennett index. Our finding is
comparable with that of Laakso (4), who
demonstrated that fasting insulin alone
was less variable and had a higher corre-
lation with ISI in individuals with normo-
glycemia than in individuals with IGT and
diabetes. Thus, any method to predict in-
sulin sensitivity in normoglycemic indi-
viduals should be compared with fasting
insulin. Cross-reactivity with proinsulin
is unlikely to alter these findings, because
proinsulin levels are low in insulin-resistant
normoglycemic individuals, and the pat-
tern of response is similar to that of spe-
cific insulin (18,19).
A number of clinical and metabolic
abnormalities have been associated with
insulin resistance (1). Mykkanen et al. (6)
has confirmed that low insulin sensitivity
is associated with “clusters” of metabolic
disorders and that the ISI (measured by
an IVGTT) decreased with an increased
number of disorders. The metabolic dis-
orders were classified as dyslipidemia, hy-
pertension, and IGT. Howard et al. (17)
compared several alternative methods for
measuring insulin sensitivity to predict
cardiovascular risk. Many of the methods,
including the modified Galvin method
and other methods based on the fre-
quently sampled IVGTT, are invasive and
time consuming, and they are not appro-
priate for general population screening.
However, they did confirm that the best
method was dependent on glucose status.
The Galvin and the HOMA methods were
the most useful across all glucose levels.
In addition, the data for individuals with
normal glucose tolerance from these two
methods were consistent with our own
method, showing that a fasting insulin
was as good, if not better, than HOMA,
insulin-to-glucose ratio, or the Bennett in-
dex. Berglund and Lithell (7) have used
BMI and either triglycerides (TAG) or a
serum alanine-amino transferase to pre-
dict insulin resistance in hypertensive pa-
tients. They found either combination to
be as useful as a fasting insulin, but they
did not add fasting insulin to either com-
bination. However, glucose status was not
reported in all of the groups they evalu-
ated. More recently, Strumvoll et al. (8)
assessed 104 nondiabetic individuals to
determine whether age, BMI, WHR, and
glucose and insulin levels during an oral
glucose tolerance test could predict insu-
lin sensitivity (measured using a euglyce-
mic insulin clamp). They found that BMI,
insulin at 120 min, and glucose at 90 min
best predicted insulin sensitivity. These
parameters appeared robust in individu-
als with normal glucose tolerance (n
65), as well as in individuals with IGT.
Other studies have not found a 120-min
insulin measurement as useful as a fasting
insulin measurement (3,4). It is worth
noting that in the Strumvoll et al. study
(8), the correlation coefficient between
fasting insulin and ISI (–0.59) was re-
markably similar to that between ISI and
120-min insulin (–0.62), suggesting little
difference between fasting insulin and
120-min insulin as predictors. In sum-
mary, simple predictors of insulin resis-
tance have often been applied to those
with a range of glucose tolerance, and
there is evidence that the best method de-
pends on glucose status.
When measures of obesity, i.e., BMI,
waist circumference, and WHR, are as-
sessed as predictors of insulin resistance
(measured by a euglycemic insulin clamp),
it is important that glucose disposal is
expressed for fat-free mass rather than for
total body weight, because the contribu-
tion of fat mass to glucose disposal is
small (12). Muscle is the primary site for
glucose disposal, and if total body weight
is used in the calculation of glucose dis-
posal, then insulin resistance is consider-
ably overestimated in overweight
individuals (12). It is possible that if other
Table 3—Sensitivity, specificity, PSEP, and shrinkage analysis for measures of insulin sen-
sitivity (expressed Mffm/I) using three different methods
Cutoff Sensitivity Specificity PSEP
Method A
Insulin 12.2 mU/l 0.57 0.82 0.42
BMI 29.3 kg/m
2
0.56 0.76 0.33
TAG 1.5 mmol/l 0.53 0.75 0.29
Insulin, BMI, and TAG 0.63 0.82 0.46
Validation estimates 0.64 0.81 0.46
95% CI 0.53–0.73 0.74–0.88 0.40–0.52
Insulin and TAG 0.62 0.84 0.40
Validation estimates 0.64 0.83 0.49
95% CI 0.53–0.73 0.74–0.88 0.44–0.52
Method B
Insulin 12.1 mU/l 0.56 0.84 0.45
BMI 30.0 kg/m
2
0.52 0.78 0.31
TAG 1.6 mmol/l 0.53 0.79 0.34
Insulin, BMI, and TAG 0.61 0.85 0.50
Validation estimates 0.45 0.91 0.50
95% CI 0.32–0.6 0.85–0.96 0.45–0.55
Insulin and TAG 0.61 0.84 0.49
Validation estimates 0.49 0.92 0.51
95% CI 0.32–0.6 0.86–0.95 0.47–0.54
Method C
Insulin 12.0 mU/l 0.71 0.69 0.39
BMI 30.0 kg/m
2
0.69 0.64 0.33
TAG 1.5 mmol/l 0.63 0.58 0.20
Insulin, BMI, and TAG 0.65 0.79 0.45
Validation estimates 0.52 0.83
95% CI 0.0–0.65 0.72–1.0
Method A, ISI and predictors analyzed continuously; method B, categorical ISI (6.3 MmU
1
l
1
) and
continuous predictors; method C, categorical ISI and predictors. Cutoffs have been derived from the equa-
tions for each variable.
MCAuley and Associates
DIABETES CARE,VOLUME 24, NUMBER 3, MARCH 2001 463
studies had corrected for fat-free mass,
BMI would not have been as important in
predicting insulin resistance. In our
study, the correction for fat-free mass re-
duced the correlation between ISI and
BMI from – 0.59 to –0.41. The correlation
of ISI and waist circumference, a crude
measure of central adiposity, was reduced
to a lesser extent (from –0.55 to –0.43)
after the correction for fat-free mass. It is
difficult to draw any conclusions about
the usefulness of waist circumference
compared with BMI, given the reduced
number of measurements of waist cir-
cumference in this sample; however, this
should be considered in future studies.
In our study, a fasting insulin of
12.2 mU/l in normoglycemic individu-
als is a remarkably specific test for insu-
lin resistance. Our aim was to discover
whether the addition of other variables
that correlate well with insulin resistance
improved on a fasting insulin alone. The
variables (other than fasting insulin) that
best predicted insulin sensitivity (Mffm/I)
were fasting triglycerides, AST, waist cir-
cumference, and BMI. The addition of tri-
glycerides to a fasting insulin increases sen-
sitivity from 0.57 to 0.64 and maintains
good specificity. When glucose disposal is
corrected for fat-free mass, rather than for
total body weight, BMI does not increase
the sensitivity or specificity of this combi-
nation. Waist circumference and AST also
appear to be important associated vari-
ables and warrant further assessment;
however, because of the small number of
these measurements in this study, no fur-
ther conclusions can be made.
This study shows that the best predic-
tors of insulin sensitivity in the general
population were the log-transformed val-
ues for fasting insulin and fasting triglyc-
erides. We found that it is important to
use continuous variables, because in our
study the shrinkage was small, which
means the score (1B) should perform well
on future samples. Many prediction mod-
els suggested for clinical use prefer to use
categories, rather than continuous varia-
bles, for simplicity and ease of calculation.
However,this study shows that whencate-
gories are used, there is considerable shrink-
age, which means the score may not pre-
dict well in new samples. The equation
predicting insulin sensitivity using con-
tinuous variables is just as easily applied
as that for categories using a simple com-
puter program and does not sacrifice
valuable information.
Acknowledgments— This study was funded
by the Health Research Council, Otago Uni-
versity; Novo Nordisk; and the Otago Diabetes
Research Trust, New Zealand.
Many thanks to Maggie Oakley, Margaret
Waldron, and Val Burke (research nurses),
and Gretchen Story and Damon Bell (research
assistants).
References
1. DeFronzo RA, Ferrannini E: Insulin resis-
tance: a multifaceted syndrome respon-
sible for NIDDM, obesity, hypertension,
dyslipidemia, and atherosclerotic cardio-
vascular disease. Diabetes Care 14:173–194,
1991
2. Ferrannini E: Insulin resistance is central
to the burden of diabetes. Diabetes Metab
Rev 13:81–86, 1997
3. Ferrannini E, Mari A: How to measure in-
sulin sensitivity. J Hypertens 16:895–906,
1998
4. Laakso M: How good a marker is insulin
level for insulin resistance? Am J Epidemiol
137:959–965, 1993
5. Matthews D, Hosker J, Rudenski A, Nay-
lor B, Treacher D, Turner R: Homeostasis
model assessment: insulin resistance and
B-cell function from fasting plasma glu-
cose and insulin concentrations in man.
Diabetologia 28:412–419, 1985
6. Mykkanen L, Haffner S, Ronnemaa T,
Bergman R, Laakso M: Low insulin sensi-
tivity is associated with clustering of car-
diovascular disease risk factors. Am J
Epidemiol 146:315–321, 1997
7. Berglund L, Lithell H: Prediction models
for insulin resistance. Blood Press 5:274
277, 1996
8. Strumvoll M, Mitrakou A, Pimenta W,
Jenssen T, Yki-Jarvinen H, Van Haeften T,
Renn W, Gerich J: Use of the oral glucose
tolerance test to assess insulin release and
insulin sensitivity. Diabetes Care 23:295–
301, 2000
9. Russell D, Parnell W, Wilson N: Key Re-
sults of the 1997 National Nutrition Sur-
vey. In NZ Food: NZ People. Dunedin, New
Zealand, University of Otago for the Min-
istry of Health, 1999, p. 1–268
10. DeFronzo R, Tobin J, Andres R: Glucose
clamp technique: a method for quantify-
ing insulin secretion and resistance. Am J
Physiol 237:E214–E223, 1979
11. Pratt M, Lewis-Barned N, Walker R: A
comparison between enalapril and capto-
pril on insulin sensitivity in normotensive
healthy volunteers. AustNZJMed23:
652–655, 1993
12. Ferrannini E, Natali A, Bell P: Insulin re-
sistance and hypersecretion in obesity.
J Clin Invest 100:1166–1173, 1997
13. Anderson RL, Hamman RF, Savage PJ,
Saad MF, Laws A, Kades WW, Sands RE,
Cefalu W: Exploration of simple insulin
sensitivity measures derived from the fre-
quently sampled intravenous glucose tol-
erance (FSIGT) tests. Am J Epidemiol 142:
724–732, 1995
14. Altman D, Andersen P: Bootstrap investi-
gation of the stability of a Cox regression
model. Stat Med 8:771–783, 1989
15. Harrell F, Lee K, Mark D: Multivariable
prognostic models: issues in developing
models evaluating assumptions and ade-
quacy and measuring and reducing er-
rors. Stat Med 15:361–387, 1996
16. Altman D, Royston P: What do we mean
by validating a prognostic model?. Stat
Med 19:453–473, 2000
17. Howard G, Bergman R, Wagenknecht LE,
Haffner SM, Savage PJ, Saad MF, Laws A,
D’Agostino RB Jr: Ability of alternative in-
dices of insulin sensitivity to predict car-
diovascular risk: comparison with the
“minimal model.” Ann Epidemiol 8:358
369, 1998
18. Snehalatha C, Ramachandran A, Satya-
vani K, Vijay V, Haffner S: Specific insulin
and proinsulin concentrations in nondia-
betic South Indians. Metabolism 47:230
233, 1998
19. Mykka¨nen L, Haffner S, Hales C, Ro¨nne-
maa T, Laakso M: The relation of proin-
sulin, insulin, and proinsulin-to-insulin
ratio to insulin sensitivity and acute insu-
lin response in normoglycemic subjects.
Diabetes 46:1990–1995, 1997
Predicting insulin resistance
464 DIABETES CARE,VOLUME 24, NUMBER 3, MARCH 2001
... All of them were built based on the non-pregnant diabetic population, which indicates that there could be some bias when used to gauge IR among GDM. [11][12][13] Fetal ultrasound is not only a relatively easy-to-do imaging scan but also a safe method for the assessment of fetal growth. In Vietnam, it is recommended that all pregnant women have a pregnancy ultrasound scan at least 3 times during pregnancy at the end of each quarter to manage pregnancy and help the prognosis of delivery. ...
Article
Full-text available
Background In pregnant women with gestational diabetes mellitus (GDM), insulin resistance (IR) increases the risk of developing manifest type 2 diabetes mellitus and is associated with complications in both mother and fetus. Objectives This research aimed to evaluate the associations between IR evaluated by 3 indices (namely updated homeostasis model assessment model (HOMA2), QUICKI, and McAuley’s index) and the diabetes risk factors and the fetal growth indices in Vietnamese women with GDM. Methods A cross-sectional descriptive study was conducted on 370 women with GDM and 40 healthy pregnant women from January 2015 to May 2019. IR was calculated by HOMA2 (HOMA2-IR), QUICKI, and McAuley’s index. Fetal anthropometric measurements were assessed via ultrasound which was performed and interpreted by ultrasound experts. Results In the simple regression analysis, McAuley’s index illustrated had statistically significant correlations to the highest number of risk factors of diabetes mellitus compared with HOMA2-IR and QUICKI indices. Moreover, McAuley’s index correlated statistically significantly to the highest number of fetal ultrasound measurements factors such as including biparietal diameter (BPD) ( r = −0.271, P < .001), head circumference (HC) ( r = −0.225, P < .001), abdominal circumference (AC) ( r = −0.214, P < .001), femur length (FL) ( r = −0.231, P < .001), estimated fetal weight (EFW) ( r = −0.239, P < .001) and fetal estimated age ( r = −0.299, P < .001). In the multivariable analysis, the McAuley’s index contributed the greatest to AC (Standardized B of −0.656, P < .001). Conclusion The McAuley’s index was significantly associated with a higher number of more risk factors for diabetes mellitus as well as fetal ultrasound sonography findings measurements than compared with HOMA2-IR and QUICKI indices.
... The coefficients of variation (Cv) were as follows: intra-assay leptin 3.7 to 5.5% and insulin 2.9 to 6.2%, and inter-assay leptin 5.8 to 6.8% and insulin 5.4 to 8.6%. The homeostasis model assessment for insulin resistance (HOMA-IR) was used to estimate changes in insulin resistance using standard methods [35]. ...
Article
Full-text available
Aim: To compare the efficacy of two popular weight loss approaches on weight loss, body composition, and markers of health in sedentary obese women. Methods: In total, 51 sedentary women (age 34.5 ± 7.7 yrs.; weight 90.0 ± 14.5 kg; BMI 34.0 ± 5.1 kg/m2; 46.5 ± 7.0% fat) were matched and randomized to participate in the Weight Watchers® Momentum™ (WW) or Curves® (CV) Fitness and Weight Management program for 16 weeks. Participants in the WW group (n = 27) were provided a point-based diet program, received weekly progress checks and counseling, and were encouraged to exercise. Participants in the CV group (n = 24) followed a menu-based higher protein/low-fat diet (1200 kcal/d) for 1 week; 1500 kcal/d diet for 3 weeks; and 2000-2500 kcals/d for 2 weeks that was repeated three times (except the last segment) while participating in a supervised circuit-style resistance training program (3 d/wk). A general linear model (GLM) with repeated measures was used to analyze data and are presented as mean changes from baseline (mean [UL, LL]). Results: Supervised CV training resulted in greater amounts of vigorous and total physical activity. After 16 weeks, both groups lost weight (WW -6.1 [-7.8, -4.6], CV -4.9 [-6.2, -3.2] kg, p = 0.264). Participants in the CV group observed greater reductions in fat mass (WW -2.9 [-6.7, -0.2], CV -6.4 [-9.2, -3.6] kg, p = 0.081) and increases in lean mass (WW -2.5 [-4.3, -0.7], CV 1.3 [-0.6, 3.2] kg, p = 0.005) resulting in more favorable changes in percent body fat (WW -1.4 [-4.1, 1.2], CV -4.7 [-7.5, -1.8]%, p = 0.098). Both groups observed improvements in peak aerobic capacity and muscular endurance, although bench press lifting volume was greater in the CV group. Those in the CV group experienced a greater increase in HDLc and reduction in the CHL-HDLc ratio and triglycerides. Conclusion: Both interventions promoted weight loss and improvements in fitness and markers of health. The CV program, which included supervised resistance training and higher protein diet menus, promoted greater fat loss, increases in lean mass, and improvements in percent body fat and blood lipids. Trial registration: clinicaltrials.gov, #NCT04372771, registered retrospectively 1 May 2020.
... All OB underwent bariatric surgery for the first time and the lean individuals underwent surgery for non-inflammatory surgical conditions. The OB were divided into two subgroups as follows: a) MODM: 16 adults with morbid obesity and T2D; and b) MOW: 20 adults with morbid obesity and hyperinsulinemia (defined as fasting insulin ≥12.2 μU/mL in the presence of euglycemia) 16 . In the MODM group, the majority of adults had long-standing T2D with a mean duration of 3.69 years and were under treatment with metformin (13/14) or glimepiride (1/14). ...
Article
Full-text available
Objective: Obesity is characterized by hypertrophy and pathological expansion of adipocytes with impaired insulin signaling causing insulin resistance (IR) and metabolic dysfunction. We recently reported decreased expression of glucose transporter-4 (GLUT4) in cultured adipocytes from visceral and abdominal subcutaneous fat depots from patients with morbid obesity and hyperinsulinemia (MOW) and with Type 2 diabetes (MODM). Subsequently, we wanted to study the molecular mechanisms of the glucose transport regulators, p85PI3K, Rab5 and Gapex5 in morbid obesity. Patients and methods: Primary in vitro adipocyte cultures were developed from surgical biopsies from visceral (Visc) and abdominal (Sub) and gluteal subcutaneous (Glut) fat depots from 20 lean adults and 36 adults with morbid obesity divided into two groups: 20 with MOW and 16 MODM). mRNA and protein expression (P) of p85PI3K, Rab5 and Gapex5 were studied with RT-PCR and Western Immunoblotting (WI), respectively. Results: In Sub, the P of (1) p85PI3K and Gapex5 were increased in MODM and (2) Rab5 was decreased in MOW and MODM compared to the lean. In Glut, the P of p85PI3K, Rab5 and Gapex5 showed no difference between the lean and MODM. Conclusions: In Sub of MODM (1) reduced RAB5 may possibly contribute to IR and glucose transport dysfunction, (2) increased Gapex5 may be a response to decreased Rab5 in an attempt to increase glucose transport and (3) increased p85PI3K may enhance IR mediating lipid accumulation in MODM. In Glut of MODM, though, the expression of p85PI3K, Rab5 and Gapex5 seems to be similar to that found in lean individuals.
Article
Full-text available
Objective: Head and Neck cancer (HNC) is the commonest malignancy encountered by otolaryngologists. Surgery is the primary treatment modality for most of them. These cancers have good prognosis whenappropriate treatment is provided timely. This study aimed to evaluate clinicohistopathological profile of HNCs treated during COVID-19 pandemic at a tertiary referral hospital. Methods: This retrospective study was conducted at the department of ENT-Head and Neck Surgery, Shree Birendra Hospital. A total of 83 biopsy proven HNC patients with complete medical records who underwent treatment between February 2020 and September 2021 were included. Results: There were 58 males with mean age of 51.4 years and 25 females with mean age of 48.4 years having male to female ratio of 2.3:1. Head and neck cancer was significantly higher (78.3%) in patients with age ≥ 40 years as compared to patients ≤ 40 years (p = 0.000). Oral cavity cancer was the most common malignancy followed by thyroid cancer. Squamous cell carcinoma (SCC) was the commonest histological subtype (65%) in oral cavity. Papillary cancer was the commonest among thyroid cancers. Multimodality treatment was received by 98.8% patients. Conclusions: Oral cavity SCC was the most common malignancy in males of 40-60 years age group.Papillary thyroid cancer was the commonest amongst females of 20-40 years age group. HNCs were significantly higher in 40 years and above age group compared to younger patients. Half of the patients underwent surgery with adjuvant post-operative radiotherapy/ radioiodine ablation or concurrent Chemoradiotherapy (CTRT). Almost all patients required multimodality treatment signifying presentation of HNCs at advanced stages. Keywords: covid, pandemic, ENT-head and neck cancers, profile
Article
Objective: Hepatic glucose release plays a potential role in hyperglycemia in type 2 diabetes (T2D) patients. The aim of this experimental study was to determine the effect of 6 weeks of high-intensity interval training (HIIT) on fasting levels of glucose and insulin as well as glucokinase (GCK) expression in liver tissue in obese T2D rats. Materials and Methods: T2D was induced by a high-fat diet (HFD) and streptozotocin (STZ) intraperitoneal injection in 14 male wistar rats, then were randomly divided into HIIT (n=7) and control (n=7) groups. The HIIT group practiced 6-week HIIT (5 days/ weekly). Finally, 48 hours after the last session, fasting levels of glucose, insulin, and GCK expression in liver hepatocytes of both groups were measured and compared by independent T-test (SPSS, Version 22.0). Results: HIIT resulted in a significant decrease of fasting glucose compared to the control group (P< 0.0001). Compared with the control group, serum insulin (P: 0.018) and GCK expression in hepatocytes (P: 0.030) were significantly increased. Conclusion: Based on these findings, the improvement in glucose in response to HIIT may be rooted in increased insulin and GCK expression in hepatocytes. However, understanding the mechanisms responsible for the effect of exercise training on the processes affecting hepatic glucose release requires further studies.
Article
Among all endocrine diseases complicating pregnancy, gestational diabetes mellitus (GSD) is the most common. Th is violation of carbohydrate metabolism poses a serious threat to the health of the mother and fetus, associated with a high risk of perinatal complications. At the same time, the eff ective achievement of normoglycemia in a woman suff ering from GSD can signifi cantly improve the prognosis. Th e choice of GSD therapy depends on a number of factors and is decided individually in each case. One of the factors determining the eff ectiveness of non-drug therapy and the need for pharmacological correction may be related to the pathophysiological aspects of the formation of hyperglycemia during pregnancy. Currently, they talk about the heterogeneity of GSD and distinguish its various subtypes depending on the predominance of pancreatic beta-cell dysfunction, insulin resistance (IR) or a combination of these factors in the pathogenesis. Since the prevailing criterion for the verifi cation of GSD subtypes is the presence and severity of IR, various methods of its verifi cation are considered in this review. It is shown that the currently available methods for detecting IR have a number of disadvantages, consisting both in the complexity and complexity of implementation (hyperinsulinemic euglycemic clamp) and in the absence of clear reference intervals (mathematical models). It is necessary to continue research aimed at studying IR methods for the subsequent identifi cation of GSD subtypes.
Article
Background and Aims Bile acid malabsorption (BAM) is a debilitating disease characterized by loose stools and high stool frequency. The pathophysiology of BAM is not well-understood. We investigated postprandial enterohepatic and gluco-metabolic physiology, as well as gut microbiome composition and fecal bile acid content in patients with BAM. Methods Twelve participants with selenium-75 homocholic acid taurine test–verified BAM and 12 healthy controls, individually matched on sex, age, and body mass index, were included. Each participant underwent 2 mixed meal tests (with and without administration of the bile acid sequestrant colesevelam) with blood sampling and evaluation of gallbladder motility; bile acid content and microbiota composition were evaluated in fecal specimens. Results Patients with BAM were characterized by increased bile acid synthesis as assessed by circulating 7-alpha-hydroxy-4-cholesten-3-one, fecal bile acid content, and postprandial concentrations of glucose, insulin, C-peptide, and glucagon. The McAuley index of insulin sensitivity was lower in patients with BAM than that in healthy controls. In patients with BAM, colesevelam co-administered with the meal reduced postprandial concentrations of bile acids and fibroblast growth factor 19 and increased 7-alpha-hydroxy-4-cholesten-3-one concentrations but did not affect postprandial glucagon-like peptide 1 responses or other gluco-metabolic parameters. Patients with BAM were characterized by a gut microbiome with low relative abundance of bifidobacteria and high relative abundance of Blautia, Streptococcus, Ruminococcus gnavus, and Akkermansia muciniphila. Conclusion Patients with BAM are characterized by an overproduction of bile acids, greater fecal bile acid content, and a gluco-metabolic profile indicative of a dysmetabolic prediabetic-like state, with changes in their gut microbiome composition potentially linking their enterohepatic pathophysiology and their dysmetabolic phenotype. ClinicalTrials.gov number NCT03009916.
Article
Polycystic ovary syndrome (PCOS) is one of the most common endocrine disorders, affecting 5%-10% of women of reproductive age. The importance of this syndrome lies in the magnitude of associated comorbidities: infertility, metabolic dysfunction, cardiovascular disease (CVD), plus psychological and oncological complications. Insulin resistance (IR) is a prominent feature of PCOS with a prevalence of 35%-80%. Without adequate management, IR with compensatory hyperinsulinemia contributes directly to reproductive dysfunction in women with PCOS. Furthermore, epidemiological data shows compelling evidence that PCOS is associated with an increased risk of impaired glucose tolerance, gestational diabetes mellitus and type 2 diabetes. In addition, metabolic dysfunction leads to a risk for CVD that increases with aging in women with PCOS. Indeed, the severity of IR in women with PCOS is associated with the amount of abdominal obesity, even in lean women with PCOS. Given these drastic implications, it is important to diagnose and treat insulin resistance as early as possible. Many markers have been proposed. However, quantitative assessment of IR in clinical practice remains a major challenge. The gold standard method for assessing insulin sensitivity is the hyperinsulinemic euglycemic glucose clamp. However, it is not used routinely because of the complexity of its procedure. Consequently, there has been an urgent need for surrogate markers of IR that are more applicable in large population-based epidemiological investigations. Despite this, many of them are either difficult to apply in routine clinical practice or useless for women with PCOS. Considering this difficulty, there is still a need for an accurate marker for easy, early detection and assessment of IR in women with PCOS. This review highlights markers of IR already used in women with PCOS, including new markers recently reported in literature, and it establishes a new classification for these markers.
Article
Prognostic models are used in medicine for investigating patient outcome in relation to patient and disease characteristics. Such models do not always work well in practice, so it is widely recommended that they need to be validated. The idea of validating a prognostic model is generally taken to mean establishing that it works satisfactorily for patients other than those from whose data it was derived. In this paper we examine what is meant by validation and review why it is necessary. We consider how to validate a model and suggest that it is desirable to consider two rather different aspects – statistical and clinical validity – and examine some general approaches to validation. We illustrate the issues using several case studies. Copyright © 2000 John Wiley & Sons, Ltd.
Article
Multivariable regression models are powerful tools that are used frequently in studies of clinical outcomes. These models can use a mixture of categorical and continuous variables and can handle partially observed (censored) responses. However, uncritical application of modelling techniques can result in models that poorly fit the dataset at hand, or, even more likely, inaccurately predict outcomes on new subjects. One must know how to measure qualities of a model's fit in order to avoid poorly fitted or overfitted models. Measurement of predictive accuracy can be difficult for survival time data in the presence of censoring. We discuss an easily interpretable index of predictive discrimination as well as methods for assessing calibration of predicted survival probabilities. Both types of predictive accuracy should be unbiasedly validated using bootstrapping or cross-validation, before using predictions in a new data series. We discuss some of the hazards of poorly fitted and overfitted regression models and present one modelling strategy that avoids many of the problems discussed. The methods described are applicable to all regression models, but are particularly needed for binary, ordinal, and time-to-event outcomes. Methods are illustrated with a survival analysis in prostate cancer using Cox regression.
Article
Prognostic models are used in medicine for investigating patient outcome in relation to patient and disease characteristics. Such models do not always work well in practice, so it is widely recommended that they need to be validated. The idea of validating a prognostic model is generally taken to mean establishing that it works satisfactorily for patients other than those from whose data it was derived. In this paper we examine what is meant by validation and review why it is necessary. We consider how to validate a model and suggest that it is desirable to consider two rather different aspects – statistical and clinical validity – and examine some general approaches to validation. We illustrate the issues using several case studies. Copyright © 2000 John Wiley & Sons, Ltd.
Article
Methods for the quantification of beta-cell sensitivity to glucose (hyperglycemic clamp technique) and of tissue sensitivity to insulin (euglycemic insulin clamp technique) are described. Hyperglycemic clamp technique. The plasma glucose concentration is acutely raised to 125 mg/dl above basal levels by a priming infusion of glucose. The desired hyperglycemic plateau is subsequently maintained by adjustment of a variable glucose infusion, based on the negative feedback principle. Because the plasma glucose concentration is held constant, the glucose infusion rate is an index of glucose metabolism. Under these conditions of constant hyperglycemia, the plasma insulin response is biphasic with an early burst of insulin release during the first 6 min followed by a gradually progressive increase in plasma insulin concentration. Euglycemic insulin clamp technique. The plasma insulin concentration is acutely raised and maintained at approximately 100 muU/ml by a prime-continuous infusion of insulin. The plasma glucose concentration is held constant at basal levels by a variable glucose infusion using the negative feedback principle. Under these steady-state conditions of euglycemia, the glucose infusion rate equals glucose uptake by all the tissues in the body and is therefore a measure of tissue sensitivity to exogenous insulin.
Article
INTRODUCTION: Although recognition of insulin sensitivity as a risk factor for cardiovascular disease is growing, a deeper understanding of its role is impeded by the cost and complexity of currently available measures. This report evaluates previously described alternative indices of insulin sensitivity with the goal of identifying a reliable, but logistically simpler, alternative.METHODS: Data from 1460 participants in the Insulin Resistance Atherosclerosis Study (IRAS) were used to assess the proportion of the relationship between a recognized measure of insulin sensitivity (Bergman’s SI) and cardiovascular risk factors that is contained in each of nine alternative measures.RESULTS: A number of the alternative indices contained a substantial proportion of the information available in Bergman’s SI. The Galvin’s index and the homeostasis model were most promising. However, there remained a significant amount of the information in Bergman’s SI that was not contained in any of the alternative indices.DISCUSSION: There are simpler alternative indices of insulin sensitivity for use in epidemiological studies, but each alternative is associated with some loss of information. It may be possible that this loss can be overcome with an increased sample size; however, using the alternative indices may also confound the assessment of insulin sensitivity with other underlying factors (i.e., hyperinsulinemia). The alternative indices are not recommended for the clinical assessment of insulin sensitivity for an individual patient or subject.
Article
Multivariable regression models are powerful tools that are used frequently in studies of clinical outcomes. These models can use a mixture of categorical and continuous variables and can handle partially observed (censored) responses. However, uncritical application of modelling techniques can result in models that poorly fit the dataset at hand, or, even more likely, inaccurately predict outcomes on new subjects. One must know how to measure qualities of a model's fit in order to avoid poorly fitted or overfitted models. Measurement of predictive accuracy can be difficult for survival time data in the presence of censoring. We discuss an easily interpretable index of predictive discrimination as well as methods for assessing calibration of predicted survival probabilities. Both types of predictive accuracy should be unbiasedly validated using bootstrapping or cross-validation, before using predictions in a new data series. We discuss some of the hazards of poorly fitted and overfitted regression models and present one modelling strategy that avoids many of the problems discussed. The methods described are applicable to all regression models, but are particularly needed for binary, ordinal, and time-to-event outcomes. Methods are illustrated with a survival analysis in prostate cancer using Cox regression.
Article
Prognostic models are used in medicine for investigating patient outcome in relation to patient and disease characteristics. Such models do not always work well in practice, so it is widely recommended that they need to be validated. The idea of validating a prognostic model is generally taken to mean establishing that it works satisfactorily for patients other than those from whose data it was derived. In this paper we examine what is meant by validation and review why it is necessary. We consider how to validate a model and suggest that it is desirable to consider two rather different aspects – statistical and clinical validity – and examine some general approaches to validation. We illustrate the issues using several case studies. Copyright © 2000 John Wiley & Sons, Ltd.