Berman pcori challenge document

A Conceptual Model of Using Medical Measures
To Match Individuals for Health Research

Note: This work is derived from my Doctoral Dissertation, completed May 2011 at George
Washington University.

Lewis E. Berman, PhD, MS
April 15, 2013

Abstract
Lower survey and study response rates and higher costs provide significant challenges to
carry out biomedical and public health research. Increasingly health studies desire larger sample
sizes in order to analyze illnesses that may occur with low prevalence in the population.
Moreover, sub-group delineation is required in order to assess illness in hard to reach groups or
those groups that may occur with lower frequency in the general population.
The increasing availability of electronic medical information may serve as the foundation
for automatically matching individuals with health researchers for the purposes of advancing
health research. As electronic health records become the norm in the delivery of care, the record
and feature space for this data will become quite large. This will provide the basis for accurately
matching individuals with health researchers and projects.
This paper proposes a conceptual model to match individuals using filtering, data
reduction, and similarity coefficients. The filtering and data reduction steps reduce the scale of
the problem from a computational perspective. A simulation of the conceptual model is
illustrated. The findings from the simulation demonstrate that the record and feature space can be
significantly reduced and automated.

1 Introduction
There has been an increase in the demand for information access due to the widespread
use and ubiquitous nature of the Internet. Concurrently, medicine has undergone significant
change in equipment, procedures, treatments, monitoring, and specialization. In addition, the
federal government of the United States (U.S.) is investing in health information technology
(HIT) and electronic health records (EHR) with the hope that it will improve health [1].
Currently, individuals self-select into online health communities or pre-defined groups.
An alternative to self-selection is automated formation of health communities using medical
measurements. In essence, a “matchmaking mechanism” between patients can be automated
using medical measurements from an electronic health record [2, page 6]. While matching may
be done for social support, it may also be done for the purposes of health research.

1.1 Problem Statement
A common problem across disparate disciplines is matching and grouping objects based
on feature similarity. This is a classification. In the biological sciences, classification has been
emphasized to develop taxonomies such as the well-defined classification of the animal kingdom.

Currently, health studies utilize phone calling, mailing, and door-to-door visits to recruit
and match individuals for health research studies. It is widely agreed that health studies, and
studies in general, are achieving lower response rates for a variety of reasons. Moreover, in
attempting to recruit participants into these studies, the participant selection criterion is typically
limited by time and money. While this approach has some merit when considering the trade-off
between screening detail and cost, it is limiting since a study may be interested in recruiting large
numbers of individuals into a study and may need very detailed information for selection
purposes. So, an alternative to manual matching and selection is needed.
Therefore, this paper proposes to build a conceptual model for grouping individuals
based on electronically available medical measurements. The model consists of filtering, data
reduction, and similarity computation.

1.2 Research Approach and Organization of the Paper
The research approach in this paper is to develop the conceptual model and simulate the
model with a database of medical measurements. Section 2 is review of the relevant literature.
Section 3 presents the conceptual model and a simulation example. Section 4 presents the
simulation results. Section 5 is discusses the results. The last chapter is the conclusion.

2 Literature Review
This chapter is a review of the computational techniques related to the development of a
conceptual model for matching individuals. The topics cover medical measurement data types,
data reduction, and similarity coefficients.

2.1 Medical Measurement Data Types
Measurement is defined as the assignment of a number to an attribute of some instance of
an object. An important consideration in measurement is that the “properties of the attribute are
faithfully represented as numerical properties” as described by Krantz [3, page 1]. Medical
measurements are the result of tests, procedures, treatments, health history questions, or
diagnoses, and articulate an individual’s health state.
In general, there are four measurement types that may be assigned to medical
measurements. The first type is nominal measurement, which separates data into discrete groups
that are mutually exclusive. The second type is ordinal measurement. Ordinal measurement
assigns objects to categories such that these categories have a meaningful rank. In
epidemiological research, people may be pooled into different fitness groups such as poor, good,
and outstanding based on an individual’s perception of fitness level. While there is an ordering
and a sense of the magnitude difference between fitness groups, it is not possible to determine the
actual difference between groups. A third measurement type is interval. An example of an
interval measurement is Fahrenheit temperature. A temperature of 80° F is greater than a
temperature of 60° F. However, temperature, like all interval measurements, has two interesting
distinctions. First, a temperature of 0° F does not suggest the absence of temperature. Secondly,
even though temperature measurements possesses equal intervals it is not the case that there is a
true zero point and as a result, ratios between interval measures do not exist. Thus, 100° F is not
twice as hot as 50° F. The fourth measurement type is ratio. Ratio is much like interval except is
has an absolute zero point. Thus, a person who weighs 200 pounds is twice as heavy as a person
weighing 100 pounds and a 50-pound difference between any two weights always has the same
meaning [4, 5].

2

2.2 Data Reduction
The definition of data reduction is the process of converting large sets of data into a
smaller number of data points. Mathematically, data reduction is the transformation of an n-
dimensional vector of observed data points or measurements, m = (m1, m2, …, mn), to a k-
dimensional vector of variables t = (t1, t2, …, tk) such that k≤n. In addition, the transformation
from m to t adheres to some criterion [6].
Data reduction methods fall into linear and non-linear methods. Some well-used linear
methods include Principal Component Analysis (PCA) and Factor Analysis (FA). Non-linear
methods include Principal Curves (PC), Multidimensional Scaling (MDS), and Neural Networks
(NN). The linear methods are considered easier to implement than non-linear methods [6]. PCA
has been applied in biology, medicine, chemistry, meteorology, and the social sciences [6, 7].

2.3 Similarity
Similarity is the basis for classification and is defined to be the amount of resemblance
between two objects based on the distinct information pertaining to the variables (i.e., features) of
the objects [8]. Similarity coefficients have been applied to several fields such as manufacturing
systems, plant breeding, seed bank management, high throughput screening of chemical datasets,
and determining the molecular markers of genetic relationships between individuals [9, 10, 11,
12].
In 1901, Jaccard created the earliest similarity coefficient [13, 14]. There are a number of
other similarity coefficients. However, some coefficients such as geometric and ontological are
not suitable for this work because they restrict the type of measurement types that can be used or
a single feature may adversely skew the results. Therefore, this paper explores three commonly
used coefficients, developed by Jaccard, Gower, and Tversky, which are not as susceptible to
these issues.

2.3.1 Jaccard Coefficient
The Jaccard Coefficient (JC) is feature-based model (FBM) which uses common and
unique features to compute similarity between objects. As shown in Equation 1 JC computes the
ratio of the number of features in common between two objects and the total number features in
common plus the number of features possessed uniquely by each of the two objects.
Jaccard a Where:
 (1)
Coefficient:
a  b  c  a = # of features in common
st
b = # of features possessed by 1 object
c = # of features possessed by 2nd object

2.3.2 Tversky Feature Contrast Similarity Model
Tversky suggested using a set-theoretical approach known as the feature contrast model.
The Tversky Feature Contrast Model Coefficient (TFCMC) computes similarity as a linear
combination of the common and unique features of individual objects. Thus, for two objects A
and B, there is a similarity function S; non-negative set functions f and g that define the weights
of individual features and how they are combined; and two constants θ, α, β ≥ 0 such that [16]:
𝑆(𝐴, 𝐵) = ∅𝑔(𝐴 ∩ 𝐵) − (𝛼𝑓(𝐴 − 𝐵) + 𝛽𝑓(𝐵 − 𝐴)) (2)

3

2.3.3 Gower’s Model
In 1971, Gower proposed a similarity coefficient that could simultaneously use variables
of different measurement scales [8]. Gower computed the similarity between two objects, A and
B, as follows:
 p 
  S ( A, B) k 
 
S  A, B    p 
k 1
(3)
 
 W ( A, B) k 
 
 k 1 
For nominal or ordinal data S(A,B)k = 1 when the feature values are the same and 0
otherwise. For interval or ratio data S(A,B)k = 1 - | fAk – fBk | / Rk such that fAk and fBk are the
values of the features for objects A and B; Rk equals the range for feature k across all objects (i.e.,
persons). In essence, this function scales the real valued features. A second feature of the Gower
coefficient (GC) is the denominator, W(A,B)k, which is a type of binary weighting variable. It
takes a value of 1 when the comparison between feature fAk and fBk, for objects A and B, is
considered valid. Otherwise, it is equal to 0.

3 Conceptual Model
This paper proposes a conceptual model to match individuals for medical research. As
illustrated in Figure 1, the conceptual model progresses through candidate measurement vector
(CMV) selection, rule-based filtering, principal component analysis (PCA) data reduction, and
similarity computation. This chapter will describe the steps in the conceptual model, criteria for
selection of a simulation dataset, and a description of the simulation example.

3.1 Candidate Measurement Vector Selection
It is assumed that individuals are being grouped together to match with the objective of a
research study proposed by a research scientist. To match individuals a hypothetical “candidate”
individual is created to represent the features of a typical member of the group. The “candidate”
consists of a specific set of medical measurements related to the features of people needed for the
research study. In a typical research study, the investigator and their team define the features of
interest for the patient population. However, this algorithm allows the selection process to be
sensitive to the desires of the patient population by augmenting the feature set of the “candidate”.
For example, a research scientist might be interested in recruiting individuals with type 2
diabetes into a study on diabetes co-morbidity factors. In this conceptual model the first step is
for the research scientist to prepare a candidate measurement vector (CMV) that includes the type
2 diabetes co-morbidity measurement vector. In this case, a CMV could include measurements
for the history of smoking, high blood pressure, body mass index equaling overweight, and
medication used to control high blood pressure and diabetes. Conversely, the patient population
might be interested in issues such as quality of life and familial history. These patient selected
features are included in the CMV. The data reduction step uses the CMV as input.

4

Figure 1. Conceptual model for matching individuals.

3.2 Rule-Based Filtering
The first step in the conceptual model is to filter out individuals using a rule set. The
rules are declarative statements that in affect constrain the individuals that may be used for
matching. A rule is a declarative statement as shown in equation 4. The predicates of R, (P1,P2,
…,Pj), are operators used to express the logic of the filter. The operators are typically {>, <, ≠, =,
≥, ≤}. Filtering is O(N), where N is the number of records in the dataset.
𝑅: 𝐼𝑓 (𝑃1 ⋀ 𝑃2 … ⋀ 𝑃𝑗 ) 𝑡ℎ𝑒𝑛 {𝑅𝑒𝑡𝑎𝑖𝑛 | 𝐷𝑒𝑙𝑒𝑡𝑒} (4)
Filtering is computed in two ways. First, a database is filtered according to demographic
information such as age ranges, gender, and geography. Secondly, the database is filtered
according to temporal criteria delineating when medical events or measurements must occur. For
example, a CMV containing elevated total cholesterol may be grouped with an individual having
a similar diagnosis during the same time. Total cholesterol measurements less than 200 are
considered desirable [17]. Figure 2 illustrates this situation with a temporal overlap between two
individuals based on a similar total cholesterol value.

TCHOL=
Potential Match
265

TCHOL= TCHOL=
Candidate
185 260

t→
Figure 2. Simple events with temporal overlap.

3.3 Data Reduction
The third step in the computational model is data reduction. Data reduction is used to
improve efficiency by reducing the number of measurements used to compute similarity.
Principal Component Analysis (PCA) is used specifically for data reduction [6] and has been used
in health research [18].

5

PCA takes independent measurements and reduces them to a smaller set of elements
known as principal components (PC). The PCs are uncorrelated and represent most of the
information in the original set of measurements [7]. The goal of PCA is to summarize the
interrelationships for a set of measurements with a smaller set of uncorrelated orthogonal PCs that
are linear combinations of the original measurements [19]. The PCs explains the maximum
amount of variance possible in the observed measurements with a smaller set of linearly
transformed variables [6, 7]. If only a few principal components explain a high proportion of the
variance in the observed variables and only a few of the measurements are highly correlated with
these PCs, than the dataset can be reduced with a small loss of information.
PCA results in a correlation matrix in which each element has a range of -1.0 to +1.0,
representing the correlation, rxy, between two elements. The higher the absolute value of rxy the
stronger the relationship is between two types of measurements. An absolute value of rxy between
.50 - .69 is a moderate strength of relationship, between .70 - .89 is considered a strong
relationship, and between .90 – 1.00 is considered a very strong relationship [18].
PCA also produces a solution to the characteristic equation of the correlation matrix.
Solving this equation results in eigenvalues and an eigenvector representing the variance in the
measurements and loadings associated with each item in the correlation matrix. The loadings
represent the correlation of an item with a PC. The sum of the loadings is equal to the total
variance that is explained by a PC. Similarly, since the total variance is known, the proportion of
the total variance explained by a PC is equal to the sum of the loadings on a PC divided by the
total variance, where the total variance is equal to the number of measurements [18].

3.4 Similarity
The last step in the conceptual model is similarity computation. The JC, TFCMC, and
GC coefficients are used and compared in the simulation. GC is appealing since it is computed
on the raw data and can use all measurement types directly. Conversely, a drawback of JC and
TFCMC is that they operate on binary datasets. Each measurement is recoded to a binary value
to accommodate this requirement. Similarity computation results in a value that assigns a value to
the degree of likeness between two objects.

3.4.1 Tolerance Ranges
A single measurement from two individuals, of the same data type, can be an exact
match. However, these two values may differ but be considered equivalent from a clinical
perspective. For example, an individual with a blood pressure of 110/80 and another with 115/80
would both have normal blood pressure. However, if JC or TFCMC is used, than these two
individuals would not be considered a match unless some procedure is used to account for the
blood pressure readings being essentially the same.
There are two approaches to this problem. The first approach is to define a percentage-
based tolerance range (PBTR). A PBTR is determined by a tolerance level, τ, which is defined
for the set of measurements. The tolerance level establishes a lower and upper value for each
measurement. This establishes the range of values for a measurement that are considered equal to
that in the CMV. As shown in equation 5, the tolerance range for the jth measurement is
determined by the value of that measurement for the CMV, c, and the tolerance level τ.

𝑇𝑗 = (𝑚 𝑐𝑗 ∗ (1 − 𝜏), 𝑚 𝑐𝑗 ∗ (1 + 𝜏)) (5)

6

For example, assume a tolerance of 20% is used for body weight. If the CMV has a body
weight measurement of 200 pounds, than the PBTR for body weight is Tj = (180, 220). Thus, an
individual with a body weight in this range is considered similar to the CMV for this feature.
Conversely, someone with a body weight of 245 is not considered similar to the CMV for this
feature.
The second approach is to set a cut point tolerance range (CPTR) for each of the medical
measures. Often a medical measure has a clinically relevant cut-point, which establishes a
threshold between healthy and un-healthy values. For example, the National Heart Lung and
Blood Institute Obesity Education Initiative defined six classifications for body mass index
(BMI). These classifications are cut points ranging from less than 18.5 kg/m2 for underweight,
18.5 - 24.9 kg/m2 for normal weight, to greater than or equal to 40 kg/m2 for extreme obesity
[20]. Thus, for a BMI value of 22 the CPTR Tj = (18.5, 24.9).
Both the PBTR and the CPTR approaches can be applied to interval and ratio data. For
ordinal data, a tolerance range can be chosen as a range on the ordinal scale of potential values.
For example, Figure 3 illustrates a question on mental health. The responses are ordered in
ascending order of intensity. If the CMV includes item response two to this question, than
grouping would be with people who have the same response or perhaps a subset of the possible
categories. For instance, the tolerance set might be categories 2 and 3, represented as Tj = (2, 3).

Figure 3. Ordinal measurement type.

For nominal data, there are two approaches. First, each response category of a nominal
data item may be converted into an independent item. For example, if the nominal data item is a
checklist of the prescription medications used by an individual this can be converted into 10
binary data items on the usage of each specific medication (e.g., using Lipitor / not using Lipitor,
using aspirin / not using aspirin). Disuniting each element of a nominal data item in this manner
has the possibility of overwhelming the similarity computation. An alternative approach for
nominal data is to associate a tolerance with this feature such as "X out of the Y nominal
categories must be the same" for the binary data item to show agreement. This would preclude
the possibility of overwhelming the similarity computation by a disunited single nominal
variable.

3.4.2 Similarity Computation
For a dataset of individuals I = {I1, I2, … In} each with a set of measurements M = {m1,
m2, … mk} an NxN similarity matrix can be computed between each pair of objects. This is
O(N2). The computation can be simplified under three conditions. First, pair-wise computation
of an object with itself is (i.e., on the diagonal) is not needed. Second, it is reasonable to assume
that there is a symmetric relationship between two objects, thus S(A,B) = S(B, A). Under these
two conditions, the computation is reduced to the lower half of the matrix and thus there are

7

N2  N
computations. Note, that the objective of this work is to match similar individuals. As
2
such, the computation can be reduced to O(N) since only the similarity coefficient between the
CMV and the list of individuals is computed.

3.4.3 Simulation Dataset
The United States National Institutes of Health (NIH) and the United States Centers for
Disease Control and Prevention (CDC) operate clinical trials, cross-sectional studies, and
surveillance activities either through intramural or extramural research. For the purposes of this
work the dataset must be public use, contain a large number of individuals, and contain a variety
of measures. Therefore, data from the National Health and Nutrition Examination Survey
(NHANES) have been selected.
NHANES is a nationally representative cross-sectional survey of the non-institutionalized
population of the United States. Each year the NHANES enrolls approximately 5,000 individuals
of all age ranges, genders, race, and ethnicities. Study participants participate in an interview in
their home. After the home interview, a participant receives an extensive physical exam at one of
three mobile examination centers. Content on the study includes cardiovascular disease,
environmental exposures, eye disease, kidney disease, obesity, physical fitness, physical
functioning, and many other health indicators [21, 22].

3.4.4 Missing Data
Surveys such as NHANES may have missing data for some individual's measurements.
This can arise because individuals refuse to participate in the survey or because they refuse to
participate in portions of the survey [23]. Missing data affects two elements of the computational
model. First, it affects the data reduction piece, as PCA requires complete records for
computation. However, PCA will automatically remove incomplete records to determine the
variance structure.
Secondly, similarity computation needs to account for missing data. Conceptually, it is
unknown if a measurement is missing because it was never observed or recorded, it is a feature
that does not exist for an individual, or some other reason. The reasons for missing data are not
encoded in the NHANES database and therefore it cannot be concluded that a person with a
missing measurement has a value similar to the CMV. In this research, missing data is re-coded
to NULL and is considered different from another person’s measurement.

3.5 SHN Simulation
Publicly available data from NHANES 1999-2003 is used in the simulation. The dataset
includes 31,124 individuals at birth age and older. This dataset comprises measures related to
self-report questions on health, physical measures, and the results of laboratory tests [24, 25, 26].
The simulation is evaluated on type 2 diabetes. Tables 3 and 4 describe the data items and the
data files used for the simulation.

3.5.1 Type 2 Diabetes
Type 2 diabetes (T2D) usually occurs in individuals who are older, obese, or lacking in
physical activity. It occurs as insulin resistance such that the muscle, liver, and fat cells do not
use insulin properly. As a result, the body needs additional insulin to get glucose into cells for

8

energy [27]. T2D can be controlled with healthy eating habits, physical activity, weight loss, and
for some individuals, with the use of medications [28].
A primary risk factor for T2D is age, with those individuals over 45 being at increased
risk. Some other risk factors associated with type 2 diabetes are abdominal obesity, ethnicity,
HDL values lower than the normal range, history of gestational diabetes, hypertension, insulin
resistance, overweight, physical inactivity, and a family history of diabetes [29, 30].
Symptoms of T2D include infections, blurry vision, and tingling or numbness in the
hands and feet [31]. There are numerous health effects resulting from diabetes such as cataracts,
glaucoma, or retinopathy; foot ulcers, amputations; hearing loss; heart disease, or hypertension;
nervous system diseases; skin infections; or stroke [31, 32, 33].
Diabetes is diagnosed with a fasting plasma glucose (FPG) test, a regular plasma glucose
test or an oral glucose tolerance test (OGTT). All three tests assess the level of glucose in the
blood. A normal value is less than 100 mg/dL for people without diabetes. Values between 100
and 125 mg/dL is labeled as "impaired fasting glucose", while values greater 125 mg/dL are given
a label of "provisional diagnosis of diabetes". A non-fasting plasma glucose test may also be
used. If the value from this test is above 200 mg/dL, than an individual may have diabetes.
Confirmatory tests are usually required [34, 35, 36].
T2D is monitored with laboratory tests such as total cholesterol, HDL cholesterol, LDL
cholesterol, triglycerides, and insulin [37]. Many of the T2D related self-reported questions,
physical measures, and laboratory tests are available in the NHANES dataset.

3.5.2 Simulation Software
The computational model and software for the simulation runs on a Hewlett-Packard
model p6210y personal computer with an AMD Athlon ™ II X4 620 Processor. The processor
runs at 2.60 GHz and there is 6GB of installed RAM. Windows 7 64-bit operating system is
installed on the personal computer. Filtering and data reduction is computed with software
written in the SAS Statistical Software v9.1. Similarity is computed with software written in
Java.

4 Results
The dataset was prepared by merging several datasets from NHANES 1999-200,
NHANES 2001-2002, and NHANES 2003-2004. As shown in Table 3, the dataset includes 28
medical measurements. One can imagine that a research scientist studying T2D would select the
items in this dataset. Perhaps the patient population would select items related to family history
and pain. Therefore, both the researcher and patients can influence the matching process without
affecting the conceptual model.
The simulation is examined from two different perspectives: 1) reduction in the record
and feature space resulting from filtering and PCA, and 2) the correlation between the three
similarity coefficients.

4.1 Filtering
T2D occurs mostly in adults, thus the datasets were filtered in the first stage for
individuals ages 20 and above. This resulted in the original dataset of 31,124 individuals being
reduced to 49.2% of the original size. The dataset does not include temporal information due to
confidentiality and disclosure concerns. Therefore, temporal matching is not utilized for this
problem.
9

4.2 Data Reduction
The second step in the process is to conduct the principal component analysis (PCA) to
reduce the scale of the feature space (i.e., medical measures). Figure 4 shows the value of the
principal components for T2D. The first 11 principal components (PC) are greater than 1.0.
Figure 5 shows the unique and cumulative proportion that each PC contributes to the overall
variance. The first 11 PCs uniquely contribute between 3.8% and 12.2% of the overall variance.
In addition, the T2D PCs cumulatively contribute 70.7% to the overall variance. Thus, following
the criteria for selection of PCs the first 11 T2D PCs are used for data reduction.

6

4

2

0
1 5 9 13 17 21 25
Principal Component #

Figure 4. Type 2 diabetes principal component values.

100%

80%

60% Unique Proportion

40% Cumulative Proportion

20%

0%
1 5 9 13 17 21 25
Principal Component

Figure 5. Type 2 diabetes principal component unique and cumulative
proportions.

Figure 6 shows 18 of the original 28 measures related to T2D. Fourteen of these
measures have a loading of 0.70 or greater on a PC. Four measures are loaded very close to 0.70
and are thus retained. Thus, PCA reduces the measurement space for T2D by 35.7%.

10

1
0.8
0.6
0.4
0.2
0

LBXSKSI
LBXSGL
LBXGLU

BPXSY1
URXUMA

DIQ070

DIQ080
LBXHCT

LBDLDL
LBDHDL

LBXSPH
LBXHGB

BMXBMI

FAMDIA
LBXGH

LBXTC

LBXTR
BMXWAIST
Figure 6. Type 2 diabetes loadings.

4.3 Similarity
Similarity coefficients are computed in the third step of the model. TFCMC, JC, and GC
are used. For the TFCMC and JC the binary datasets are computed with PBTRs of 5%, 10%,
25%, and 50% as shown in Table 1. The PBTRs for each measurement (i.e., variable) are
calculated as described in Equation 5. Thus, in the T2D example the CMV has a body mass index
measurement (BMI) of 27 and a 5% PBTR of (26.201, 28.959). As the tolerance level increases
the tolerance range around each measure becomes larger. For categorical data, individual
categories may be selected; for ordinal data, ranges may be selected. FAMDIA is an example of a
categorical measurement, which can be coded, with value of zero or one. A zero represents a
CMV without a family history of diabetes. In T2D example, the FAMDIA PBTR range across all
tolerance levels is essentially (0,0).
For the CPTR approach, the tolerance range used is one that is medically relevant. For
example, the CMV has a BMI measurement of 27.5 and a systolic blood pressure reading of 120.
The literature describes a BMI of 27.5 to be in the overweight classification range of 25.0 – 29.9
[20]. Thus, the CPTR for BMI is (25, 29.9). Similarly, systolic blood pressure is considered
normal if it is less than or equal to 120 mmHg. The CMV blood pressure is exactly 120, so the
CPTR can be set as less than or equal to 120 mmHg.
Table 1 also delineates the CPTRs. For several measurements, the literature describes a
CPTR delineating healthy and unhealthy levels (refer to the references noted in Table 1). Some
measurements do not have a specific set of cut points for healthy and unhealthy values. Instead,
these measurements have a reference range that denotes where the values of the measurement fall
for a large percentage of the population. All reference ranges for these measurements are
consistent with the CMV age and are inclusive of differences between males and females.
Table 2, Figure 7, and Figure 8 illustrate the descriptive statistics for the example.
TFCMC can produce negative similarity scores when the majority of measurements between the
CMV and an individual are dissimilar. In both examples, the similarity score at each percentile
increases as the PBTR tolerance level increases. For example, at the 5% tolerance level and 95th
percentile, the TFCMC similarity score results in a value of negative six; and at the 50%
tolerance level and 95th percentile TFCMC has a similarity score of 14. Thus, higher similarity
scores occur by increasing the tolerance level around a measurement. One must be careful in
setting the tolerance level because high similarity scores can result between the CMV and an
individual who is in all likelihood dissimilar. In addition, the cut point tolerance ranges produce
similarity scores at the different percentiles that fall between the 10% and 50% tolerance level.

11

20
15
10

Similarity Score
5
0
-5
-10
-15
-20
Min 25th % 50th % 75th % 95th % Max Mean

5% PBTR 10% PBTR 25% PBTR
50% PBTR CPTR

Figure 7. Type 2 diabetes TFCMC descriptive similarity statistics.

1
0.9
0.8
Similarity Score

0.7
0.6
0.5
0.4
0.3
0.2
0.1
0
Min 25th % 50th % 75th % 95th % Max Mean

5% PBTR 10% PBTR 25% PBTR
50% PBTR CPTR GC

Figure 8. Type 2 diabetes JC and GC descriptive similarity statistics.

Figure 9 illustrates the correlation coefficients between each combination of similarity
coefficients at PBTRs of 5%, 10%, 25%, and 50% and the correlation coefficient for the CPTR.
This figure shows that the correlation strength between (TFCMC, GC) and (JC, GC) increases as
the tolerance level increases. Note however that (TFCMC, JC) are strongly correlated at all
PBTRs and the CPTR. For (TFCMC, GC), and (JC, GC) the correlation coefficient for CPTR is
between the correlation coefficients at the 25% and 50% tolerance levels.

12

1.0

0.8

0.6

R
0.4

0.2

0.0
TFCMC, JC TFCMC, GC JC, GC

5% 10% 25% 50% CPTR

Figure 9.Correlation coefficients associated with type 2 diabetes similarity
coefficients.

5 Discussion
The purpose of this paper is to propose a conceptual model for grouping similar
individuals together, based on their medical measurements, and demonstrate it with an example.
The conceptual model consists of candidate measurement vector (CMV) selection, rule-based
filtering, principal component analysis (PCA) data reduction, and similarity computation.
Different techniques for computing similarity were compared. This research is significant
because, to date, a conceptual model for the purpose of automatically grouping individuals for
health research has not been defined.
The simulation uses a publicly available dataset and successfully demonstrates that the
scale of the problem, in terms of the number of observations and feature space, can be reduced
using filtering and principal component analysis (PCA). In the example chosen, filtering for a
specific age range reduced the number of observations by about one-half. This will vary based on
the filtering critera and the population of individuals in the dataset. The feature space was
reduced from 28 to 18 medical measurements using PCA, a reduction of 35.7%.
The mean similarity scores for TFCMC, JC, and GC all increased as the PBTR increased.
The increased scores imply a higher degree of likeness between the CMV and each of the other
observations (i.e., individuals) in the dataset. The mean similarity score for TFCMC is low at all
PBTRs and with the CPTR. It should be noted however, that a higher similarity score is balanced
against the tolerance level used with PBTRs. Using a high tolerance level may in practice bring
dissimilar individuals together into an RHN. Therefore, caution is recommended in setting the
tolerance level.
The strong correlation between TFCMC and JC is an unexpected finding as TFCMC
lowers the similarity score due to disimilar measurements. However, JC in some sense takes
dissimilar features into account as well in the denominator (refer to equation 1). Thus, the two
similarity coefficients track together and are thus correlated. This may not be the case however
if TFCMC is weighted.
Similarity computation showed strong positive correlations between JC and GC for
PBTRs of 10%, 25%, 50%, and CPTRs. For PBTRs the correlation was a little bit below a
moderate correlation. Using a 50% PBTR is not likely to be a good approach as it may result in
13

ranges that cross over many cut points of healthy values for a specific medical measurement. For
example, one person with an unhealthy blood pressure level might pool with people who have
healthy blood pressure levels.
The similarity score results highlight two points. First, in the case of TFCMC and JC it is
important to establish a threshold similirty score for grouping individuals. This may be based on
a minimum number of measurements that are considered the same. Arbitrary assignment of a
threshold value should be avoided. Intuitively, one might consider that at least half the
measurements should be equivalent. This would establish a TFCMC floor of zero and a JC floor
of 0.50. An alternative approach is to consider the statistical distribution of the similarity scores
and choose those scores at the 95th percentile or higher. In practice, the assignment of a
threshold may be based on empirical evidence. Secondly, GC scales each measurement by the
range and is conceptually appealing as it is desgiend to work with mixed data types. It is true that
as the GC score increases two individuals are considered more similar. However, it is not clear
how the scores are to be interpreted and thus GC presents a problem. Moreover, the
interpretation of the GC similarity score is not as intuituive as JC and TFCMC.

6 Conclusion
Developing a conceptual model for matching individuals with the appropriate research
program is an important contributor to improving the research process and engaging individuals.
While research programs have selected individuals for participation in their programs for many
years, it is plausible to re-think this approach to improve matching of a study respondent and
researcher. Therefore, this paper proposes a conceptual model that automatically groups
individuals based on a filtering the data space, reducing the feature space with PCA, and
computing the likeness between individuals with similarity coefficients. An example was used to
simulate the conceptual model, and illustrate the effectiveness of filtering and PCA in reducing
the scale of the problem. Based on the results, two next steps include evaluation of the
conceptual model with a large-scale problem and temporal filtering to refine the matching.

References
[1] Blumenthal D. Launching HITECH. New England Journal of Medicine, vol. 362, no. 5,
February 2010, pp. 382-385.
[2] Halamka JD, Mandl KD, Tang PC. Early experiences with personal health records.
Journal of the American Medical Informatics Association, vol. 15, no. 1, Jan / Feb 2008, pp. 1-7.
[3] Krantz DH, Luce RD, Suppes P, and Tversky A. Foundations of Measurement: Volume
1, Additive and Polynomial Representations. Dover Publications, Mineola, NY, 1999.
[4] McCall RB. Fundamental Statistics for Psychology. Second Edition, Harcout Brace
Jovanovich, Inc. New York. 2nd Edition, 1975. pp. 6-9.
[5] Friedman CP, Wyatt JC. Evaluation Methods in Medical Informatics. Springer-Verlag,
New York, 1997, pp. 107-108.
[6] Fodor IK. A Survey of Dimension Reduction Techniques. U.S. Department of Energy,
Lawrence Livermore National Laboratory, UCRL-ID-148494. May, 9, 2002.
[7] Dunteman GH. Principal Component Analysis, Series: Quantitative Applications in the
Social Sciences, Sage Publications, 1989, Newbury Park, CA.
[8] Gower JC. A general coefficient of similarity and some of its properties. Biometrics,
December 1971, vol. 27, pp. 857-874.
14

[9] Yin Y and Yasuda K. Similarity coefficient methods applied to cell formation problem: a
comparative investigation. Computers & Industrial Engineering, 2005, vol. 48,pp. 471-489.
[10] Reif, JC, Melchinger, AE, Frisch, M. Genetical and Mathematical Properties of Similarity
and Dissimilarity Coefficients Applied in Plant Breeding and Seed Bank Management Crop Sci,
2005, vol. 45, pp. 1-7.
[11] Willett P. Similarity-based virtual screening using 2D fingerprints. Drug Discovery
Today, December 2006, vol 11, no. 23/24, pp. 1046-1053.
[12] Kosman E., Leonard KJ. Similarity coefficients for molecular markers in studies of
genetic relationships between individuals for haploid, diploid, and ployploid species. Molecular
Ecology, 2005, vol. 14, pp. 415-424.
[13] Goodall DQ. A new similarity index based on probability. Biometrics, December 1966,
pp. 882-907.
[14] Jaccard P. The distribution of the flora in the alpine zone. The New Phytologist, vol. XI,
no. 2, pp. 37-50, Feb. 1912.
[15] Alderderfer MS and Blashfield RK. Cluster Analysis, Series: Quantitative Applications
in the Social Sciences. Series/Number 07-044. Newbury Park: Sage Publications, 1984.
[16] Tversky A. Features of Similarity. Psychological Review, July 1977, vol. 84, no. 4, pp.
327 – 352.
[17] National Cholesterol Education Program. Detection, Evaluation, and Treatment of High
Cholesterol in Adults (Adult Treatment Panel III): Executive Summary. U.S. Department of
Health and Human Services, NIH Publication No. 01-3670, May 2001, pp. 3.
http://www.nhlbi.nih.gov/guidelines/cholesterol/atp3xsum.pdf. Accessed on April 6, 2010.
[18] Pett MA, Lackey NR, Sullivan JJ. Making Sense of Factor Analysis: The Use of Factor
Analysis for Instrument Development in Health Care Research. Sage Publications, Thousand
Oaks, California, 2003.
[19] Goddard J and Kirby A. An introduction to factor analysis. Norwich, UK: Geo
Abstracts, 1976.
[20] The Practical Guide Identification, Evaluation, and Treatment of Overweight and Obesity
in Adults. U.S. Department of Health and Human Services, Public Health Service, National
Institutes of Health, National Heart, Lung, and Blood Institute. NIH Publication No. 00-4084.
October 2000. Available at http://www.nhlbi.nih.gov/guidelines/obesity/prctgd_c.pdf. Accessed
on January 4, 2011.
[21] About the National Health and Nutrition Examination Survey (NHANES). United States
Centers for Disease Control and Prevention, National Center for Health Statistics.
http://www.cdc.gov/nchs/nhanes/about_nhanes.htm. Accessed on April 6, 2010.
[22] National Health and Nutrition Examination Survey: 1999-2010 Survey Content. United
States Centers for Disease Control and Prevention, National Center for Health. Statistics.
http://www.cdc.gov/nchs/data/nhanes/survey_content_99_10.pdf. Accessed April 6, 2010.
[23] Brick JM and Kalton G. Handling missing data in survey research. Stat Methods Med
Res. September 1996, vol. 5, pp. 215-238.

15

[24] National Health and Nutrition Examination Survey: NHANES 1999-2000. Centers for
Disease Control and Prevention. http://www.cdc.gov/nchs/nhanes/nhanes1999-
2000/nhanes99_00.htm. Accessed on January 4, 2011.
[27] Diagnosis of Diabetes. National Institutes of Health, National Institute of Diabetes and
Digestive and Kidney Diseases. http://diabetes.niddk.nih.gov/dm/pubs/diagnosis/index.htm.
Accessed on January 4, 2011.
[28] National Diabetes Fact Sheet, 2007. Centers for Disease Control and Prevention.
http://www.cdc.gov/diabetes/pubs/pdf/ndfs_2007.pdf. Accessed on January 4, 2011.
[29] Medline Plus: Type 2 Diabetes - Risk Factors. National Institutes of Health, National
Library of Medicine. http://www.nlm.nih.gov/medlineplus/ency/article/002072.htm. Accessed
on January 4, 2011.
[30] Diabetes Health Center: Risk Factors for Diabetes. WedMD.
http://diabetes.webmd.com/risk-factors-for-diabetes. Accessed on January 4, 2011.
[31] Diabetes Basics: Symptoms. American Diabetes Association.
http://www.diabetes.org/diabetes-basics/symptoms/. Accessed on January 4, 2011.
[32] Living with Diabetes: Complications. American Diabetes Association.
http://www.diabetes.org/living-with-diabetes/complications/. Accessed on January 4, 2011.
[33] Complications of Diabetes. National Institutes of Health, National Institute of Diabetes
and Digestive and Kidney Diseases. http://diabetes.niddk.nih.gov/complications/. Accessed on
January 4, 2011.
[34] Diabetes Guide: Diabetes Testing. WebMD.
http://diabetes.webmd.com/guide/diagnosing-type-2-diabetes. Accessed on January 4, 2011.
[35] Mayfield, J. Diagnosis and Classification of Diabetes Mellitus: New Criteria. American
Family Physician. http://www.aafp.org/afp/981015ap/mayfield.html. Accessed on January 4,
2011.
[36] American Diabetes Association. Position Statement: Diagnosis and Classification of
Diabetes Mellitus. Diabetes Care. Volume 27, Supplement 1, January 2004, pp. s5-s10.
http://care.diabetesjournals.org/content/27/suppl_1/s5.full.pdf+html. Accessed on January 4,
2011.
[37] Diabetes. Lab Tests Online.
http://www.labtestsonline.org/understanding/conditions/diabetes-6.html. Accessed on January 4,
2011.
[38] Healthy Weight - it's not a diet, it's a lifestyle!: About BMI for Adults.
http://www.cdc.gov/healthyweight/assessing/bmi/adult_bmi/index.html. Accessed on January 4,
2011.

16

[39] Weight-control Information Network: Weight and Waist Measurement: Tools for Adults.
National Institutes of Health, National Institute of DIabetes and Digestive and Kidney Disease.
http://www.win.niddk.nih.gov/publications/tools.htm#circumf. Accessed on January 4, 2011.
[40] Medline Plus: High Blood Pressure. National Institutes of Health, National Library of
Medicine. http://www.nlm.nih.gov/medlineplus/highbloodpressure.html. Accessed on January 4,
2011.
[41] Tietz NW. Clinical Guide to Laboratory Tests. 3rd Edition. Edited by Norbert W.
Tietz. W. B. Saunders Company, Philadelphia, 1995.
[42] Diabetes Health Center: Blood Glucose. WebMD. http://diabetes.webmd.com/blood-
glucose?page=3. Accessed on January 4, 2011.
[43] Diabetes Health Center: Microalbumin Urine Test. WebMD.
http://diabetes.webmd.com/microalbumin-urine-test?page=2. Accessed on January 4, 2011.
[44] Diabetes Health Center: Hyperglycemia and Diabetes. WebMD.
http://diabetes.webmd.com/diabetes-hyperglycemia. Accessed on January 4, 2011.

17

APPENDIX A
Table 1. Percentage-based and clinically relevant cut-point tolerance ranges for type 2 diabetes measures.
τ = 5% τ = 10% τ = 25% τ = 50%

Variable CMV Value Cut-Point Tolerance Ranges Min Max Min Max Min Max Min Max

BMXBMI 27.5 ≥ 25 is overweight, thus 25 - 29.9 is used [20]. 26.2 28.9 24.8 30.3 20.6 34.4 13.7 41.3
Higher risk category is ≥ 88 for women and ≥ 101 for
BMXWAIST 101 95.9 106.0 90.9 111.1 75.7 126.2 50.5 151.5
men, thus ≥ 88 is used [38, 39].
BPXSY1 120 ≤ 120 normal [40] 114 126 108 132 90 150 60 180
LBXGH 7.4 > 5.2% [41] 7.0 7.77 6.6 8.14 5.5 9.25 3.7 11.1
LBXGLU 178.4 > 99 is abnormal [42] 169.4 187.3 160.5 196.2 133.8 223 89.2 267.6
LBXTC 167 < 200 is normal [41] 158.6 175.3 150.3 183.7 125.2 208.7 83.5 250.5
LBDHDL 32 < 35 is at risk [41] 30.4 33.6 28.8 35.2 24 40 16 48
LBXTR 218 < 250 is desirable [41] 207.1 228.9 196.2 239.8 163.5 272.5 109 327
LBDLDL 92 < 130 is desirable [41] 87.4 96.6 82.8 101.2 69 115 46 138
URXUMA 26.4 ≥ 20 is abnormal [43] 25.0 27.7 23.7 29.0 19.8 33 13.2 39.6
LBXSGL 179 > 180 is abnormal [44] 170.0 187.9 161.1 196.9 134.2 223.7 89.5 268.5
Reference range is 2.8 - 4.1 for women and 2.3 - 3.7
LBXSPH 2.6 2.4 2.7 2.34 2.8 1.95 3.25 1.3 3.9
for men. Thus, 2.3 - 4.1 is used [41].
LBXSKSI 3.8 Reference range is 3.5 - 5.1 [41]. 3.6 4.0 3.5 4.27 2.91 4.86 1.9 5.8
Reference range is 11.7-16.0 for women and 13.1-17.2
LBXHGB 17 16.1 17.8 15.3 18.7 12.7 21.2 8.5 25.5
for men. Thus, 11.7 - 17.2 is used [41].
Reference range is 35-47 for women and 39 - 50 for
LBXHCT 51.1 48.5 53.6 45.9 56.2 38.3 63.8 25.5 76.6
men. Thus, 35 - 50 is used [41].
DIQ080 1 1 1 1 1 1 1 1 1 1
DID060MN 0 0 0 0 0 0 0 0 0 0
FAMDIA 0 0 0 0 0 0 0 0 0 0

18

Table 2. Type 2 diabetes descriptive statistics.
Percentile
Tolerance Similarity
Min 25 50 75 95 Max Mean
Level Measure
TFCMC -18 -14 -12 -10 -6 2 -11.02
5%
JC 0 0.111 0.166 0.222 0.333 0.555 0.193
TFCMC -18 -10 -8 -6 -2 6 -8.365
10%
JC 0 0.222 0.277 0.333 0.444 0.667 0.267
TFCMC -18 -4 0 2 4 16 -1.942
25%
JC 0 0.388 0.5 0.555 0.611 0.944 0.446
TFCMC -18 2 6 8 14 18 4.432
50%
JC 0 .556 .667 .722 .889 1 .0623
TFCMC -18 -6 -2 0 4 12 -2.940
Cut Points
JC 0 0.333 0.444 0.5 0.611 0.833 0.418
GC 0.056 0.673 0.717 0.841 0.881 0.963 0.689

Table 3. NHANES 1999-2003 data items that correspond to diabetes.

Measurement NHANES Measure Variable Name
Item # Measurement Notes
Area 1999-2000 2001-2002 2003-2004

Demographic 1 Gender RIAGENDR RIAGENDR RIAGENDR
Data 2 Age RIDAGEYR RIDAGEYR RIDAGEYR
3 Systolic Blood Pressure BPXSY1 BPXSY1 BPXSY1
Examination 4 Diastolic Blood Pressure BPXDI1 BPXDI1 BPXDI1
Data 5 Body Mass Index BMXBMI BMXBMI BMXBMI
6 Waist circumference BMXWAIST BMXWAIST BMXWAIST

19

Recoded to months on insulin
7 How long taking insulin DIQ060U/Q DIQ060U/Q DIQ060U/Q which is measure variable
name DID060MN
Take diabetic pills to lower blood
8 DIQ070 DIQ070 DIQ070
sugar
Diabetes affected eyes / had
9 DIQ080 DIQ080 DIQ080
retinopathy
Ulcers / sores not healed within 4
DIA090 DIA090 DIA090
weeks
Numbness in hands / feet past 3
DIQ100 DIQ100 DIQ100
months
Numbness in hands / feet or both DIQ110 DIQ110 DIQ110 Merged into 1 data item
10 reflecting pain / numbness /
Pain in hands / feet past 3 months DIQ120 DIQ120 DIQ120 tingling
Where was pain or tingling DIQ130 DIQ130 DIQ130
Pain in either leg while walking DIQ140 DIQ140 DIQ140
Questionnaire Pain in calf or calves DIQ150 DIQ150 DIQ150
Data
Mother with diabetes MCQ260AA MCQ260AA MCQ260AA
Father with diabetes MCQ260AB MCQ260AB MCQ260AB
Mat. grandmother with diabetes MCQ260AC MCQ260AC MCQ260AC
Pat. grandmother with diabetes MCQ260AE MCQ260AE MCQ260AE Merged into 1 data item
11 Mat. grandfather with diabetes MCQ260AD MCQ260AD MCQ260AD reflecting family history of
Pat. grandfather with diabetes MCQ260AF MCQ260AF MCQ260AF diabetes

Brother with diabetes MCQ260AG MCQ260AG MCQ260AG
Sister with diabetes MCQ260AH MCQ260AH MCQ260AH
Other relative with diabetes MCQ260AI MCQ260AI MCQ260AI
Mother with hypertension MCQ260FA MCQ260FA MCQ260FA
Merged into 1 data item
12 Father with hypertension MCQ260FB MCQ260FB MCQ260FB reflecting family history of
Mat. grandmother with hypertension
MCQ260FC MCQ260FC MCQ260FC
hypertension

20

Pat. grandmother with hypertension MCQ260FE MCQ260FE MCQ260FE
Mat. grandfather with hypertension MCQ260FD MCQ260FD MCQ260FD
Pat. grandfather with hypertension MCQ260FF MCQ260FF MCQ260FF
Brother with hypertension MCQ260FG MCQ260FG MCQ260FG
Sister with hypertension MCQ260FH MCQ260FH MCQ260FH
Other relative with hypertension MCQ260FI MCQ260FI MCQ260FI
13 Told to take medicine for BP BPQ040A BPQ040A BPQ040A
14 Glycohemoglobin LBXGH LBXGH LBXGH
15 High Density Lipoprotein LBDHDL LBDHDL LBXHDD
16 Hematocrit LBXHCT LBXHCT LBXHCT
17 Hemoglobin LBXHGB LBXHGB LBXHGB
18 Hepatitis C LBDHCV LBDHCV LBDHCV
19 Insulin LBXIN LBXIN LBXIN
20 Low Density Lipoprotein LBDLDL LBDLDL LBDLDL
Laboratory
21 Phosphorus LBXSPH LBDSPH LBXSPH
Data
22 Plasma Glucose LBXGLU LBXGLU LBXGLU
23 Potassium LBXSKSI LBXSKSI LBXSKSI
24 Serum Glucose LBXSGL LBXSGL LBXSGL
25 Total Cholesterol LBXTC LBXTC LBXTC
26 Triglyceride LBXTR LBXTR LBXTR
27 Urine Albumin URXUMA URXUMA URXUMA
28 White Blood Cell Count LBXWBCSI LBXWBCSI LBXWBCSI

21

Berman pcori challenge document

More Related Content

What's hot (19)

Viewers also liked (18)

Similar to Berman pcori challenge document (20)

More from Lew Berman (16)

Recently uploaded (20)

Berman pcori challenge document