A Network Learning Method for Functional Disability Prediction
from Health Data
Riccardo Dondi
1 a
and Mehdi Hosseinzadeh
1,2 b
1
Universit
`
a degli Studi di Bergamo, Bergamo, Italy
2
University of Calabria, Rende(CS), Italy
Keywords:
Network Analysis, Disability Classification, Learning Algorithms, Healthcare Analytics, Graph Data Mining.
Abstract:
This contribution proposes a novel network analysis model with the goal of predicting a classification of in-
dividuals as either ‘disabled’ or ‘not-disabled’, using a dataset from the Health and Retirement Study (HRS).
Our approach is based on selecting features that span health indicators and socioeconomic factors due to their
pivotal roles in identifying disability. Considering the selected features, our approach computes similarities
between individuals and uses this similarity to predict disability. We present a preliminary experimental eval-
uation of our method on the HRS dataset, where it shows an enhanced average accuracy of 62.48%.
1 INTRODUCTION
A relevant problem for supporting elderly individuals
is the prediction of their health status. In this context,
it is extremely valuable to predict the risk of function-
ally disability, in order to provide the needed support
(Stuck et al., 1999).
Current studies demonstrate the advancements in
the use of knowledge graphs and network analysis in
the fields of biology and healthcare (Hosseinzadeh,
2020; Hosseinzadeh et al., 2022) and (Pham et al.,
2018; Tao et al., 2020; Wang et al., 2020; Pham et al.,
2022; Cui et al., 2023). In this context, the develop-
ment of prediction models based on graphs in health-
care is essential for improving disease diagnosis and
reducing human error. In particular, (Wang et al.,
2020) developed a predictive model that classifies in-
dividuals according to their disability risk, using a
network to represent disease progression. (Tao et al.,
2020) introduced a novel classification model that
uses a heterogeneous knowledge graph for conceptu-
alizing medical domain knowledge. The developed
model was used to forecast possible health risks for
patients using data from the National Health and Nu-
trition Examination Survey (NHANES). (Cui et al.,
2023) provided a comprehensive review of knowl-
edge graph applications in healthcare, highlighting
the instruments, applications, and possibilities for
a
https://orcid.org/0000-0002-6124-2965
b
https://orcid.org/0000-0003-3275-6286
improved understanding and prediction of complex
medical scenarios.
Our contribution aims to build a prediction
method inspired by approaches for classifying indi-
viduals based on their risk of becoming disabled. Our
approach proposes a novel network analysis model
based on the features presented in a dataset from
the Health and Retirement Study (HRS)
1
(Health and
Study, 2008). We select some features that span
health indicators and socioeconomic factors due to
their pivotal roles in identifying disability. Each in-
dividual is then represented as a vector on the se-
lected features and similarity between two individuals
is evaluated by computing a function of the difference
in the values of the features. A user is then assigned
to the category (‘disabled’, meaning high-risk of be-
coming disable, or ‘non-disabled’, meaning low-risk
of becoming disable) based on the average similarity
with each single group (disable individuals and non
disable individuals).
We present some preliminary experimental eval-
uation of our method on the HRS dataset. We se-
lect 10 samples of 100 individuals extracted randomly
from the HRS dataset and on each of this sample we
evaluate the performance of our method. The method
shows a moderate accuracy in the classification (aver-
age accuracy of 62.48%).
The remainder of the paper is organized as fol-
lows. In section 2, we start by introducing some defi-
1
https://hrs.isr.umich.edu.
358
Dondi, R. and Hosseinzadeh, M.
A Network Learning Method for Functional Disability Prediction from Health Data.
DOI: 10.5220/0012991400003838
Paper published under CC license (CC BY-NC-ND 4.0)
In Proceedings of the 16th International Joint Conference on Knowledge Discovery, Knowledge Engineering and Knowledge Management (IC3K 2024) - Volume 1: KDIR, pages 358-362
ISBN: 978-989-758-716-0; ISSN: 2184-3228
Proceedings Copyright © 2024 by SCITEPRESS – Science and Technology Publications, Lda.

nitions and by providing the formal definition by for-
mally introduces the research problem. In Section 3,
we present the computational approach used to ad-
dress the research problem, including the construc-
tion and application of the bipartite graph model. In
Section 4, we present the results from the experimen-
tal analysis, discussing the implications and insights
gained from applying our methodology to the HRS
dataset. Finally, In Section 5, we conclude the main
outcomes with some future directions.
2 DEFINITIONS AND RESEARCH
PROBLEM
In this section, we define the main concepts needed
for our methodology, mainly graph theory and net-
work analysis, and we present the formal research
problem our study addresses.
All the graphs we consider in this paper are undi-
rected. A graph G is defined as a pair G = (V, E),
where V is a set of nodes and E is a set of edges. Each
edge e ∈ E is an unordered pair {v,w}, indicating a
connection between nodes v and w in V (Bondy and
Murty, 2008). We mainly consider bipartite graph,
defined in the following.
Definition 1. A graph is bipartite if there exist two
disjoint sets X ⊆ V and Y ⊆ V such that X ⊎Y = V
2
,
and every edge in E links a node of X and a node of
Y .
We now provide the definition of the neighbor-
hood of a node, which is a relevant concept needed
for describing our method.
Definition 2. Given a graph G = (V,E) and a node
v ∈ V , the neighborhood of v in G is defined as
N
G
(v) = {u ∈ V | {u, v} ∈ E}.
We now introduce a specific bipartite weighted
graph we consider to represent the relations between
features and individuals, called Feature-Individual
Classification Network (FICN).
Definition 3. The Feature-Individual Classification
Network is a bipartite weighted graph denoted as
G = (V,E,W
E
,W
F
) where:
• V is the set of nodes, partitioned into two disjoint
subsets V
I
and V
F
. The subset V
I
represents indi-
viduals, and V
F
represents features.
• E ⊆ V
I
×V
F
is the set of edges, each connecting
an individual node in V
I
and a feature node in V
F
.
• W
E
: E → R
+
is a weight function for the edges,
where each weight represents the value of an in-
dividual for a specific feature (these weights are
2
We recall that ⊎ denotes the disjoint union of sets.
derived from the input data, their computation is
described later).
• W
F
: V
F
→ R
+
is a weight function for the fea-
ture nodes, assigning a weight to each feature;
this weight represents the relevance of the specific
feature for classification. Unlike W
E
, W
F
is not
obtained from the input data but it is computed
using an optimization technique (that we will de-
scribe in Section 3).
Given an edge uv ∈ E, w
E
(uv) denotes the edge
weight of uv. Given a node u ∈ V
F
, w
F
(u) denotes the
feature weight of node u.
Note that G = (V, E,W
E
,W
F
) is not a complete bi-
partite graph as some edges may not defined, reflect-
ing possible missing data of individuals.
The classification of an individual is based on the
similarity value between the node related to the indi-
vidual, and the nodes of the Feature-Individual Clas-
sification Network, as defined in the following.
Definition 4. Let G = (V,E,W
E
,W
F
) be a Feature-
Individual Classification Network, where V = V
I
⊎V
F
.
Consider a candidate node c (note that c /∈ V ) such
that c is connected with a subset F
c
of feature nodes
(hence F
c
⊆ V
F
). The similarity measure σ(c,G) of c
with respect to G is defined as:
σ(c,G) =
∑
f ∈F
c
w
I
( f )z
c f
∑
f ∈F
c
w
I
( f )
,
where w
I
( f ) is the weight function of individual I for
feature f , and for each f ∈ N
G
(c), z
c f
is the z-score
3
of w
E
(c f ) in the following set:
{w
E
(u f ) : u ∈ N
G
( f )}.
Note that the similarity measure σ(c,G) is based
on the weight of the features that are computed by the
optimization technique described in Section 3.
2.1 Research Problem
Next we describe our problem. Given two disjoint
sets of individuals (‘disable’ and ‘non-disable’), we
define a Feature-Individual Classification Network
for each of these sets.
G
1
= (V
I
1
⊎V
F
,E
1
,W
E
1
,W
F
1
)
represents the graph consisting of the set V
I
1
of
individuals identified as ‘disabled’.
G
2
= (V
I
2
⊎V
F
,E
2
,W
E
2
,W
F
2
)
3
The z-score for feature f is computed as z
c f
=
v
c f
−µ
f
σ
f
,
where v
c f
is the value of feature f for the candidate node
c, µ
f
is the mean value of feature f across all the neigh-
bor nodes N
G
( f ) in G that have f , and σ
f
is the standard
deviation of f among the same nodes.
A Network Learning Method for Functional Disability Prediction from Health Data
359

represents the set V
I
2
of ‘non-disabled’ individuals.
We introduce now the main research problem we
consider in this paper, which aims to compute the fea-
ture weights in order to optimize the classification.
Problem 1. Weight Feature Optimization Problem.
Input: Two Feature-Individual Classification Net-
works G
1
= (V
I
1
⊎ V
F
,E
1
,W
E
1
,W
F
) and G
2
= (V
I
2
⊎
V
F
,E
2
,W
E
2
,W
F
).
Output: Compute W
F
: V
F
→ R
+
for the feature
nodes so that the classification in ‘disabled’ or ‘non-
disabled’ individuals is optimized.
Assume that the feature weights are known. For
each individual i
class
to be classified, we compute the
similarity σ(i
class
,G
i
), with i ∈ {1,2}. After calculat-
ing these similarity scores, the overall classification of
i
class
as either ‘disabled’ or ‘non-disabled’ is obtained
by aggregating these scores:
Classification(i
class
) = arg max
G∈{G
1
,G
2
}
σ(i
class
,G).
This aggregate score assigns i
class
to the group with
the highest computed similarity score. So, in order
to apply the classification, we need to compute the
feature weights.
The Weight Feature Optimization Problem is
solved by considering a training set of individuals for
which we already know the classification and then
compute the value of the weights W
F
in order to maxi-
mize the correct classification. Formally, we consider
the following problem.
Problem 2. Training Weight Feature Optimization
Problem.
Input: Two Feature-Individual Classification Net-
works G
1
= (V
I
1
⊎ V
F
,E
1
,W
E
1
,W
F
) and G
2
= (V
I
2
⊎
V
F
,E
2
,W
E
2
,W
F
); two sets X
1
, X
2
of candidate nodes
that are classified as disable and non-disable, respec-
tively.
Output: Compute W
F
: V
F
→ R
+
so that the number
of individuals of X
1
⊎ X
2
correctly classified is maxi-
mized.
3 METHODOLOGY
In the preliminary phase of our study, we define our
classification criteria based on the established guide-
lines from (Li et al., 2017; Rossetti and Cazabet,
2018). Specifically, an individual is considered ‘dis-
abled’ when encounters two or more difficulties in
any of the six identified Activities of Daily Living
(ADL). In order to address the classification problem,
we structure our data into two disjoint sets, i.e. a train-
ing set and a test set.
The training set consists of distinct subsets for dif-
ferent phases of the model development process. The
first subset of the training set consists of (1) 250 in-
dividuals randomly selected from the ‘disabled’ indi-
viduals and used to build the Feature-Individual Clas-
sification Network G
1
representing ‘disabled’ indi-
viduals, (2) 250 individuals randomly selected from
the ‘disabled’ individuals and used to build G
2
repre-
senting ‘non-disabled’ individuals. The second subset
consists of 100 ‘disabled’ individuals, and 100 ‘non-
disabled’ individuals used for the weight optimization
phase of our model. Note that the first and second
subsets are disjoint.
The test set is a subset consisting of individuals
whose disability status is also known; it is not utilized
to compute feature weights, but for assessing the ac-
curacy of our model. This set will be described in
Section 4.
In order to solve the Training Weight Feature Op-
timization Problem, we implemented an optimization
technique. This process starts with uniform initial
weights for each feature. Then we adopt a greedy al-
gorithm to incrementally adjust these weights, one at
a time. This process is mathematically formulated as
follows:
Initial Setup: The optimization starts with uniform
initial weights for each feature: W
F
( f ) = 1 for all f ∈
V
F
, where recall that V
F
is the set of all feature nodes.
Greedy Algorithm for Weight Adjustment: We
adopt a greedy algorithm to incrementally adjust these
weights, where each iteration focuses on changing the
value of a single feature weight. The adjustment pro-
cess is mathematically formulated as follows:
W
F
( f ) ← W
F
( f ) + ∆w
f
,
where ∆w
f
is the change in weight for feature f . After
applying this change, we evaluate its effectiveness on
the model’s classification accuracy.
Accuracy Assessment: The impact of each weight
adjustment is assessed by recalculating the classifica-
tion accuracy. Each individual i
class
is classified based
on the highest similarity score for the networks G
1
and G
2
.
Each individual i
class
in the dataset has a ‘Ground
Truth’ label, denoted by τ(i
class
), which indicates
whether the individual is ‘disabled’ or ‘non-disabled’.
After all individuals have been classified, we evaluate
the accuracy of the model by determining the fraction
of individuals that have been correctly classified ac-
cording to the ‘Ground Truth’. The accuracy of the
KDIR 2024 - 16th International Conference on Knowledge Discovery and Information Retrieval
360

Table 1: Features in the HRS Dataset That Have Been Used in This Study.
Variables
Years of education
Ever had cancer
Body Mass Index (BMI)
Ever drinks alcohol
Ever had high blood pressure
Total of all assets
Ever had lung disease
Ever had cancer
Ever had arthritis
Any difficulty-Using the toilet
Any difficulty-Walk across room
Any difficulty-Dressing
Any difficulty-Bathing or showering
Any difficulty-Eating
Any difficulty-Get in/out of bed
Table 2: Experimental Results Summary.
Experiment TP TN FP FN FPR FNR Accuracy (%)
1 25 31 18 21 0.37 0.46 58.95
2 27 34 16 22 0.32 0.50 61.62
3 25 30 19 23 0.39 0.48 56.70
4 30 35 14 20 0.29 0.40 65.66
5 29 34 16 20 0.32 0.41 63.64
6 27 31 19 21 0.38 0.44 59.18
7 28 38 12 21 0.24 0.43 66.67
8 30 38 12 19 0.24 0.39 68.69
9 25 31 19 25 0.38 0.50 56.00
10 28 37 11 20 0.23 0.42 67.71
Average 27 34 16 21 0.32 0.44 62.48
classification model is calculated as follows:
Accuracy =
1
N
N
∑
j=1

Classi f ication(i
class
j
) = τ(i
class
j
)

,
where N is the number of classified nodes, i
class
j
is
the node being classified, G
1
and G
2
are the two net-
works. Classification(i
class
j
), and τ(i
class
j
) are the pre-
dicted classification outcome and the ‘Ground Truth’
label for the j-th individual, respectively.
At each iteration, we increased the weight of a sin-
gle feature, evaluating the impact of this change on
the model’s accuracy. Then the following steps are
applied:
• If the accuracy of the model increases following
the weight change, we update the selected weight
with the change made. Then we randomly select
a weight feature to further explore potential im-
provements in model performance.
• If there is no improvement in accuracy the weight
is reverted to its previous value, and the adjust-
ment process randomly select a feature different
from the one that has been considered in this iter-
ation.
This process is repeated for all features, until the
method converges, that is each weight feature change
does not improve accuracy.
4 EXPERIMENTAL RESULTS
In this section, we present the outcomes of a series of
preliminary experiments to evaluate the performance
of our predictive model. For each experiment, a dis-
tinct random sample of 100 individuals was selected
from a larger dataset, which included 50 disabled in-
dividuals and 50 non-disabled individuals. It is im-
portant to note that some selected individuals (at most
5%) may have incomplete information available in
the dataset, hence they are not used for the validation
A Network Learning Method for Functional Disability Prediction from Health Data
361

phase.
We use RAND U.S. Health and Retirement Study
(HRS) data
4
. The used dataset comprises health sta-
tus and risk factor details from 42,406 survey partic-
ipants born between the years 1890 and 1995. The
features in the HRS dataset that were used in this re-
search are described in Table 1.
Table 2 presents the performance of our methods
for the data random samples from the test dataset.
The performance metrics considered include True
Positives (TP), True Negatives (TN), False Positives
(FP), False Negatives (FN)
5
, the False Positive Rate
(FPR), the False Negative Rate (FNR), and the over-
all accuracy in percentage. The FPR and FNR pro-
vide insights into the model’s tendency to categorize
negative and positive cases erroneously, which are re-
spectively calculated as: FPR = FP/(FP + T N), and
FNR = FN/(T P + FN). Finally, accuracy quantifies
the percentage of actual findings in the dataset that
match the ground truth.
In Table 2, experiment 10, which has the highest
accuracy at 68.69%, shows a balance between iden-
tifying true positives and true negatives while min-
imizing both false positives and false negatives. In
contrast, Experiment 5 shows the lowest accuracy,
indicating a higher misclassification rate. On aver-
age, these experiments have the accuracy 62.48%, and
across the 10 experiments, the model achieved a TP
rate of 27, a TN rate of 34, with FP and FN averag-
ing at 16 and 21, respectively. The average FPR was
observed at 0.32, with the FNR at 0.44.
In Table 2, a notable pattern across all experiments
is the higher number of TN compared to TP, and FN
compare to FP. This trend shows that the model has
a tendency to classify individuals as ‘not-disabled’.
In particular, the methods has better performances in
correctly identifying individuals who are not disabled
than it is at identifying those who are disabled.
5 CONCLUSION
This preliminary study explores feature weight opti-
mization for disability classification and shows how
learning and network approaches can be integrated
into healthcare frameworks in a potentially fruitful
way. We plain to compare the results of our method
with other prediction methods. Another possible fu-
4
https://hrs.isr.umich.edu.
5
The TP refers to when an individual’s ground truth is
‘not disabled’, but they are incorrectly classified as ‘dis-
abled’, and the FN refers to when an individual’s ground
truth is ‘disabled’, but they are incorrectly classified as ‘not
disabled’.
ture direction is to improve the ability to classify ‘dis-
abled’ individuals. Extending our dataset to include
a wider variety of demographic and geographic char-
acteristics is expected to enhance the generalizability
and relevance of our findings.
REFERENCES
Bondy, J. A. and Murty, U. S. R. (2008). Graph theory.
Springer Publishing Company, Incorporated.
Cui, H., Lu, J., Wang, S., Xu, R., Ma, W., Yu, S.,
Yu, Y., Kan, X., Ling, C., Ho, J., et al. (2023).
A survey on knowledge graphs for healthcare: Re-
sources, applications, and promises. arXiv preprint
arXiv:2306.04802.
Health and Study, R. (2008). Public use dataset. produced
and distributed by the university of michigan with
funding from the national institute on aging (grant
number nia u01ag009740).
Hosseinzadeh, M. M. (2020). Dense subgraphs in biologi-
cal networks. In International conference on current
trends in theory and practice of informatics, pages
711–719. Springer.
Hosseinzadeh, M. M., Cannataro, M., Guzzi, P. H., and
Dondi, R. (2022). Temporal networks in biology and
medicine: a survey on models, algorithms, and tools.
Network Modeling Analysis in Health Informatics and
Bioinformatics, 12(1):10.
Li, Z., Shao, A. W., and Sherris, M. (2017). The impact of
systematic trend and uncertainty on mortality and dis-
ability in a multistate latent factor model for transition
rates. North American Actuarial Journal, 21(4):594–
610.
Pham, T., Tao, X., Zhanag, J., Yong, J., Zhang, W., and
Cai, Y. (2018). Mining heterogeneous information
graph for health status classification. In 2018 5th
International Conference on Behavioral, Economic,
and Socio-Cultural Computing (BESC), pages 73–78.
IEEE.
Pham, T., Tao, X., Zhang, J., Yong, J., Li, Y., and Xie,
H. (2022). Graph-based multi-label disease prediction
model learning from medical data and domain knowl-
edge. Knowledge-based systems, 235:107662.
Rossetti, G. and Cazabet, R. (2018). Community discov-
ery in dynamic networks: a survey. ACM computing
surveys (CSUR), 51(2):1–37.
Stuck, A. E., Walthert, J. M., Nikolaus, T., B
¨
ula, C. J.,
Hohmann, C., and Beck, J. C. (1999). Risk factors for
functional status decline in community-living elderly
people: a systematic literature review. Social science
& medicine, 48(4):445–469.
Tao, X., Pham, T., Zhang, J., Yong, J., Goh, W. P., Zhang,
W., and Cai, Y. (2020). Mining health knowledge
graph for health risk prediction. World Wide Web,
23:2341–2362.
Wang, T., Qiu, R. G., Yu, M., and Zhang, R. (2020). Di-
rected disease networks to facilitate multiple-disease
risk assessment modeling. Decision Support Systems,
129:113171.
KDIR 2024 - 16th International Conference on Knowledge Discovery and Information Retrieval
362