Prediction of Response to Intra-Articular Injections of

Hyaluronic Acid for Knee Osteoarthritis

Eva K. Lee

1,2,3 a

, Fan Yuan

, Barton J. Mann

and Marlene DeMaio

4,5

Center for Operations Research in Medicine and Healthcare, The Data and Analytics Innovation Institute, Atlanta, U.S.A.

Georgia Institute of Technology, Atlanta, U.S.A.

AccuHealth Technologies, Atlanta, Georgia, U.S.A.

The American Orthopedic Society for Sports Medicine, Chicago, U.S.A.

Medical Corps, United States Navy, U.S.A.

Keywords: Knee Osteoarthritis, Injections of Hyaluronic Acid, Machine Learning for Evidence-Based Practice,

Branch-and-Bound, Particle Swarm Optimization.

Abstract: Osteoarthritis (OA) is a degenerative joint disease, with the knee the most frequently affected joint. Fifty

percent of knee OA patients eventually undergo surgical procedures such as knee replacement to address pain

and functional limitations. A significant number of these surgeries may be unnecessary, with intra-articular

injections of hyaluronic acid (HA) serving as a non-invasive, cost-effective alternative. Although research

studies have clearly demonstrated that HA improves knee function, the efficacy of this treatment remains

controversial. Many physicians have observed that effects depend on several patient characteristics such as

age, weight, gender, severity of the OA, and technical issues such as injection site and placement. In this study,

a multi-stage, multi-group machine learning model is utilized to uncover discriminatory features that can

predict the response status of knee OA patients to different types of HA treatment. The algorithm can identify

certain subgroups of knee OA patients who respond well to HA therapy. The baseline results, based on factors

such as patients’ weight, smoking status and frequency, identifies the patients most suitable for HA injection.

The model can achieve more than 89% blind prediction accuracy. The data derived from this study allows

physicians to administer HA products more selectively, resulting in a higher therapy success rate. Information

on the predicted responses could also be shared with patients beforehand to incorporate their values and

preferences into treatment selection. The model’s decision support tools also allow physicians to quickly

determine whether a patient is exhibiting at least the expected treatment response, and if not, to potentially

take corrective action. To the best of our knowledge, this work represents the first machine learning approach

that predicts patient responses to HA injections for knee osteoarthritis. The model is generalizable and can be

used to predict patient responses to other treatments and conditions.

1 INTRODUCTION

Osteoarthritis (OA) is a degenerative joint disease

that can affect the many tissues of the joint. It is one

of the most prevalent and costly chronic medical

conditions. affecting more than 32.5 million adults in

the United States (United States Bone and Joint

Initiative 2018). During 2019–2021, 21.2% of U.S.

adults (53.2 million) reported an arthritis diagnosis.

(Elgaddal, et al., 2022; Fallon, et. al., 2023) and by

2040, it is projected to increase to 78.4 million

Americans.

https://orcid.org/0000-0003-0415-4640

Arthritis increasingly is reported as the main

cause of disability among U.S. adults (Theis, K.A. et

al., 2019). Annual direct medical care expenditures

for osteoarthritis in the U.S. is estimated to exceed

$495.5 billion (United States Bone and Joint Initiative,

2019; Lo, et al., 2020). Worldwide, about 528 million

people were living with osteoarthritis in 2019 (WHO

2023, GBD 2019). It is estimated that those with OA

pain lost 31% of productive time at work due to

presenteeism and 8% due to absenteeism, compared

to 16% and 4%, respectively, for those who did not

report OA pain (Leifer et al., 2022).

Lee, E., Yuan, F., Mann, B. and DeMaio, M.

Prediction of Response to Intra-Articular Injections of Hyaluronic Acid for Knee Osteoarthritis.

DOI: 10.5220/0013071800003838

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 497-508

ISBN: 978-989-758-716-0; ISSN: 2184-3228

497

There is no known cure for OA. Instead,

treatments aim to reduce pain, maintain or improve

joint mobility, and limit functional impairment.

Treatments are usually non-operative, such as

physical therapy, rest, modification of daily activities,

analgesics, and anti-inflammatory medication. For

individuals who desire or require a high level of

physical activity, rest and activity reduction are not

viable treatment options. Oral non-steroidal anti-

inflammatory drugs (NSAIDs) are often

recommended, although frequent and serious adverse

effects of NSAIDs have been reported (Zhang et al.,

2010, Salis and Sainsbury, 2024). Over the past 25

years, intra-articular injection of hyaluronic acid (and

similar hyaluronan preparations) has emerged as an

additional tool for managing the symptoms of OA for

patients who fail to respond to other conservative

treatments. However, controversies exist regarding

its safety and efficacy, the number of injections and

courses, type of preparation, duration of its effects,

and combining it with other drugs or molecules

(Chavda et al., 2022). Other factors include patient

characteristics such as age, weight, gender, and

severity of the OA.

Knee OA happens when the cartilage in the knee

joint breaks down, enabling the bones to rub together.

The friction makes the knees hurt, become stiff, and

sometimes swell. Knee OA is a leading cause of

arthritis disability (Cui et al., 2020). Of significance

for sport medicine, heavy physical activity,

participation in high intensity contact sports,

participation in certain elite level sports, and knee

injury have all been linked to the development of knee

OA (Chan, et al., 2020; Driban, et al., 2017;

Lohmander, et al., 2007; McAlindon et al., 1999;

Sharma, 2001; Spector et al., 1996; Turner, et al.,

2000). Although it cannot be cured, treatments are

available to slow its progression and ease the

symptoms. Knee OA alone results in the loss of an

average of 13 days of work per year (versus 3 days

for those without Knee OA (Ayis & Dieppe, 2009).

Knee osteoarthritis affects more than 14 million

Americans, and its symptoms often lead to physical

inabilities, disabilities, and all sorts of inconveniences

for patients. It is estimated that knee osteoarthritis is

associated with approximately $27 billion in total

healthcare costs every year, with about 800,000 knee

surgeries performed annually. Specifically, 99% of

these knee replacements are done to address pain and

functional limitations (Barbour et al., 2017).

In a

multicenter longitudinal cohort study, it was reported

that about one-third of knee replacements may be

unnecessary (Riddle et al., 2014).

The management of knee pain depends on the

diagnosis, inciting activity, underlying medical

conditions, body mass, and chronicity. In general,

non-operative management is the mainstay of initial

treatment and includes rehabilitation, activity

modification, weight loss when indicated, shoe

orthoses, local modalities, and medication. The oral

medication often prescribed is an analgesic, usually

with anti-inflammatory properties. Supplements,

such as chondroitin sulfate and glucosamine, have

been shown to have a role. Since 1997, the regimen

has expanded to include viscosupplementation.

These agents are preparations of hyaluronic acid or

their derivatives (HA) which are sterilely injected into

the knee. Although research studies have clearly

demonstrated that HA improves knee function, the

efficacy of this treatment remains controversial.

Many physicians have observed that effects seem to

depend on several patient characteristics, such as age,

weight, gender, severity of the OA and technical

issues such as injection site and placement (Mora et

al., 2018).

This study aims to answer an important question:

whether different types of patients may respond

differently to HA treatment. Is it possible to identify

certain subgroups of knee OA patients who respond

well (or those who don’t) to HA therapy? Further, we

question whether it is possible prior to treatment to

predict a patient’s response to HA injections based on

patient and treatment characteristics. Physicians

could then make empirically informed decisions

about whether to treat a particular patient with HA

and perhaps which type of HA preparation is most

likely to produce the best treatment response for that

individual patient.

The goal of this study is to evaluate which patient

population, or patient characteristics, would benefit

most from HA injection. Since at least 18% of out-

patient visits to military treatment facilities by active-

duty personnel are attributed to painful knee disorders,

our study focuses on these patients. The study uses a

prospective, double-blinded clinical trial. A multi-

stage, multi-group machine learning model (Lee et al.,

2016b; Lee, 2017; Lee & Egan, 2022; Lee et al., 2021,

2023a, 2023b) described in Section 2.3 is used to

uncover discriminatory patterns that can predict

suitability of treatment and outcomes. The resulting

predictive rule can be implemented as part of a

clinical practice guideline for evidence-based

intervention. The model enables physicians to

administer HA products more selectively and

effectively to the targeted population to maximize

cost effectiveness and the percentage of patients who

experience a successful HA injection.

KDIR 2024 - 16th International Conference on Knowledge Discovery and Information Retrieval

498

2 METHODS AND STUDY

DESIGN

2.1 Patient Cohort, Treatment, and

Outcome Measures

2.1.1 Patient Data

Three group of patients (active-duty military

personnel, military retirees, and their families)

through the Department of Orthopaedics at the Naval

Medical Center Portsmouth were included. The

cohort includes those between 18 and 65 who sought

treatment for symptomatic osteoarthritis of the knee.

All patients were evaluated by a board-certified

orthopaedic surgeon. Each patient has had

radiographic evidence of knee OA with a minimum

Kellgren-Lawrence score of 1, has experienced

symptoms for more than three months, has failed a

minimum of three months of non-operative treatment,

including, but not limited to, analgesic and anti-

inflammatory medication, cortisone injection,

physical therapy, bracing, and/or heel wedge. The

cohort excludes patients with precautions or

contraindications for viscosupplementation, those

who had a cortisone injection within the past three

months, those who had prior HA injections at any

point, those with a history of deep knee infection,

those currently experiencing peripheral neuropathy,

chondrocalcinosis, or knee ligament instability, and

those who were candidates for knee surgery.

Patients were randomly assigned to receive either

Hylan G-F 20 (Synvisc®) [Sanofi Biosurgery,

Cambridge, MA, USA], a high molecular weight

(MW = 6000 kDa) cross-linked HA product derived

from an avian source, or EUFLEXXA®

[bioengineered 1% sodium hyaluronate (IA-BioHA);

Ferring Pharmaceuticals, Inc., Parsippany, NJ], a

medium weight (MW = 2400 - 2600 kDa) HA product

derived from bacterial fermentation.

Treatment allocations were randomly assigned by

the study pharmacist using the

RANDBETWEEN(0,1) function in Microsoft Excel.

Physicians, physicians performing the injections,

patients, and research personnel were blinded to

treatment assignment. To maintain blinding, the

pharmacy removed the original manufacturer's label

prior to dispensing and relabelled with the protocol

title, subject identifier and expiration date. The two

HA products had the same volume and color, so there

was no ability to discern one from the other at the time

of injection.

During a baseline evaluation before the first

injection, the following data were collected:

 patient demographic data: age, sex, height,

weight, BMI (as calculated from height and

weight), and smoking history.

 the Western Ontario and McMaster

Universities Osteoarthritis Index (WOMAC;

Bellamy, 2002) as a measure of knee OA

symptoms and functioning.

 the RAND-36 (Hays et al., 1993) as a measure

of general health status.

 the MARX Knee Activity Rating Scale (Marx

et al., 2001) to assess activity level (running,

deceleration, cutting (changing directions

while running) and pivoting.

 patient-rated health conditions (a) using a

comorbidity questionnaire (Sangha et al.,

2003) and (b) quality of life as measured by the

EuroQOL EQ-5D (Brooks, 1996).

 a patient-completed Arthritis Self-Efficacy

Scale (Lorig et al., 1989), an eight-item

instrument that assesses patient’s perceived

ability to manage arthritis symptoms.

Specific patient treatment expectations (e.g.,

“Improve ability to go up and down stairs”) and the

importance of these expectations were evaluated with

the scale developed by Mancuso (Mancuso et al.,

2001). Patients were also asked to rate their global

expectation for their response to the HA injections on

a seven-point scale ranging from “No improvement. I

don’t have much hope that this treatment will help my

symptoms at all” to “Excellent improvement. I expect

complete or nearly complete relief from knee

symptoms.” Patients with bilateral OA were

instructed to rate only the knee they perceived to be

more severe in terms of pain and functional

impairment on all instruments and to rate the same

knee at baseline and follow-ups.

Prior to the first injection, a physician assessed

quadriceps atrophy, presence of antalgic gait, knee

effusion, pain on palpation of the knee, range of

motion and alignment, and use of medication.

Patients also received four baseline radiographs.

These included (a) a standing anteroposterior (AP) of

the knee weight-bearing view; (b) weight-bearing

flexed view 400 posterior-anterior (PA) Rosenberg

view; (c) a lateral x-ray at 300; and (d) a Merchant

view. Digitized radiographs were evaluated for

osteoarthritis severity and for alignment by a board-

certified musculoskeletal radiologist and an

orthopaedic surgeon blinded to assigned treatment or

other patient characteristics. OA severity was rated

using the Kellgren-Lawrence Grading System which

incorporates joint space narrowing, osteophyte

formation, sclerosis and bony deformation observed

Prediction of Response to Intra-Articular Injections of Hyaluronic Acid for Knee Osteoarthritis

499

on x-rays. Scores range from 0 (no radiographic

features of OA) to 4 (large osteophytes, marked joint

space narrowing, severe sclerosis, and definite bony

deformity). Alignment was determined by measuring

the following angles from x-rays: (a) condylar-hip

angle of the femoral condylar tangent with respect to

the mechanical axis of the femur expressed as degrees

of deviation from 90°, negative for varus and positive

for valgus; (b) plateau-ankle angle between the tibial

margin tangent and the mechanical axis of the tibia

expressed as degrees of deviation from 90°, negative

for varus and positive for valgus; (c) condylar-plateau

angle between the femoral and tibial joint surface

tangents; and (d) hip-knee-ankle angle between a line

drawn from the center of the femoral head to the

midpoint of the tibial eminential spine and another

line from this midpoint to the center of the talus

surface of the ankle joint. The medial angle between

the lines is the HKA angle (varus < 180°).

2.1.2 HA Treatments

Patients received injections every seven days for a

total of three injections. Physicians received specific

instructions to standardize injection technique. All

injections were performed using an anteromedial

approach with a 21-gauge 1½” needle. Physicians

aspirated the knee joint prior to injection of the HA

product to ensure needle placement. Patients were

asked to flex and extend their knee a few times

following injection to maximize dispersal into the

joint. Patients were provided with written post

injection and standardized physical therapy

instructions. Patients were allowed full weight

bearing and full range of motion (active and passive)

after injections but were advised to avoid strenuous

activity (such as jogging, tennis, etc.) or prolonged

weight bearing for the first 48 hours after injection.

Patients were also instructed to use ice 30 minutes on

and 30 minutes off for 48 hours and take up to 4 gram

of acetaminophen per day as need for knee pain, but

not to take any 24 hours prior to each visit.

Patients were not offered a second course of HA

treatment within the first six months following the

final injection. Following the standard clinical

practice, those who received a second series of

injections after the first six months were not

considered treatment failures. Patients who had

surgery on the target knee to relieve arthritis

symptoms within the first six months following the

last HA injection were considered treatment failures.

The protocol was approved by the Institutional

Review Board at the data collection site and was

registered with ClinicalTrials.gov (identifier:

NCT01557868). A physician at the site served as the

medical monitor and an independent data and safety

board monitored the study.

2.1.3 Primary and Secondary Outcomes

The primary outcome was treatment responder status

defined a priori by improvement in the Western

Ontario and McMaster Universities Osteoarthritis

Index (WOMAC) Pain Scale (Hochberg et al., 1997;

Riddle & Perera, 2020) between baseline and 3-

month assessments. The WOMAC Pain Scale is

comprised of 5 items and the response format used in

this study was the 5-point rating scale. Scores were

calculated to range from 0 (worst) to 100 (best). The

reliability, validity and responsiveness of the

WOMAC Pain Scale have been supported in

numerous studies (Bellamy, et al., 2011; Burgers, et

al. 2015) and the WOMAC is one of the most widely

used outcome instruments in arthritis research.

Patients whose pain scores decreased by 20% or more

compared with their baseline scores were classified as

treatment responders and those whose scores did not

meet this criterion were classified as non-responders.

2.2 Machine Learning Predictive

Analysis

We apply a multi-stage machine learning approach to

analyze how different types of patients may respond

differently to HA treatment. The system will uncover

discriminatory features in the HA data that will reveal

patient and treatment characteristics that predict

optimal response to intra-articular injections of

hyaluronic acid for knee osteoarthritis. The model

determines which patient variables lead to the best

outcomes of HA.

Detail of the multi-stage multi-group discriminant

analysis via mixed-integer program (DAMIP) model

and computational framework is reported in Lee et al.

(Lee, 2017; Lee & Egan, 2022; Lee, Wang, et al.,

2016; Lee et al., 2021, 2023a, 2023b). Briefly we

include the DAMIP formulation below.

Let 𝑢



represent the binary variable that

indicates whether observation i in group g is

classified to group h, ℎ∈





∪𝒢. Thus, 𝑢



denotes a correct classification for observation i in

group g. The multi-group model with a reserved

judgement region is formulated as:

max   𝑢



∈𝒪



 ∈ 𝒢

(𝐃𝐀𝐌𝐈𝐏)

subject to

𝐿



= 𝜋



𝑓



𝒙



−

∑

𝜆



𝑓



𝒙





∈𝒢,

, ∀ ℎ,𝑔 ∈ 𝒢,𝑗 ∈ 𝒪



(1)

𝑦



−𝐿



≤𝑀1−𝑢



, ∀ ℎ, 𝑔 ∈ 𝒢, 𝑗 ∈ 𝒪



(2)

KDIR 2024 - 16th International Conference on Knowledge Discovery and Information Retrieval

500

𝑦



≤𝑀1−𝑢



, ∀ 𝑔 ∈ 𝒢, 𝑗 ∈ 𝒪



(3)

𝑦



−𝐿



≥𝜀1−𝑢



, ∀ ℎ, 𝑔 ∈ 𝒢, 𝑗 ∈ 𝒪



(4)

𝑦



≥𝜀 𝑢

,

∀ ℎ,𝑔 ∈ 𝒢, 𝑗 ∈ 𝒪



(5)

∑

𝑢

∈







∪𝒢

= 1, ∀ 𝑔 ∈ 𝒢, 𝑗 ∈ 𝒪



(6)

∑

𝑢

∈𝒪



≤𝛼



𝑛



, ∀ ℎ,𝑔 ∈ 𝒢,𝑔 ≠ ℎ (7)

𝑢



∈



0,1



∀ ℎ ∈





∪𝒢,𝑔∈𝒢,𝑗∈𝒪



(8)

𝑦



≥ 0, ∀ ℎ,𝑔 ∈ 𝒢, 𝑗 ∈ 𝒪



(9)

𝜆



≥ 0 ∀ ℎ, 𝑔 ∈ 𝒢,𝑔 ≠ ℎ (10)

Here, 𝜋



is the prior probability of group 𝑔 and

𝑓



(𝒙) is the conditional probability density function

of group 𝑔, 𝑔∈ 𝒢 for the data point 𝒙∈ℝ



. 𝒪



denote the set of observations in group g, and 𝑛



denote the number of observations in group g ∈𝒢.

𝛼



∈ (0, 1) , h, 𝑔∈𝒢, ℎ≠𝑔 represents the

predetermined limit on the inter-group

misclassification rate where the observations of group

𝑔 are misclassified to group h. The group assignment

decisions of observations that are classified into a

reserved judgment region are denoted by group g = 0.

Constraints (1) define the loss functions;

constraints (2)-(6) guarantee an observation is

uniquely assigned to the group with the maximum

value of 𝐿



(

𝒙

)

among all group, and constraints (7)

set the misclassification limits. With the reserved

judgment region in place, the mathematical system

ensures that a solution that satisfies the pre-set

misclassification rate always exists.

Theorem 1. Given prior probabilities 𝜋



and

conditional group density functions 𝑓



(𝒙), allocation

according to modified posterior probabilities defined

by the solution to (DAMIP) is a universally strongly

consistent method for classification.

Theorem 2. The DAMIP optimization problem is

𝒩𝒫 − 𝐶𝑜𝑚𝑝𝑙𝑒𝑡𝑒 when the number of groups is

greater than 2. The theoretical result holds for

DAMIP variants: (a) maximize the minimum value of

correct classification rates among all groups; (b)

maximize the minimum difference between correct

classification and misclassification; and (c) maximize

correct classification while constraining the

percentage of reserved judgment for each group.

The multi-stage classification approach utilizes

the reserved judgment region in DAMIP to improve

the classification performance, especially among

highly inseparable data. At each stage, DAMIP

partitions the observations into an ‘easy–to-classify’

subset that is classified to specific groups, and a

‘difficult-to-classify’ subset that is classified to a

reserved judgment region. The group assignment of

the difficult-to-classify observations are delayed, thus

allowing the DAMIP classifier to maintain a low

misclassification error. The observations in the

reserved judgment region are moved to the next stage

where a new feature set is selected and a new DAMIP

classifier is developed. In this way, the multi-stage

framework constructs a chain of successive classifiers

using different subsets of features. The classifier at

the ith stage, denoted by 𝑓



, can be represented by a

discriminant function 𝑓

(

𝒙



,𝝀

𝒊

)

, which is determined

by the feature subset 𝒙



, and the decision variables 𝝀

𝒊

in DAMIP.

At each stage, two models are performed: a

single-stage model that solves a DAMIP model

without a reserved judgment region and a multi-stage

model that solves a DAMIP model with a reserved

judgment region. The computational framework

selects the better of the two results. The algorithm

naturally terminates when there are no observations

in the reserved judgment region. To avoid overfitting

using too few observations for training, two

additional stopping criteria are used to terminate the

process: (a) the number of observations is less than a

preset minimum value, n, and (b) the maximum

allowed depth, d, is reached. The parameters n and d

are predetermined according to the number of

observations and the number of input features in the

given data.

Computationally, DAMIP classifier has some

distinct characteristics: (a) it is applicable for

classification of any number of groups; (b) there is

always a feasible solution to the model; (c) the

reserved judgement region facilitates successive

stage of classification to be performed; (d) DAMIP is

able to establish classification rules with good

predictive accuracy even when the training set is

relatively small; (e) DAMIP classifier can handle

imbalanced data; and (f) DAMIP classifier is totally

universally consistent.

Figure 1 shows the machine learning framework

where features are first selected via an exact branch-

and-bound algorithm (BB) and a fast heuristic particle

swarm optimization (PSO) (Lee et al., 2023a). The

resulting classification rule is subsequently

established via the DAMIP classifier. To quantify the

accuracy, ten-fold cross validation evaluation is

performed. If the results satisfy some pre-set accuracy

level, the classification rule is reported. Blind

prediction using this rule is then performed. We

contrast the BB-PSO/DAMIP results with eight

commonly used classifiers: Bernoulli Naïve Bayes,

Decision Tree, Gradient Boosting, K-nearest

neighbors, Logistic Regression, Neural Network,

Random Forest, and Support Vector Machine (SVM).

Prediction of Response to Intra-Articular Injections of Hyaluronic Acid for Knee Osteoarthritis

501

Figure 1: Multi-stage machine learning framework for HA

predictive analytics.

In 10-fold cross validation, the training set is

partitioned into 10 roughly equal subsets. In each run,

9-fold are selected to train and establish the rule, and

the remaining 1-fold is then tested, counting how

many of them are classified into which group.

Through 10 folds procedure (where each fold is being

validated exactly once), we obtain an unbiased

estimate of the classification accuracy.

Blind prediction is performed on patients that are

independent of the training set to gauge the predictive

power of the established rule. These patients have

never been used in the feature selection and the

machine learning analysis. We run each patient in the

blind set through the rule, which returns a group

status of the patient. The status is then checked

against the clinical status to confirm the accuracy.

The classifier response and outcome prediction

rules will culminate in a clinical decision algorithm

for the use of viscosupplementation in the treatment

of knee OA. For example, a physician determines that

HA is indicated for a particular patient. The physician

would then enter specific variables (those

discriminatory features identified by the classifier)

into a clinical computer program and a response set

would be generated for the potential outcome after

using hyaluronic acid injections. The optimal HA

agent(s) would be ranked. The physician would then

take this information into account as part of the

clinical decision process to select the HA agent for

the individual patient.

3 RESULTS

3.1 Patient Characteristics

Of the 273 patients assessed for eligibility, 45 did not

meet study criteria, 13 eligible patients declined to

participate, and 12 eligible patients could not

complete study participation due to anticipated

deployment or relocation. The other 203 eligible

patients were randomized to treatment: 107 assigned

to the Synvisc group and 96 to the Euflexxa group.

After randomization, 6 patients were non-compliant

with the study protocol, 9 received an excluded

intervention, 6 were reassigned, 10 were lost to

follow-up and 6 missed the follow-up appointment.

Consequently, these patients were not included in the

analyses, leaving a total of 166 (87 in the Synvisc

group and 79 in the Euflexxa group).

Table 1 summarizes the baseline characteristics of

the study participants. The Synvisc and the Euflexxa

groups did not differ on demographic or

anthropometric variables. The groups also did not

differ on co-morbid conditions with the exception that

a greater proportion of patients in the Euflexxa group

reported depression (21% vs. 10%, p = 0.02). The

baseline

scores from the patient report measures did

Table 1: Baseline Characteristics of the Study Participants.

Characteristic

Synvisc

(N = 107)

Euflexxa

(N = 96)

Combined

Sample

(N = 203)

Age – year 46+10 43+10 45+10

Female sex – no. (%) 44 (41) 36 (38) 80 (39)

Body mass index 30+5 29+5 30+5

Race

Asian 1 (1) 5 (5) 6 (3)

Black/African-

American

36 (34) 21 (22) 57 (28)

Hispanic 5 (5) 6 (6) 11 (5)

White 63 (59) 63 (66) 126 (62)

Other 2 (2) 1 (1) 3 (2)

Married – no. (%) 86 (80) 79 (82) 165 (81)

Current smoker – no. (%) 16 (15) 11(12) 27 (13)

Kellgren-Lawrence Score

Grade I – no. (%) 28 (26) 37 (39) 65 (32)

Grade II – no. (%) 44 (41) 33 (34) 77 (38)

Grade III – no. (%) 29 (27) 18 (19) 47 (23)

Grade IV–no. (%) 6 (6) 8 (8) 14 (7)

WOMAC Pain Scale 59+17 61+19 60+18

SF-36

Physical functioning 51+23 54+24 53+23

Mental health 79+15 74+18 77+17

Marx Activity Scale 5+5 5+5 5+5

EuroQOL EQ-5D Health

Rating

71+16 70+20 70+19

Arthritis Self-Efficacy Scale 6+2 6+2 6+2

Treatment response

expectation

5+1 5+1 5+1

Bilateral HA injections –

no. (%)

54 (51) 45 (47) 99 (49)

KDIR 2024 - 16th International Conference on Knowledge Discovery and Information Retrieval

502

not significantly differ between the two treatment

groups either.

3.2 Primary End Points

Of the 166 patients who completed the 3-month

assessment, 84 (50.6%) were classified as treatment

responders. Within the Synvisc group, 57.5% were

responders compared to 43% of the Euflexxa group

(p = 0.04). This outcome, as well as those at the 2-

week and 6-month follow-ups, is shown in Table 2.

Table 3 displays the percentage of patients who were

classified as “recovered” based on both statistically

reliable improvement in WOMAC Pain Scale scores

and a follow-up score that fell within the range of age-

and sex-matched patients who reported having no

knee problems or any history of knee surgery (see

Mann, et al., 2012).

Table 2: Treatment Responders (20% Reduction in

WOMAC Pain) by Treatment Group.

Follow-Up Synvisc Euflexxa P Value

2 weeks 56.3% 56.3% 0.55

3 months 57.5% 43.0% 0.04

6 months 51.3% 41.5% 0.31

Table 3: Return to Normal on WOMAC Pain Scale by

Treatment Group.

Follow-Up Synvisc Euflexxa P Value

2 weeks 36.5% 25.4% 0.20

3 months 38.0% 22.6% 0.06

6 months 33.9% 30.0% 0.31

3.3 Response and Outcome Prediction

We analyze the HA data to uncover patient and

treatment factors that predict optimal response to

intra-articular injections of hyaluronic acid for knee

osteoarthritis. The treatment responder status six

months after final injection is measured by

‘WOMACP20,” Treatment Responder Status Using

20% Reduction in WOMAC Pain Scale. Recovery

status is assessed via the KOOS Scale. The machine

learning model determines which patient variables

lead to the best outcomes of HA. We also perform the

prediction for each of the two HA products to gauge

their similarities and differences in treatment

outcome characteristics.

Table 4 shows the number of patients in the

training set and the blind prediction set for predicting

reinjection status. In this analysis, for every attribute

in which there is missing data, an associated binary

attribute is created to capture whether data is missing

or not for this field. The number of attributes at three

time-points: (a) baseline screening before first

injection; (b) prior to second injection (prefix: T0);

and (c) six months after final injection (prefix: T5) are

27, 483, and 1215 respectively. Table 5 shows the

training set and blind prediction statistics used for

predicting treatment responder status and recovery

status.

Table 4: Training set and blind prediction set characteristics

for predicting reinjection status.

Training set Blind Prediction Set

Total

reinjection

Reinjection Total

reinjection

Reinjection

150 111 39 53 40 13

Table 5: Training set and blind prediction set characteristics

for predicting treatment responder status and recovery

status.

Training set Blind Prediction Set

Total

Non-

Responder

Responder Total

Non-

Responder

Responde

71 34 37 70 41 29

Synvisc

40 18 22 36 19 17

Euflexxa

35 21 14 30 17 13

We summarize herein the best predictive rules for

each of the analyses. Table 6 shows the prediction

accuracy for no-reinjection versus re-injection using

attributes collected up to the three stated time-points.

For the baseline results, factors that appear to be

critical includes “Weight,” “Currently Smoke

Cigarettes,” and “Smoking: Number per day.”

Baseline prediction results are comparable to Pap

Smear test accuracy (~70%).

We can observe high accuracy in predicting

success for patients using screening and T0 attributes

alone (86% blind predictive accuracy). This is very

promising for identifying patients early (just after the

first injection) who should be targeted for HA

intervention (with an expected success outcome). The

discriminatory features selected includes the Marx

Activity Scale “T0MarxCuttingSymptomFree”,

“T0MarxCutting”, effectiveness of exercise

“T0ExerciseEffective”, confidence in the injector

“T0ConfidenceInjector”, and other medications

“T0MedicationXEffective.”

Including attributes until T5 significantly

increases the accuracy for predicting the reinjection

group (from 71% to 89%). Early attributes include

“T0PhysicalTherapyEffective”,

“T0MedicationXEffective,” and overall health

“T0EQRateHealth” continue to appear among the

selected features.

Prediction of Response to Intra-Articular Injections of Hyaluronic Acid for Knee Osteoarthritis

503

Table 6: Best predictive rule for re-injection status when

using attributes (a) baseline screening before first injection,

(b) prior to second injection, and (c) 6 months after final

injection.

Input

attributes

10-fold cross

validation

blind prediction

No-

reinjecti

Re-

injection

No-

reinjection

Re-

injection

Baseline

screening

71% 71% 72% 71%

Prior to 2

injection

89% 74% 86% 71%

Input

attributes

84% 83% 81% 89%

Figure 2 show the 10-fold cross validation and

blind prediction accuracies for predicting treatment

responder status and recovery status for patients

injected with Synvisc and Euflexxa, respectively. For

each HA injection, four measurement frameworks are

graphed: PSO/DAMIP results for predicting

treatment responder and recovery respectively versus

the best results from the eight commonly used

classifiers, Random Forest. Our PSO/DAMIP

framework selected 3-8 discriminatory features

whereas Random Forest uses over 40 features with

poor results. Although the size of the two groups is

rather balanced, the challenge here is due to the

highly inseparable data that makes it difficult to

classify using traditional approaches. A multi-stage

approach allows the partitioning of patients from the

same group via different rules (associated with

different features).

Figure 2: Comparison of the best DAMIP classification

rules for predicting treatment responder status and recovery

status using Synvisc(top) and Euflexxa (bottom) against the

Random Forest approach.

Our study shows that early predictors can be used

to determine the group of patients who benefit the

most from HA injection. It also allows evidence-

based correction to be made during the course of

treatment. For example, after T0, the physician can

quit treatment based on results from the predictive

rule.

4 CONCLUSIONS

In 2019, about 528 million people worldwide were

living with osteoarthritis, an increase of 113% since

1990. For 365 million, the knee was the most

frequently affected joint. On average, the total cost of

knee replacement surgery ranges from $30,000 to

$50,000. This includes the cost of the surgery itself,

the hospital stays, anesthesia and other associated

medical expenses. HA treatment, on the other hand,

costs about $900 to $3,000 for a full course (three to

five injections administered over several weeks). The

range reflects the variations due to the type of HA

product and the physician's fees. Although 50% of

knee osteoarthritis patients eventually receive

surgical procedures, almost one third of these

surgeries are unnecessary. Hence intra-articular

injections of hyaluronic acid can serve as a non-

invasive cost-effective alternative to surgery for knee

osteoarthritis.

Unlike surgical options, HA injections do not

require incisions or extensive recovery periods. HA is

a substance that naturally occurs in the synovial fluid

of the joints, which helps lubricate and cushion them.

In osteoarthritis, this fluid becomes less effective,

leading to pain and reduced mobility. Thus, HA

20%

40%

60%

80%

100%

DAMIP /

WOMACP20

DAMIP /

KOOSRecovery

RandomForest /

WOMACP20

RandomForest /

KOOSRecovery

Percentage of accuracy

Name of classifier

10-fold cross validation (solid) and Blind

prediction (dotted) results for Synvisc

10-fold, Non-Responder

10-fold, Responder

Blind prediction, Non-Responder

Blind prediction, Responder

20%

40%

60%

80%

100%

DAMIP /

WOMACP20

DAMIP /

KOOSRecovery

RandomForest /

WOMACP20

RandomForest /

KOOSRecovery

Percentage of accuracy

Name of classifier

10-fold cross validation (solid) and Blind

prediction (dotted) results for Euflexxa

10-fold, Non-Responder

10-fold, Responder

Blind prediction, Non-Responder

Blind prediction, Responder

KDIR 2024 - 16th International Conference on Knowledge Discovery and Information Retrieval

504

injected directly into the knee joint helps restore the

lubricating properties of the synovial fluid and reduce

inflammation. By restoring lubrication, HA injections

can help improve joint mobility and reduce stiffness.

The procedure is relatively low risk, with mild

potential side effects, such as temporary swelling or

discomfort.

However, the benefits of HA injections are not

permanent; they typically last for several months.

Repeated injections may be needed for ongoing relief.

More importantly, controversies exist regarding its

safety and efficacy, the number of injections and

courses, type of preparation, duration of its effects,

and combining it with other drugs or molecules. Other

factors include patient characteristics such as age,

weight, gender, and severity of the OA. The study

uses a prospective, double-blinded clinical trial. A

multi-stage, multi-group DAMIP-based machine

learning model is utilized to uncover discriminatory

features that can predict the response status of knee

OA patients to different types of HA treatment. The

algorithm can identify certain subgroups of knee OA

patients who respond well (or those who don’t) to HA

therapy. The study’s baseline result, including factors

such as patients’ weight, smoking status and smoking

frequency, gives physicians insight for patient

treatment recommendations by identifying those

most suitable for HA injection.

To the best of our knowledge, this work presents

the first machine learning approach that predicts

patient responses to HA injections for knee

osteoarthritis. Another uniqueness of this study is that

this is the first prospective clinical trial designed such

that in addition to clinical data, patient self-reporting

data is also carefully collected. The latter is

challenging since patients often refuse or bypass

questionnaires or miss filling in forms. Self-reported

answers may be exaggerated; respondents may be too

embarrassed to reveal private details; various biases

may affect the results, like social desirability bias.

However, knee pains, whether patients can move or

do certain activities are standard questions used by

physicians and are rather routine evaluation for

active-duty personnel and athletes, and hence their

self-reporting are rather reliable. Further, there has

been no study indicating that patients would

exaggerate their pain to receive treatment to their

knee pain.

Traditional data collection methods, primarily

focusing on clinical settings, limits our understanding

of drug efficacy and patient wellbeing. Patient self-

reporting data is crucial for machine learning in

healthcare because it provides a unique, subjective

perspective on a patient's health experience, including

their symptoms, quality of life, and perception of

treatment effectiveness, which can be vital for

accurate diagnosis, treatment planning, and overall

patient care, often not captured by solely objective

medical data like lab results or imaging scans. There

is growing interest and support for the utility and

importance of patient-reported outcome measures

(PROMs) (Kingsley & Patel, 2017; Verma, et al.,

2021). This is one of the strengths of our study since

it includes a broad spectrum of patient wellbeing data.

DAMIP classifier was chosen partly due to earlier

DAMIP models have produced good predictive

accuracy on blind data for numerous clinical studies

where the training patient size is relatively small (e.g.,

in early cancer detection to uncover genomic

signatures that predict CpG islands methylation

(Feltus, et al., 2003), vaccine immunogenicity

prediction that accelerates vaccine design and target

delivery (Lee, Nakaya, et al., 2016a; Nakaya, et al,

2011, 2015; Querec, et al., 2009;) in which DAMIP

results were instrumental in the eventual world-wide

clinical trial of the Malaria vaccines (Kazmin, et al.,

2017; Lee, Nakaya et al., 2016a)). DAMIP has also

been used for studies involving very large number

patient sets with equally consistent predictive

accuracy (Lee, Wang, et al., 2016b). Multi-stage is

performed herein to manage the highly inseparable

data.

With the established predictive rule, prior to

treatment physicians can predict a patient’s response

to HA injections based on patient and treatment

characteristics. Physicians can then make empirically

informed decisions about whether to treat a particular

patient with HA and perhaps which type of HA

preparation is most likely to produce the best

treatment response for that individual patient.

Predicting treatment response based on clinically

measured variables and patient-centered well-being

data will empower physicians with an evidence-based

decision-making tool to administer the most cost-

effective intervention for the patients.

The study's follow-up period is focused on six

months after the final injection. Since knee

osteoarthritis is incurable, treatment for patients

includes rehabilitation, activity modification, weight

loss when indicated, shoe orthoses, local modalities,

and medication. For more severe cases, either HA

injections or knee surgery is selected. And HA

injections are typically given as a series of 3-5

injections, spaced one week apart, with repeat

courses usually needed every six months, depending

on the individual's pain relief duration and the

severity of their arthritis; most people experience pain

relief for several months after a full course of

injections.

Prediction of Response to Intra-Articular Injections of Hyaluronic Acid for Knee Osteoarthritis

505

The data and model derived from this study

allows physicians to administer HA products more

selectively and effectively, which will increase the

percentage of patients who experience a successful

HA therapy. Information about predicted responses

could easily be shared with patients to incorporate

their values and preferences into treatment selection.

Specifically, the classification rule can be

implemented within the electronic health record

system as an Application Programming Interface

(API). In addition, this decision support tool would

allow physicians to quickly determine whether a

patient is exhibiting at least an expected treatment

response and if not, to potentially take corrective

action. Of note, this model can also be used to predict

patient responses to other forms of treatment and

conditions.

There is a clear demand for evidence-based

medical decision-making in addition to expert

opinion, clinical experience and case reports.

Additionally, there is an increased demand for

clinical studies of prospective, rather than

retrospective, treatment assessment options. While

each of these study types has a role, the value of

evidence-based, single studies or meta-analyses of

published reports is that clinical criterion or criteria

are analyzed globally with respect to outcome.

Quantified variables that are uncovered by predictive

models are evaluated and analyzed and can serve as

important decision variables to help physicians select

the best course of treatment for patients. Evidence-

based decision-making increases outcome success.

Trends, impressions and opinions are minimized and

objective, evidence-based, outcome-driven targeted

delivery is maximized.

ACKNOWLEDGEMENTS

A portion of the results from this project (the machine

learning advances) received the first runner-up prize

at the Caterpillar and INFORMS Innovative

Applications in Analytics Award. This work is

partially supported by grants from the National

Science Foundation (IIP-1361532), and the American

Orthopedic Society for Sports Medicine. Findings

and conclusions in this paper are those of the authors

and do not necessarily reflect the views of the

National Science Foundation and or the American

Orthopedic Society for Sports Medicine. The authors

extend their deepest respect and gratitude to the late

Dr. Barton J. Mann PhD, with whom we collaborated

on the knee osteoarthritis research, and to Dr. Captain

Marlene DeMaio (retired) from Medical Corps,

United States Navy, for her clinical guidance and

collaboration on the project. We thank the

anonymous reviewers for their insightful comments.

REFERENCES

Ayis S, Dieppe P. (2009). The natural history of disability

and its determinants in adults with lower limb

musculoskeletal pain. J Rheumatol. 36:583–91. doi:

10.3899/jrheum.080455

Barbour, K. E., Helmick, C. G., Boring, M., & Brady, T. J.

(2017). Vital Signs: Prevalence of Doctor-Diagnosed

Arthritis and Arthritis-Attributable Activity Limitation

— United States, 2013–2015. MMWR. Morbidity and

Mortality Weekly Report, 66(9).

Bellamy, N. (2002). WOMAC osteoarthritis index user

guide. Version V. Brisbane, Australia.

Bellamy, N., Wilson, C., Hendrikz, J., Whitehouse, S. L.,

Patel, B., Dennison, S., & Davis, T. (2011).

Osteoarthritis Index delivered by mobile phone (m-

WOMAC) is valid, reliable, and responsive. Journal of

Clinical Epidemiology, 64(2).

Brooks, R., & De Charro, F. (1996). EuroQol: The current

state of play. Health Policy, 37(1).

https://doi.org/10.1016/0168-8510(96)00822-6

Burgers, P. T. P. W., Poolman, R. W., Van Bakel, T. M. J.,

Tuinebreijer, W. E., Zielinski, S. M., Bhandari, M.,

Patka, P., & Van Lieshout, E. M. M. (2015). Reliability,

validity, and responsiveness of the Western Ontario and

McMaster Universities osteoarthritis index for elderly

patients with a femoral neck fracture. Journal of Bone

and Joint Surgery - American Volume, 97(9).

Chan, L. L. Y., Wong, A. Y. L., & Wang, M. H. (2020).

Associations between sport participation and knee

symptoms: A cross-sectional study involving 3053

undergraduate students. BMC Sports Science,

Medicine and Rehabilitation, 12(1).

Chavda, S., Rabbani, S. A., & Wadhwa, T. (2022). Role and

Effectiveness of Intra-articular Injection of Hyaluronic

Acid in the Treatment of Knee Osteoarthritis: A

Systematic Review. Cureus.

https://doi.org/10.7759/cureus.24503

Cui, A., Li, H., Wang, D., Zhong, J., Chen, Y., & Lu, H.

(2020). Global, regional prevalence, incidence and risk

factors of knee osteoarthritis in population-based

studies. EClinicalMedicine, 29–30.

Driban, J. B., Hootman, J. M., Sitler, M. R., Harris, K. P.,

& Cattano, N. M. (2017). Is participation in certain

sports associated with knee osteoarthritis? A systematic

review. In Journal of Athletic Training (Vol. 52, Issue

6). https://doi.org/10.4085/1062-6050-50.2.08

Elgaddal N, Kramarow EA, Weeks JD, Reuben C. Arthritis

in adults age 18 and older: United States, 2022. NCHS

Data Brief, no 497. Hyattsville, MD: National Center

for Health Statistics. 2024.

Fallon, E. A., Boring, M. A., Foster, A. L., Stowe, E. W.,

Lites, T. D., Odom, E. L., & Seth, P. (2023). Prevalence

of Diagnosed Arthritis — United States, 2019–2021.

KDIR 2024 - 16th International Conference on Knowledge Discovery and Information Retrieval

506

MMWR. Morbidity and Mortality Weekly Report,

72(41). https://doi.org/10.15585/mmwr.mm7241a1

Feltus, F. A., Lee, E. K., Costello, J. F., Plass, C., &

Vertino, P. M. (2003). Predicting aberrant CpG island

methylation. Proceedings of the National Academy of

Sciences of the United States of America, 100(21).

GBD 2019: Global burden of 369 diseases and injuries in

204 countries and territories, 1990–2019: a systematic

analysis for the Global Burden of Disease Study 2019.

Hays, R. D., Sherbourne, C. D., & Mazel, R. M. (1993). The

rand 36‐item health survey 1.0. Health Economics,

2(3). https://doi.org/10.1002/hec.4730020305

Hochberg, M. C., Altman, R. D., Brandt, K. D., &

Moskowitz, R. W. (1997). Design and conduct of

clinical trials in osteoarthritis: Preliminary

recommendations from a Task Force of the

Osteoarthritis Research Society. Journal of

Rheumatology, 24(4).

Kazmin, D., Nakaya, H. I., Lee, E. K., Johnson, M. J., Van

Der Most, R., Van Den Berg, R. A., Ballou, W. R.,

Jongert, E., Wille-Reece, U., Ockenhouse, C., Aderem,

A., Zak, D. E., Sadoff, J., Hendriks, J., Wrammert, J.,

Ahmed, R., & Pulendran, B. (2017). Systems analysis

of protective immune responses to RTS,S malaria

vaccination in humans. Proceedings of the National

Academy of Sciences of the United States of America,

114(9). https://doi.org/10.1073/pnas.1621489114

Kingsley, C., & Patel, S. (2017). Patient-reported outcome

measures and patient-reported experience measures.

BJA Education, 17(4).

Lee, E. K. (2017). Innovation in big data analytics:

Applications of mathematical programming in

medicine and healthcare. Proceedings - 2017 IEEE

International Conference on Big Data, Big Data 2017,

2018-January.

Lee, E. K., & Egan, B. (2022). A Multi-stage Multi-group

Classification Model: Applications to Knowledge

Discovery for Evidence-based Patient-centered Care.

International Joint Conference on Knowledge

Discovery, Knowledge Engineering and Knowledge

Management, IC3K - Proceedings, 1.

Lee, E. K., Li, Z., Wang, Y., Hagen, M. S., Davis, R., &

Egan, B. M. (2021). Multi-Site Best Practice

Discovery: From Free Text to Standardized Concepts to

Clinical Decisions. 2021 IEEE International

Conference on Bioinformatics and Biomedicine

(BIBM), 2766–2773.

https://doi.org/10.1109/BIBM52615.2021.9669414

Lee, E. K., Nakaya, H. I., Yuan, F., Querec, T. D., Burel,

G., Pietz, F. H., Benecke, B. A., & Pulendran, B.

(2016a). Machine learning for predicting vaccine

immunogenicity. Interfaces, 46(5).

Lee, E. K., Wang, Y., Hagen, M. S., Wei, X., Davis, R. A.,

& Egan, B. M. (2016b). Machine learning: Multi-site

evidence-based best practice discovery. Lecture Notes

in Computer Science (Including Subseries Lecture

Notes in Artificial Intelligence and Lecture Notes in

Bioinformatics), 10122 LNCS.

Lee, E. K., Wang, Y., He, Y., & Egan, B. M. (2019). An

efficient, robust, and customizable information

extraction and pre-processing pipeline for electronic

health records. IC3K 2019 - Proceedings of the 11th

International Joint Conference on Knowledge

Discovery, Knowledge Engineering and Knowledge

Management, 1.

https://doi.org/10.5220/0008071303100321

Lee, E. K., Yuan, F., Man, B. J., & Egan, B. (2023a). A

General-Purpose Multi-stage Multi-group Machine

Learning Framework for Knowledge Discovery and

Decision Support. Communications in Computer and

Information Science, 1842 CCIS.

https://doi.org/10.1007/978-3-031-43471-6_4

Lee, E. K., Yuan, F., Mann, B. J., & Egan, B. (2023b).

Handling Imbalanced and Poorly Separated Data: a

Multi-Stage Multi-Group Machine Learning Approach.

Proceedings - 2023 2023 IEEE International

Conference on Bioinformatics and Biomedicine, BIBM

2023.

https://doi.org/10.1109/BIBM58861.2023.10386028

Leifer, V. P., Katz, J. N., & Losina, E. (2022). The burden

of OA-health services and economics. Osteoarthritis

and Cartilage, 30(1).

https://doi.org/10.1016/j.joca.2021.05.007

Lo, J., Chan, L., & Flynn, S. (2021). A Systematic Review

of the Incidence, Prevalence, Costs, and Activity and

Work Limitations of Amputation, Osteoarthritis,

Rheumatoid Arthritis, Back Pain, Multiple Sclerosis,

Spinal Cord Injury, Stroke, and Traumatic Brain Injury

in the United States: A 2019 Update. In Archives of

Physical Medicine and Rehabilitation (Vol. 102, Issue

1). https://doi.org/10.1016/j.apmr.2020.04.001

Lohmander, L. S., Englund, P. M., Dahl, L. L., & Roos, E.

M. (2007). The long-term consequence of anterior

cruciate ligament and meniscus injuries: Osteoarthritis.

In American Journal of Sports Medicine (Vol. 35, Issue

10). https://doi.org/10.1177/0363546507307396

Long, H., Liu, Q., Yin, H., Wang, K., Diao, N., Zhang, Y.,

Lin, J., & Guo, A. (2022). Prevalence Trends of Site-

Specific Osteoarthritis From 1990 to 2019: Findings

From the Global Burden of Disease Study 2019.

Arthritis and Rheumatology, 74(7).

https://doi.org/10.1002/art.42089

Lorig, K., Chastain, R. L., Ung, E., Shoor, S., & Holman,

H. R. (1989). Development and evaluation of a scale to

measure perceived self‐efficacy in people with arthritis.

Arthritis & Rheumatism, 32(1).

https://doi.org/10.1002/anr.1780320107

Mancuso, C. A., Sculco, T. P., Wickiewicz, T. L., Jones, E.

C., Robbins, L., Warren, R. F., & Williams-Russo, P.

(2001). Patients’ expectations of knee surgery. Journal

of Bone and Joint Surgery, 83(7).

https://doi.org/10.2106/00004623-200107000-00005

Mann, B. J., Gosens, T., & Lyman, S. (2012). Quantifying

clinically significant change: A brief review of methods

and presentation of a hybrid approach. In American

Journal of Sports Medicine (Vol. 40, Issue 10).

https://doi.org/10.1177/0363546512457346

Marx, R. G., Stump, T. J., Jones, E. C., Wickiewicz, T. L.,

& Warren, R. F. (2001). Development and evaluation

of an activity rating scale for disorders of the knee.

Prediction of Response to Intra-Articular Injections of Hyaluronic Acid for Knee Osteoarthritis

507

American Journal of Sports Medicine, 29(2).

https://doi.org/10.1177/03635465010290021601

McAlindon, T. E., Wilson, P. W. F., Aliabadi, P.,

Weissman, B., & Felson, D. T. (1999). Level of

physical activity and the risk of radiographic and

symptomatic knee osteoarthritis in the elderly: The

Framingham study. American Journal of Medicine,

106(2). https://doi.org/10.1016/S0002-9343(98)00413-

Mora, J. C., Przkora, R., & Cruz-Almeida, Y. (2018). Knee

osteoarthritis: Pathophysiology and current treatment

modalities. In Journal of Pain Research (Vol. 11).

https://doi.org/10.2147/JPR.S154002

Nakaya, H. I., Hagan, T., Duraisingham, S. S., Lee, E. K.,

Kwissa, M., Rouphael, N., Frasca, D., Gersten, M.,

Mehta, A. K., Gaujoux, R., Li, G. M., Gupta, S.,

Ahmed, R., Mulligan, M. J., Shen-Orr, S., Blomberg,

B. B., Subramaniam, S., & Pulendran, B. (2015).

Systems Analysis of Immunity to Influenza

Vaccination across Multiple Years and in Diverse

Populations Reveals Shared Molecular Signatures.

Immunity, 43(6).

https://doi.org/10.1016/j.immuni.2015.11.012

Nakaya, H. I., Wrammert, J., Lee, E. K., Racioppi, L.,

Marie-Kunze, S., Haining, W. N., Means, A. R.,

Kasturi, S. P., Khan, N., Li, G. M., McCausland, M.,

Kanchan, V., Kokko, K. E., Li, S., Elbein, R., Mehta,

A. K., Aderem, A., Subbarao, K., Ahmed, R., &

Pulendran, B. (2011). Systems biology of vaccination

for seasonal influenza in humans. Nature Immunology,

12(8). https://doi.org/10.1038/ni.2067

Querec, T. D., Akondy, R. S., Lee, E. K., Cao, W., Nakaya,

H. I., Teuwen, D., Pirani, A., Gernert, K., Deng, J.,

Marzolf, B., Kennedy, K., Wu, H., Bennouna, S.,

Oluoch, H., Miller, J., Vencio, R. Z., Mulligan, M.,

Aderem, A., Ahmed, R., & Pulendran, B. (2009).

Systems biology approach predicts immunogenicity of

the yellow fever vaccine in humans. Nature

Immunology, 10(1). https://doi.org/10.1038/ni.1688

Riddle, D. L., Jiranek, W. A., & Hayes, C. W. (2014). Use

of a validated algorithm to judge the appropriateness of

total knee arthroplasty in the United States: A

multicenter longitudinal cohort study. Arthritis and

Rheumatology, 66(8).

https://doi.org/10.1002/art.38685

Riddle, D. L., & Perera, R. A. (2020). The WOMAC pain

scale and crosstalk from co-occurring pain sites in

people with knee pain: A causal modeling study.

Physical Therapy, 100(10).

https://doi.org/10.1093/ptj/pzaa098

Salis, Z., & Sainsbury, A. (2024). Association of long-term

use of non-steroidal anti-inflammatory drugs with knee

osteoarthritis: a prospective multi-cohort study over 4-

to-5 years. Scientific Reports, 14(1).

https://doi.org/10.1038/s41598-024-56665-3

Sangha, O., Stucki, G., Liang, M. H., Fossel, A. H., & Katz,

J. N. (2003). The Self-Administered Comorbidity

Questionnaire: A new method to assess comorbidity for

clinical and health services research. Arthritis Care and

Research, 49(2). https://doi.org/10.1002/art.10993

Saxon, L., Finch, C., & Bass, S. (1999). Sports

participation, sports injuries and osteoarthritis

implications for prevention. In Sports Medicine (Vol.

28, Issue 2). https://doi.org/10.2165/00007256-

199928020-00005

Sharma, L., Song, J., Felson, D. T., Cahue, S., Shamiyeh,

E., & Dunlop, D. D. (2001). The role of knee alignment

in disease progression and functional decline in knee

osteoarthritis. JAMA, 286(2).

https://doi.org/10.1001/jama.286.2.188

Spector, T. D., Harris, P. A., Hart, D. J., Cicuttini, F. M.,

Nandra, D., Etherington, J., Wolman, R. L., & Doyle,

D. V. (1996). Risk of osteoarthritis associated with

long-term weight-bearing sports: A radiologic survey

of the hips and knees in female ex-athletes and

population controls. Arthritis and Rheumatism, 39(6).

https://doi.org/10.1002/art.1780390616

Theis, K. A., Steinweg, A., Helmick, C. G., Courtney-Long,

E., Bolen, J. A., & Lee, R. (2019). Which one? What

kind? How many? Types, causes, and prevalence of

disability among U.S. adults. Disability and Health

Journal, 12(3).

https://doi.org/10.1016/j.dhjo.2019.03.001

Turner, A. P., Barlow, J. H., & Heathcote-Elliot, C. (2000).

Long term health impact of playing professional

football in the United Kingdom. British Journal of

Sports Medicine, 34(5).

https://doi.org/10.1136/bjsm.34.5.332

United States Bone and Joint Initiative: The Burden of

Musculoskeletal Diseases in the United States (BMUS)

Fourth Edition. Forthcoming;

4th:http://www.boneandjointburden.org. Accessed

December 20, 2023.

United States Bone and Joint Initiative. The Burden of

Musculoskeletal Diseases in the United States (BMUS).

In: In. Fourth ed. Rosemont, IL. 2018: Available at

https://www.boneandjointburden.org/fourth-edition.

Accessed June 12, 2023

Verma, D., Bach, K., & Mork, P. J. (2021). Application of

machine learning methods on patient reported outcome

measurements for predicting outcomes: A literature

review. In Informatics (Vol. 8, Issue 3).

https://doi.org/10.3390/informatics8030056

World Health Organization. Osteoarthritis. 14 July 2023.

https://www.who.int/news-room/fact-

sheets/detail/osteoarthritis

Zhang, W., Nuki, G., Moskowitz, R. W., Abramson, S.,

Altman, R. D., Arden, N. K., Bierma-Zeinstra, S.,

Brandt, K. D., Croft, P., Doherty, M., Dougados, M.,

Hochberg, M., Hunter, D. J., Kwoh, K., Lohmander, L.

S., & Tugwell, P. (2010). OARSI recommendations for

the management of hip and knee osteoarthritis. Part III:

Changes in evidence following systematic cumulative

update of research published through January 2009.

Osteoarthritis and Cartilage, 18(4).

KDIR 2024 - 16th International Conference on Knowledge Discovery and Information Retrieval

508