A Gait Analysis Tool Based on Machine Learning to Support the

Rehabilitation Strategy of Post-stroke Patients

Nicoletta Balletti

1,2 a

, Gennaro Laudato

1 b

and Rocco Oliveto

1,3 c

STAKE Lab, University of Molise, Pesche (IS), Italy

Defense Veterans Center, Ministry of Defense, Rome, Italy

Datasound srl, Pesche (IS), Italy

Keywords:

Gait Analysis, Motion Tracking, Post Stroke Rehabilitation, Machine Learning.

Abstract:

Stroke is a serious medical condition that can result in permanent brain damage and other pathological issues.

Conditions suffered by survivors ranged in severity from full recovery to signiﬁcant movement disability. Even

though some may recover quickly, many stroke survivors require long-term support to help them achieve as

much independence as they can. Thanks to a proper rehabilitation, patients who have experienced a stroke can

work to regain skills that are suddenly lost when a section of their brain is injured. Due to the breakdown of

neuronal networks in the motor cortex, abnormal gait patterns are a typical disability after a stroke. Therefore,

gait analysis can be a powerful tool to support stroke patients during rehabilitation. In this work we propose

GIULYO, a Machine Learning based tool that offers support in the assessment of video gait trials in stroke

patients by providing an automatic analysis on the muscle activity of the assisted subject. GIULYO is a

device-agnostic tool because it accepts motion tracking data in terms of 3d trajectories regardless of the type

of instrumentation. GIULYO has been validated on the ARRA Stroke dataset and the results showed an overall

accuracy of 0.74 while on a subset a patients—with common clinical assessment of mobility impairments—

the accuracy increased to 0.92, therefore demonstrating the feasibility of involving a ML-based approach for

the rehabilitation support of post stroke patients.

1 INTRODUCTION

Stroke is a clinically deﬁned condition of quickly

evolving symptoms or indicators of localized loss of

brain function (Warlow et al., 1997). Survivors ex-

perienced conditions that can range in severity from

complete recovery to severe movement impairment

(Warlow, 1998). As a result, stroke patients continue

to have serious locomotor deﬁcits. Additionally, pa-

tients and rehabilitation professionals continue to face

daily challenges in achieving a good gait recovery fol-

lowing a stroke (Nadeau et al., 2013).

The idea behind this work is based on the con-

sideration that biomechanical components of steady-

state walking in healthy individuals have been demon-

strated to be produced by separated, coexcited mus-

cle groups (Allen and Neptune, 2012; Neptune et al.,

2009). These speciﬁc muscle groups can be detected

https://orcid.org/0000-0002-6617-7074

https://orcid.org/0000-0002-5241-1608

https://orcid.org/0000-0002-7995-8582

thanks to the Surface Electromyography (S-EMG)

(Routson et al., 2014), by evaluating the number of

channels that registered EMG activity. This value is

referred as modules (Allen and Neptune, 2012; Nep-

tune et al., 2009).

Nevertheless, those who have had a stroke have

poor intermuscular coordination, which is deﬁned by

the merging of modules that are ordinarily indepen-

dent in healthy people (Clark et al., 2010). Follow-

ing a stroke, having more independent modules has

been linked to better performance in a variety of clin-

ical and biomechanical walking assessments, includ-

ing faster walking, a better Dynamic Gait Index (DGI)

(Jonsdottir and Cattaneo, 2007), and better step length

and propulsion symmetry (Bowden et al., 2010; Clark

et al., 2010).

Many efforts were provided by the scientiﬁc com-

munity to support the rehabilitation strategy of post-

stroke patients by proposing video gait analysis tools

(Liu et al., 2021; Mirelman et al., 2010; Swank et al.,

2020). Liu et al. (2021) developed a method based on

the Microsoft Kinect camera that could detect the ro-

400

Balletti, N., Laudato, G. and Oliveto, R.

A Gait Analysis Tool Based on Machine Learning to Support the Rehabilitation Strategy of Post-stroke Patients.

DOI: 10.5220/0011697800003414

In Proceedings of the 16th International Joint Conference on Biomedical Engineering Systems and Technologies (BIOSTEC 2023) - Volume 5: HEALTHINF, pages 400-407

ISBN: 978-989-758-631-6; ISSN: 2184-4305

 2023 by SCITEPRESS – Science and Technology Publications, Lda. Under CC license (CC BY-NC-ND 4.0)

tation and movement of the Center of Mass in many

planes. The CoM is regarded as a crucial indicator for

evaluating the impact of therapy and recovery. Swank

et al. (2020) used video gait analysis to demonstrate

the validity of Robotic Exoskeletons (EKSO) session

in rehabilitation therapies of post-stroke individuals.

However, we believe that there is still room to

contribute this research ﬁeld. Indeed, in this paper,

we present GIULYO (video-based GaIt tool for the

aUtomatic anaLysis of rehabilitation qualitY in post

strOke survivors), a tool designed to provide an auto-

matic quantitative assessment to the gait trials of a re-

habilitation session. Indeed, GIULYO is an approach

capable of analyzing a video gait trial and automati-

cally classifying it according to the number of mod-

ules detected. Furthermore, GIULYO is a device-

agnostic tool, i.e., it does not depend on any speciﬁc

motion tracking instrumentation because its workﬂow

begins with the processing of 3d trajectories. Also,

GIULYO allows the clinical evaluation of a gait trial

without the need of installing S-EMG electrodes on

the interested limb. The proposed approach has been

validated on the ARRA post-stroke database (Rout-

son et al., 2014). GIULYO was submitted to an ex-

tensive ML experimentation and validated using the

most ﬁtting validation scheme when involving data

from different subjects, i.e., the Leave 1 Subject Out

(L1SO) cross validation. Results showed that GIU-

LYO is capable of predicting the number of moduls of

a gait trial with an overall accuracy of 0.74. However,

on a speciﬁc subset—more than a third—of subjects

presenting the common characteristic of poor motor

skills, GIULYO achieves an overall accuracy of 0.92.

The rest of the paper is structured as follows: Sec-

tion 2 provides details on related works focused on

gait analysis for post-stroke rehabilitation. Section 3

presents GIULYO, our novel approach for the assess-

ment of a gait trial in terms of muscle activity. Section

4 reports the design and the results of the empirical

study we conducted to evaluate GIULYO. Section 5

reports the results achieved by GIULYO and a qual-

itative analysis in terms of features importance and

performances at subject level. Finally, Section 7 con-

cludes the paper and provides suggestions for possible

future research directions.

2 RELATED WORKS

In this section a review of the state-of-the-art related

to the main contribution in the research ﬁeld of gait

analysis in post-stroke individuals is proposed.

The study conducted by Nadeau et al. (2013) pro-

vided an overview of the gait analysis procedure as

well as the most signiﬁcant gait parameters and de-

viations in stroke survivors, with a focus on the im-

pacts of gait speed and the signiﬁcance of ground re-

sponse forces (GRFs). Reduced walking speed, an

unbalanced gait pattern, and a drop in peak moments

and powers on the hemiparetic side were all consid-

ered as characteristics of the hemiparetic gait follow-

ing stroke. The ﬁndings showed that While some

traits are shared by all patients, hemiparetic individu-

als may have markedly different gait patterns, even if

their walking speeds are equivalent. GRFs are useful

in evaluating the gait pattern abnormalities of stroke

patients as well as other neurologic groups, just as the

other gait characteristics.

The objective of the observational study con-

ducted by Ferrarin et al. (2015) was to evaluate the

inﬂuence of gait analysis on therapeutic decision-

making (both surgical and non-surgical) for adult pa-

tients with chronic walking difﬁculties due to stroke.

The idea was that clinical recommendations based

only on clinical examination and visual gait observa-

tion are considerably different from those based ad-

ditionally on gait analysis data. The present study’s

ﬁndings were consistent with the notion that gait anal-

ysis has a considerable impact on treatment planning

for chronic post-stroke patients with locomotor dys-

function, both surgically and non-surgically, and sup-

ports decision-making.

Through comparisons with traditional gait mea-

surements, the study proposed by Guzik and

Dru

zbicki (2020) examined the concurrent validity of

the Gait Deviation Index (GDI) as an outcome mea-

sure of gait deﬁcits at a chronic stage of recovery fol-

lowing a stroke. A group of 65 people with stroke and

65 healthy people without gait abnormalities were en-

rolled. An analytic system for movement was used to

measure the kinematic gait characteristics. The re-

sults supported the contemporaneous validity of the

GDI in post-stroke patients, but only for the afﬂicted

limb and mGDI. The authors concluded that GDI for

the unaffected limb, however, may be helpful in locat-

ing any compensatory mechanisms appearing in post-

stroke gait patterns.

In the work presented by Li et al. (2019), the

symmetry, regularity, and stability of post-stroke

hemiparetic gaits were obtained as features us-

ing the dynamic temporal warping (DTW) tech-

nique, sample entropy approach, and empirical mode

decomposition-based stability index. A cluster of

15 stroke survivors and 15 healthy control persons

participated in the studies. The ﬁndings achieved

by the authors suggested that hypothesized charac-

teristics were considerably able to distinguish post-

stroke hemiparetic patients from healthy control par-

A Gait Analysis Tool Based on Machine Learning to Support the Rehabilitation Strategy of Post-stroke Patients

401

ticipants.

Khera and Kumar (2020) proposed a review to

provide researchers with critical recommendations for

applying ML approaches for gait analysis and gait re-

habilitation. This review article demonstrated the ef-

fectiveness of ML approaches in identifying illnesses,

forecasting the period of rehabilitation, and control-

ling rehabilitation devices, making them appropriate

for clinical diagnosis.

Much effort was dedicated from the scientiﬁc

community to contribute the research ﬁeld of gait

analysis for the support in the rehabilitation of post-

stroke patients. To the best of our knowledge, no ef-

fort was provided in the automatic classiﬁcation of

muscle activity from video gait analysis, in terms of

S-EMG channels detected during the walking activi-

ties to support the decision-making strategy in the re-

habilitation of post-stroke individuals.

3 USING ML TO PREDICT THE

QUALITY OF WALKING

In this section we present GIULYO, a novel approach

for the automatic assessment of gait quality designed

to support post-stroke patients during speciﬁc rehabil-

itation therapy.

3.1 The Workﬂow of GIULYO

A motion capture system is needed to measure sub-

ject kinematics data and an electromyograph to mea-

sure the muscle activity during the walking tasks.

These two instruments provide the two sources of in-

formation necessary for GIULYO’s analysis. Once

the 3D trajectories are acquired, a features calcula-

tor module is activated which is in charge to evalu-

ate three sets of aggregate features: (i) the ﬁrst con-

tains stride measures and timing info for all subjects,

(ii) the second is about the measures derived from

the walking cycle, such as the single and double sup-

port and (iii) the third contains aggregated leg angle

and foot height data for Paretic (P) and Non-Paretic

(N) legs/feet, acquired within the laboratory reference

frame. These three sets of features—together with de-

mographic and clinical descriptors—compose the ﬁ-

nal feature vector for the classiﬁcation module. This

latter is the component aimed at providing the ﬁnal

assessment of the walking activity.

Two studies were conducted within the experi-

mentation proposed in this paper. The ﬁrst one aimed

at automatically assessing the raw number of inde-

pendent muscle co-excited in the set 2,3,4 while the

second is focused on a binary classiﬁcation in Low

and High activation of the muscle. Details about this

choice of clustering are offered in Section 4.

3.2 Gait Features

As already described, the features can be conceptually

divided into 3 distinct sets:

• Set A (Stride Measures and Timing Info):

treadmill speed, Paretic Step Ratio (PSR) aver-

aged over all steps, Paretic Step Length/Stride

Length, PSR Standard deviation, Paretic Stride

length (distance between same foot), Paretic

stride length standard deviation, non-paretic stride

length, non-paretic stride length standard devia-

tion, etc.

• Set B (Measures Derived from the Walking Cy-

cle): paretic affected leg side, left single support

percentage over all cycle, right single support per-

centage over all cycle, left single support time,

right single support time, step length, step time,

stride length and stride time standard deviation for

left and right sides, single and double support time

standard deviation for left and right sides.

• Set C (Aggregated Leg Angle and Foot Height

Data): paretic leg angle from pelvis to foot,

paretic frontal angle from pelvis to foot, non-

paretic leg angle from pelvis to foot, non-paretic

frotal angle from pelvis to foot, paretic Distance

from pelvis to foot, non-Paretic Distance from

pelvis to foot, paretic (leg length/pelvis height),

non-paretic (leg length/pelvis height), paretic ver-

tical leg distance, non-Paretic vertical leg dis-

tance, etc.

Details about how these features were registered

and obtained are available in Section 4 and in the pa-

per proposed by Routson et al. (2014).

3.3 Demographic and Clinical Features

A set of demographic and clinical features were inte-

grated to the features vector in order to increase the

knowledge of the model embedded in GIULYO. De-

mographic descriptors were gender, age and the af-

fected side (left or right) while clinical features in-

cluded individual and overall scores based on the DGI

(Jonsdottir and Cattaneo, 2007), the 6 Minutes Walk

Test (6MWT) (Enright, 2003), the Berg Balance Scale

(BBS) (Berg, 1992), and the Fugl-Meyer (FM) Score

(FUGL et al., 1975).

HEALTHINF 2023 - 16th International Conference on Health Informatics

402

3.4 Putting All Together

GIULYO combines all the features we previously

described. After the training phase, GIULYO is

able—given a walking activity—to classify it in

terms of muscle activation to describe the quality

of the rehabilitation therapy and the progresses

made. The ﬁnal features vector is composed by

the following features: anonymized subject ID,

Trial Number, Gender, Age, Affected-Side, Speed,

[DGI-1,...,DGI-8,DGI-TOT], 6MWT-Distance,

[BERG-1,...,BERG-14,BERG-TOT], [FM-1,...,FM-

17,FM-TOT,FM-Sinergy], Set-A, Set-B, Set-C,

number of Modules.

4 EMPIRICAL EVALUATION

The goal of this study is to evaluate the accuracy of

GIULYO in classifying walking trials in terms of

muscle activity in rehabilitation patients. The per-

spective is of a researcher who wants to understand

if combining several walking features is useful for the

automatic assessment of the muscle activity. Thus,

the study is steered by the following:

Research Question:

Can Machine Learning be used to predict the quality

of walking in post-stroke patients?

The above RQ is then divided into two sub-RQ:

: To what extent can GIULYO predict the

number of muscles activated during walking?

: To what extent can GIULYO predict the level

of muscle activation during walking?

With RQ

we aimed at verifying if ML can predict

the single number of muscle activated during a walk-

ing activity as an indicator of the quality of the gait

while with RQ

our purpose was to evaluate the capa-

bility of the ML in discriminating a walking activity

in low and high muscle activation.

4.1 Context Selection

The context of this study is represented by ARRA

Post-Stroke Database (Routson et al., 2014). The data

were obtained from 27 post-stroke participants and 17

healthy control participants. However, only 38 sub-

jects presented the data with the complete set of de-

scriptors. Five conditions were tested while walking

on a treadmill at intervals of 30 seconds. Examples of

such conditions include (i) self-Selected (SS) walking

pace, which the participant determined to be their typ-

ical walking speed, High Step (HS) conditions, where

subjects were asked to take as high a step as they

could while walking at their SS pace. Kinematics,

kinetics (from split belt treadmill force plates), and

electromyography data were gathered for each condi-

tion. The following tools were employed to gather the

data:

• to evaluate subject kinematics, a 12-camera mo-

tion capture system (PhaseSpace, Inc., San Lean-

dro, CA) with two linear detectors in each camera

was used. In order to establish the parameters of

segment size, the system additionally makes use

of active markers that produce infrared light and

are positioned on anatomical landmarks of a pa-

tient.

• split-belt treadmill (FIT, Bertec, Inc.) with an

inclination for measuring ground reaction forces

and moments in three dimensions

• electromyograph MA400, 16 channel EMG sys-

tem (Motion Lab Systems, Baton Rouge, LA).

For each subject, more than one gait trial is avail-

able. This is because in the experimental protocol

proposed by Routson et al. (Routson et al., 2014),

multiple trials were provided for each condition in

order to select the best recording for data analysis.

So, the dataset is composed of different information

and measures related to walking, captured through the

marker-based optoelectronic instrument. In addition,

each subject—who underwent the experimental pro-

tocol of this study—had to install a set of electrodes

for EMG signal acquisition. This was done bilater-

ally from the tibialis anterior, soleus, medial gastroc-

nemius, vastus medialis, rectus femoris, medial ham-

strings, lateral hamstrings, and gluteus medius (Rout-

son et al., 2014).

So, each walking trial can be classiﬁed in terms

of co-excited muscles or modules, thanks to the infor-

mation from EMG channels. To recap, each walking

trial was assigned a class in the set [2,3,4] to indi-

cate the number of independent co-excited muscles.

Class 2, 3, and 4 indicate that two, three, and four

modules were respectively detected by the instrumen-

tation. These numerical classes can be considered as

an index of quality of gait, as supported by several

studies which found that a higher number of indepen-

dent poststroke modules is associated with better gait

performance (Bowden et al., 2010; Clark et al., 2010).

4.2 Features Selection and Classiﬁcation

In the context of our study, we experimented two tech-

niques of features engineering: (i) ﬁrst a correlation

analysis was applied to the features vector in order to

discard the features with a correlation index greater

A Gait Analysis Tool Based on Machine Learning to Support the Rehabilitation Strategy of Post-stroke Patients

403

than 0.95 (ii) then, an automatic features selection al-

gorithm was triggered to evaluate the best descriptor

to be used as input to the classiﬁcation model.

A large set of algorithms for the features selec-

tion and for the training of the classiﬁcation model

of GIULYO was involved. Especially, we experi-

mented: Random Forest (RF) (Barandiaran, 1998),

MultiLayer Perceptron (MLP) (Pal and Mitra, 1992),

Logistic Regression (LR) (Cramer, 2002), K Nearest

Neighbour (KNN) (Dasarathy, 1991), Gaussian Naive

Bayes (GNB) (Zhang, 2004), Stochastic Gradient

Descent (SGD) (Ruder, 2016), Decision Tree (DT)

(Wu et al., 2008), Bagging Classiﬁer (BC) (Breiman,

1996), Gradient Boosting Classiﬁer (GBC) (Fried-

man, 2001), AdaBoost (AB) (Freund and Schapire,

1997), Passive Aggressive Classiﬁer (PAC) (Cram-

mer et al., 2006), Extra Trees Classiﬁer (ETC) (Geurts

et al., 2006), Support Vector Machine (SVM) (Cortes

and Vapnik, 1995).

For this experiment, we used the python library

Scikit-learn (Pedregosa et al., 2011).

4.3 Experimental Procedure

A typical Leave-1-Person Out (L1PO) cross-

validation was involved to assess the accuracy of

GIULYO. To do this, we divided the data into n

folds, one for each patient, and used—one at a

time—these folds as test set. This indicates that

a patient’s data were embedded n-1 times in the

training dataset and 1 time in the test dataset. This

method enables the creation of a classiﬁer that is

not tested and trained on the same patient’s data.

We did this in order to test the approach under the

most difﬁcult conditions possible because individual

patients’ gait data might vary consistently.

To answer this RQ

, a speciﬁc study with the raw

number of independent co-excited muscles was con-

ducted. This meant that the classiﬁcation component

of GIULYO was in charge of discriminating a gait

features vector in the classes 2, 3 or 4.

The dataset is composed by 50, 219, 208 instances

for the class 2,3, and 4 respectively.

To face RQ

, a binary ML experiment was con-

ducted. Two classes were created. The ﬁrst one

Low aimed at representing the gait records in terms of

low or poor muscle co-excitement; indeed, this class

groups the records with 2 or 3 number of modules.

The second class, namely High aimed at assess the

gait records with 4 modules.

The dataset is composed by 269, and 208 in-

stances for the class Low and High respectively.

For both studies, due to the class imbalance (es-

pecially in RQ

), the SMOTE technique was experi-

Table 1: Classiﬁcation performances achieved by the top 3

ML conﬁgurations in Study 1.

SMOTE Corr. Feat. Sel. Class. Overall Acc.

No No PAC RF 0.67

No Yes RF RF 0.63

No Yes Extra RF 0.63

Table 2: Detailed classiﬁcation performances achieved by

the best conﬁguration of GIULYO in Study 1.

Class

Classiﬁcation Metrics

Precision Recall F1-score

2 0.53 0.18 0.27

3 0.64 0.73 0.68

4 0.71 0.72 0.71

Overall Accuracy 0.67

mented (Chawla et al., 2002).

The results from the two studies conducted to an-

swer the above RQs are presented according to the

following class-level metrics: Precision, Recall, F1-

score.

5 ANALYSIS OF THE RESULTS

The analysis of the results is described according to

the speciﬁc RQ in the next subsections.

5.1 RQ

: To What Extent Can

GIULYO Predict the Number of

Muscles Activated During Walking?

The top 3 conﬁgurations of ML settings are depicted

in Table 1. The one with the best results is composed

by a PAC model for features selection and a Random

Forest for classiﬁcation. Detailed results—achieved

by this latter—during the experiment to answer RQ

are shown in Table 2. It is evident that the best perfor-

mances —at class level—are achieved by GIULYO

for the subjects who had a S-EMG with 4 modules de-

tected. Indeed, in this case, all the metrics are above

0.7. For what concerns class 3, the performance show

a decrease except for the recall value, which is 0.73.

Class 2 is the one with the poorest classiﬁcation per-

formances. However, the overall accuracy is 0.67.

HEALTHINF 2023 - 16th International Conference on Health Informatics

404

Table 3: Classiﬁcation performances achieved by the top 3

ML conﬁgurations in Study 2.

SMOTE Corr. Feat. Sel. Class. Overall Acc.

Yes Yes DT RF 0.74

Yes No PAC RF 0.73

Yes Yes Extra RF 0.73

Table 4: Classiﬁcation performances achieved by GIULYO

in Study 1.

Class

Classiﬁcation Metrics

Precision Recall F1-score

Low 0.79 0.73 0.76

High 0.68 0.75 0.71

Overall Accuracy 0.74

5.2 RQ

: To What Extent Can

GIULYO Predict the Level of

Muscle Activation During Walking?

The top 3 conﬁgurations of ML settings are depicted

in Table 3. The one with the best results is composed

by an operation of SMOTE, a correlation analysis,

a Decision Tree model for features selection and a

Random Forest algorithm for classiﬁcation. Detailed

results—achieved by this latter conﬁguration—during

the experiment to answer RQ

are shown in Table 4.

In this binary classiﬁcation experiment, the high class

shows classiﬁcation performances around 0.7 with a

peak of 0.75 for the Recall. On the other hand, the

low class shows deﬁnitely better performances, with

a F1-score of 0.76 and a Precision of 0.79. The overall

accuracy achieved by GIULYO is 0.74.

The results obtained show that a classiﬁca-

tion model based on the Random Forest algorithm

achieves the best classiﬁcation performance.

5.3 Qualitative Analysis

For sake of space limitation, the qualitative analysis

of the results achieved by GIULYO in this study is

focused only on the version of the approach with the

highest accuracy, i.e., GIULYO used to discriminate

a gait trial in Low or High class of modules triggered

(Study 2, RQ

5.3.1 Features Importance

From the L1SO validation scheme, a set of most infor-

mative instances were selected for each test set, i.e.,

for each set composed of trial data related to a single

subject. The top three features that were highly taken

into consideration by the model embedded in GIU-

LYO are: (i) the standard deviation of the leg length

measures on the paretic side evaluated during the gait

trial, (ii) the cadence, the number of steps over the

minute, (iii) the maximum measure of the leg length

and the minimum measure of the sagittal angle ob-

tained during the gait trial.

It is not surprising that the leg length measure-

ment, on the paretic side, was highly taken into ac-

count by the feature selection models as this is a de-

termining factor in the analysis of the balance of post-

stroke patients (Gardas and Shah, 2020).

Other examples of selected features include addi-

tional measures on the paretic side together with some

descriptor on the normal side.

5.3.2 GIULYO’s Performances on a Subset of

Subjects

Qualitative analysis included the observation of the

results at subject level. Within this study, we found

that the performances of GIULYO are better for a

subset of 14 subjects (out of the 38 that compose

the whole dataset), i.e., with an accuracy consistently

greater than the 0.74 overall accuracy.

On this speciﬁc cluster of subject, 162 out of 176

gait trials were correctly identiﬁed as low (2 or 3 mod-

ules detected) or high (4 modules detected). The over-

all accuracy on this group of subjects is therefore ap-

proximately 0.92. To foster this analysis, we tried to

ﬁnd the peculiarities that describe this group of sub-

jects. During this study, we found that they have a low

value of DGI-2 and an overall value of DGI smaller

than 18. DGI is a gait index proposed by Shumway-

Cook and Woollacott (Shumway-Cook and Woolla-

cott, 1995) that showed high reliability in persons

With chronic stroke and also evidence of concurrent

validity with other balance and mobility scales (Jons-

dottir and Cattaneo, 2007). The DGI includes many

tasks that are evaluated in terms of numeric indexes,

and their sum compose the overall DGI. The DGI-2 is

known as ”Change in gait speed” and it is a test where

the assisted patient has to start walking normally (for

5 minutes), and then walking quickly, for 5 minutes

with ﬁve slow steps at the end (Jonsdottir and Catta-

neo, 2007).

The DGI overall score of 19 or below indicates a

risk of falling in older adults and those with vestibular

impairment (Whitney et al., 2000, 1998).

A Gait Analysis Tool Based on Machine Learning to Support the Rehabilitation Strategy of Post-stroke Patients

405

6 LIMITATIONS OF THE STUDY

The followings could represent threats to validity of

this study:

• The reduction to a binary state classiﬁcation could

seem a choice far from practical use in rehabilita-

tion. However, we believe that this choice may

provide a rapid and more accurate screening of

muscle activity, with respect to the raw number

of modules. This should represent an initial indi-

cation for the medical staff.

• The dataset used in this study may be too small

for the purposes of a medical study. However, it

was one of the public datasets—found in the rel-

evant literature—with EMG, motion tracking and

clinical assessment data for a substantial number

of patients. It is our opinion that the results of this

study should be considered as preliminary.

• There is no comparison between GIULYO and a

scientiﬁc baseline. To the best of our knowledge,

there is not a study to compare with.

7 CONCLUSIONS

This paper proposed GIULYO, a tool designed for

the support of the rehabilitation strategy of post stroke

patients. The tool is device-agnostic, meaning that

no speciﬁc motion tracking instrumentation is needed

to make it work and it is designed to provide an as-

sessment of the gait in terms of a novel indication,

the number of muscle group co-excited. This infor-

mation could be raw (i.e., the exact number) or clus-

tered (i.e., low or high muscle activation). GIULYO

was validated on the ARRA post stroke dataset. Re-

sults showed an overall accuracy of 0.74 in the bi-

nary case, and of 0.92 on a speciﬁc subset of pa-

tients, therefore demonstrating higher performances

when subjects have poor locomotor skills, according

to the DGI. GIULYO aims at be embedded in mod-

ern Decision Support Systems (DSS) for the support

to the medical equipe thanks to a rapid screening of

the assisted patients.

Future works will be devoted to validate GIU-

LYO (i) on a larger set of patients and (ii) on data

acquired from another motion tracking instrument to

verify how the accuracy of the 3d trajectories may af-

fect the precision of the automatic classiﬁcation.

ACKNOWLEDGMENT

The authors have been supported by the project

“EDAM“: A Diagnosis Recommender System based

on Explainable Artiﬁcial Intelligence and the Com-

bination of Motion Analysis and Others Clinical

Biomarkers funded by the Italian Ministry of De-

fense.

Special thank to A. Ciccotelli for the support in

the generation of the features vector.

REFERENCES

Allen, J. L. and Neptune, R. R. (2012). Three-dimensional

modular control of human walking. Journal of biome-

chanics, 45(12):2157–2163.

Barandiaran, I. (1998). The random subspace method for

constructing decision forests. IEEE Trans. Pattern

Anal. Mach. Intell, 20(8):1–22.

Berg, K. (1992). Measuring balance in the elderly: Devel-

opment and validation of an instrument.

Bowden, M. G., Clark, D. J., and Kautz, S. A. (2010). Eval-

uation of abnormal synergy patterns poststroke: rela-

tionship of the fugl-meyer assessment to hemiparetic

locomotion. Neurorehabilitation and neural repair,

24(4):328–337.

Breiman, L. (1996). Bagging predictors. Machine learning,

24(2):123–140.

Chawla, N. V., Bowyer, K. W., Hall, L. O., and Kegelmeyer,

W. P. (2002). Smote: synthetic minority over-

sampling technique. Journal of artiﬁcial intelligence

research, 16:321–357.

Clark, D. J., Ting, L. H., Zajac, F. E., Neptune, R. R.,

and Kautz, S. A. (2010). Merging of healthy motor

modules predicts reduced locomotor performance and

muscle coordination complexity post-stroke. Journal

of neurophysiology, 103(2):844–857.

Cortes, C. and Vapnik, V. (1995). Support-vector networks.

Machine learning, 20(3):273–297.

Cramer, J. S. (2002). The origins of logistic regression

(technical report). In Tinbergen Institute.

Crammer, K., Dekel, O., Keshet, J., Shalev-Shwartz, S.,

and Singer, Y. (2006). Online passive aggressive al-

gorithms.

Dasarathy, B. V. (1991). Nearest neighbor (nn) norms: Nn

pattern classiﬁcation techniques. IEEE Computer So-

ciety Tutorial.

Enright, P. L. (2003). The six-minute walk test. Respiratory

care, 48(8):783–785.

Ferrarin, M., Rabuffetti, M., Bacchini, M., Casiraghi, A.,

Castagna, A., Pizzi, A., and Montesano, A. (2015).

Does gait analysis change clinical decision-making in

poststroke patients? results from a pragmatic prospec-

tive observational study. Eur J Phys Rehabil Med,

51(2):171–84.

Freund, Y. and Schapire, R. E. (1997). A decision-theoretic

generalization of on-line learning and an application

HEALTHINF 2023 - 16th International Conference on Health Informatics

406

to boosting. Journal of computer and system sciences,

55(1):119–139.

Friedman, J. H. (2001). Greedy function approximation: a

gradient boosting machine. Annals of statistics, pages

1189–1232.

FUGL, M. et al. (1975). The post-stroke hemiplegic patient.

i. a method for evaluation of physical performance.

Gardas, S. and Shah, H. (2020). Inﬂuence of leg length dis-

crepancy on balance and gait in post-stroke patients:

a correlational study. Bulletin of Faculty of Physical

Therapy, 25(1):1–9.

Geurts, P., Ernst, D., and Wehenkel, L. (2006). Extremely

randomized trees. Machine learning, 63(1):3–42.

Guzik, A. and Dru

zbicki, M. (2020). Application of the gait

deviation index in the analysis of post-stroke hemi-

paretic gait. Journal of Biomechanics, 99:109575.

Jonsdottir, J. and Cattaneo, D. (2007). Reliability and valid-

ity of the dynamic gait index in persons with chronic

stroke. Archives of physical medicine and rehabilita-

tion, 88(11):1410–1415.

Khera, P. and Kumar, N. (2020). Role of machine learning

in gait analysis: a review. Journal of Medical Engi-

neering & Technology, 44(8):441–467.

Li, M., Tian, S., Sun, L., and Chen, X. (2019). Gait analysis

for post-stroke hemiparetic patient by multi-features

fusion method. Sensors, 19(7):1737.

Liu, Y., Liu, B., Zhou, Z., Cai, S., and Xie, L. (2021).

A novel center of mass (com) perception approach

for lower-limbs stroke rehabilitation. In Interna-

tional Conference on Social Robotics, pages 606–615.

Springer.

Mirelman, A., Patritti, B. L., Bonato, P., and Deutsch,

J. E. (2010). Effects of virtual reality training on gait

biomechanics of individuals post-stroke. Gait & pos-

ture, 31(4):433–437.

Nadeau, S., Betschart, M., and Bethoux, F. (2013). Gait

analysis for poststroke rehabilitation: the relevance

of biomechanical analysis and the impact of gait

speed. Physical Medicine and Rehabilitation Clinics,

24(2):265–276.

Neptune, R. R., Clark, D. J., and Kautz, S. A. (2009). Mod-

ular control of human walking: a simulation study.

Journal of biomechanics, 42(9):1282–1287.

Pal, S. and Mitra, S. (1992). Multilayer perceptron, fuzzy

sets, and classiﬁcation. IEEE Transactions on Neural

Networks, 3(5):683–697.

Pedregosa, F., Varoquaux, G., Gramfort, A., Michel, V.,

Thirion, B., Grisel, O., Blondel, M., Prettenhofer,

P., Weiss, R., Dubourg, V., Vanderplas, J., Passos,

A., Cournapeau, D., Brucher, M., Perrot, M., and

Duchesnay, E. (2011). Scikit-learn: Machine learning

in Python. Journal of Machine Learning Research,

12:2825–2830.

Routson, R. L., Kautz, S. A., and Neptune, R. R. (2014).

Modular organization across changing task demands

in healthy and poststroke gait. Physiological reports,

2(6):e12055.

Ruder, S. (2016). An overview of gradient de-

scent optimization algorithms. arXiv preprint

arXiv:1609.04747.

Shumway-Cook, A. and Woollacott, M. H. (1995). Theory

and practical applications. Motor Control.

Swank, C., Almutairi, S., Wang-Price, S., and Gao, F.

(2020). Immediate kinematic and muscle activity

changes after a single robotic exoskeleton walking

session post-stroke. Topics in Stroke Rehabilitation,

27(7):503–515.

Warlow, C. (1998). Epidemiology of stroke. The Lancet,

352:S1–S4.

Warlow, C. P., Dennis, M., Gijn, J. v., Hankey, G., Sander-

cock, P., Bamford, J., Wardlaw, J., and Brown, M. M.

(1997). Stroke: a practical guide to management.

BMJ-British Medical Journal-International Edition,

314(7097):1840.

Whitney, S., Hudak, M., and Marchetti, G. (2000). The

dynamic gait index relates to self-reported fall history

in individuals with vestibular dysfunction. Journal of

Vestibular Research, 10(2):99–105.

Whitney, S. L., Poole, J. L., and Cass, S. P. (1998). A review

of balance instruments for older adults. The American

Journal of Occupational Therapy, 52(8):666–671.

Wu, X., Kumar, V., Ross Quinlan, J., Ghosh, J., Yang, Q.,

Motoda, H., McLachlan, G. J., Ng, A., Liu, B., Yu,

P. S., et al. (2008). Top 10 algorithms in data mining.

Knowledge and information systems, 14(1):1–37.

Zhang, H. (2004). The optimality of naive bayes. Aa,

1(2):3.

A Gait Analysis Tool Based on Machine Learning to Support the Rehabilitation Strategy of Post-stroke Patients

407