Integrating Gait and Clinical Data with Explainable Artiﬁcial

Intelligence for Parkinson’s Prediction: The EDAM System

Nicoletta Balletti

1,2

, Emanuela Guglielmi

, Gennaro Laudato

Rocco Oliveto

2,3

, Jonathan Simeone

and Roberto Zinni

Center for Biotechnology, Institute of Biomedical Sciences of the Ministry of Defense, Rome, Italy

University of Molise, Pesche (IS), Italy

Datasound srl, Pesche (IS), Italy

WordPower, San Salvo (CH), Italy

Keywords:

Gait Analysis, Clinical Biomarkers, Explainable Artiﬁcial Intelligence, Parkinson’s Disease Detection.

Abstract:

Several machine learning (ML) approaches have been introduced for gait and posture analysis, recognized as

crucial for early diagnosing neurological disorders, particularly Parkinson’s disease. However, these exist-

ing methods are often limited by their lack of integration with other clinical biomarkers and their inability to

provide transparent, explainable predictions. To overcome these limitations, we introduce EDAM (Explain-

able Diagnosis Recommender), a system that leverages Explainable Artiﬁcial Intelligence (XAI) techniques

to deliver both accurate predictions and clear, interpretable explanations of its diagnostic decisions. We eval-

uate the capabilities of EDAM in two main areas: distinguishing between healthy individuals and those with

Parkinson’s disease, and classifying abnormal gait patterns that may indicate early-stage Parkinson’s disease.

To ensure a comprehensive evaluation, we constructed one of the largest known dataset by merging and stan-

dardizing several existing datasets. This dataset includes 557 features and 7,303 labelled instances, covering

a wide range of gait patterns and clinical features. Results show that EDAM achieves high accuracy in both

tasks, demonstrating its potential for early detection of neurological disorders.

1 INTRODUCTION

The analysis of gait and posture (motion analysis)

is crucial for the early diagnosis of several patholo-

gies, especially neurological disorders, as well as for

monitoring disease progression and evaluating a pa-

tient’s therapeutic response (Buckley et al., 2019).

Research has shown that by monitoring upper-body

movements, it is possible to differentiate between

healthy individuals and those with Parkinson’s dis-

ease, while patients with ataxic symptoms, such as

those with multiple sclerosis, exhibit deﬁcits in postu-

ral control. Moreover, slow gait has been identiﬁed as

a predictor of dementia, with early signs manifesting

up to nine years before an ofﬁcial diagnosis is made

(Buckley et al., 2019).

In the literature, several machine learning (ML)

approaches have been proposed for monitoring and

predicting speciﬁc diseases using gait and posture

data acquired through motion analysis systems (e.g.,

(Abdulhay et al., 2018; Costa et al., 2016; Cuzzolin

et al., 2017; Mannini et al., 2016; Raknim and Lan,

2016)). However, these approaches often rely on uni-

variate analyses, treating posture and gait data in iso-

lation from other clinical biomarkers. While this sim-

pliﬁcation can streamline processing, it may overlook

crucial insights that could emerge from a more com-

prehensive, multivariate analysis (Holzinger et al.,

2017a). Additionally, the robustness of these systems

is a recurring concern. Many studies are based on rel-

atively small patient samples, which limits the gener-

alizability of the ﬁndings (Buckley et al., 2019).

Another signiﬁcant limitation of ML-based mo-

tion analysis systems is their lack of transparency.

The predictions made by these models are frequently

perceived as “black-box” decisions, leaving special-

ists unable to grasp the reasoning behind the outputs

(Holzinger et al., 2017b). This opacity can reduce

trust in the system, prompting specialists to dismiss

even accurate predictions. To address this challenge,

there is an increasing need for systems that offer ex-

Balletti, N., Guglielmi, E., Laudato, G., Oliveto, R., Simeone, J. and Zinni, R.

Integrating Gait and Clinical Data with Explainable Artiﬁcial Intelligence for Parkinson’s Prediction: The EDAM System.

DOI: 10.5220/0013179400003911

Paper published under CC license (CC BY-NC-ND 4.0)

In Proceedings of the 18th International Joint Conference on Biomedical Engineering Systems and Technologies (BIOSTEC 2025) - Volume 2: HEALTHINF, pages 129-140

ISBN: 978-989-758-731-3; ISSN: 2184-4305

129

plainable and transparent predictions. Explainable

Artiﬁcial Intelligence (XAI) techniques have emerged

as a promising solution, enabling users to understand

the rationale behind ML model outputs (Edwards and

Veale, 2017). These techniques are particularly valu-

able in the medical ﬁeld, where practitioners must in-

terpret complex, heterogeneous data. AI methods, es-

pecially ML, are crucial for extracting actionable in-

sights from such data, and XAI enhances this process

by making predictions interpretable and useful for hu-

man experts.

To overcome the limitations of current ap-

proaches, in this paper, we present EDAM (Explain-

able Diagnosis Recommender), a system designed to

support predictions related to neurological disorders

by integrating posture and gait data with other clini-

cal biomarkers. The Decision Support System (DSS)

within EDAM—based on advanced ML techniques

(Hastie et al., 2009)—assists specialists in two closely

related tasks: distinguishing between healthy individ-

uals and those with Parkinson’s disease, and detecting

abnormal gait patterns, thereby facilitating the early

detection of neurological pathologies.

EDAM applies XAI techniques to generate trans-

parent, interpretable predictions. By leveraging ex-

plainable machine learning algorithms, EDAM pro-

vides specialists with not only accurate predictions

but also visual and natural-language explanations that

clearly outline the factors inﬂuencing the diagnosis.

Additionally, EDAM generates preliminary diagnos-

tic reports (pre-reports) to help specialists analyze the

results both quantitatively and qualitatively. These

user-friendly predictions enhance the interpretation

of data-driven insights, allowing specialists to seam-

lessly integrate their intuition, judgment, and experi-

ence into the decision-making process.

To evaluate the accuracy of EDAM, we con-

ducted a study using a dataset that was constructed

by homogenizing and merging several gait datasets

from the literature (i.e., datasets provided by Mehrizi

et al., 2019,Schreiber and Moissenet, 2019,Jun et al.,

2020, Kour et al., 2020). This process enabled us to

create the largest dataset in the literature, consisting

of 557 features and 7,303 labeled instances, making

it a highly comprehensive resource for gait analysis

and Parkinson’s disease prediction. Besides evaluat-

ing the accuracy of EDAM DSS, the study allowed for

an extensive evaluation of the importance of all gait-

related features in both predicting Parkinson’s dis-

ease and classifying gait types. The results demon-

strated that EDAM achieved high accuracy in both

tasks, with notable performance in detecting early

signs of Parkinson’s disease. Speciﬁcally, the classiﬁ-

cation model was able to distinguish between Parkin-

sonian and healthy subjects with a high degree of pre-

cision, and it accurately classiﬁed various gait pat-

terns, including those associated with early neurolog-

ical symptoms. These ﬁndings underscore the poten-

tial of EDAM in supporting early diagnosis through

gait analysis.

Thus, the speciﬁc contributions of the paper can

be summarized as follows:

• the introduction of EDAM, emphasizing its key

features and the integration of explainable AI

techniques for clinical decision support;

• the creation of the largest dataset (to our knowl-

edge) by merging several existing datasets, en-

abling the most comprehensive validation of

machine learning-based prediction models for

Parkinson’s disease detection and gait classiﬁca-

tion;

• an extensive evaluation of the effectiveness of

EDAM in predicting Parkinson’s disease and clas-

sifying various gait patterns. The evaluation

leverages a wide range of features, such as 3D

joint trajectories, rotations, step analysis, and

energy images (Han and Bhanu, 2005)—more

than any prior study has combined simultane-

ously—allowing for a comprehensive analysis of

their collective impact on model performance.

The remainder of the paper is organized as fol-

lows. In Section 2 we provide an overview of the

EDAM system, while in Section 3 and Section 4 we

present the study we conducted to evaluate the EDAM

DSS and the achieved results, respectively. Section 6

concludes the paper, after a discussion of the related

literature (Section 5).

2 EDAM OVERVIEW

EDAM (Explainable Diagnosis Recommender) is an

advanced diagnosis support system integrating mo-

tion analysis with other clinical biomarkers, enhanc-

ing disease prediction and early diagnosis through

XAI. The system addresses several limitations of tra-

ditional ML-based diagnostic approaches, which of-

ten focus narrowly on single data streams and deliver

predictions without offering comprehensible explana-

tions to specialists.

2.1 Architecture and Functionality

The EDAM system is built upon a modular architec-

ture that enables it to process data from various mo-

tion analysis systems and clinical devices, providing

HEALTHINF 2025 - 18th International Conference on Health Informatics

130

both diagnostic insights and explainability for medi-

cal professionals. The key components and function-

alities of EDAM are depicted in Figure 1.

EDAM is designed to capture data from a wide

range of devices, including motion-tracking systems

such as Vicon and Azure Kinect DK, heart rate mon-

itors like the Polar H10, and EEG/EMG sensors such

as DSI-7 and Cometa Mini Wave. The system is

device-agnostic, allowing seamless integration of new

sensors with minimal effort through the use of dedi-

cated drivers. These drivers ensure that data from dif-

ferent devices is formatted into a standardized struc-

ture, enabling consistent and uniform processing by

the system. This ﬂexibility allows EDAM to collect

a comprehensive set of data, including gait parame-

ters, heart rate, EEG, and EMG signals, providing a

richer, multidimensional analysis that surpasses tradi-

tional univariate motion analysis approaches.

The acquired data are stored in a hybrid database

system: A relational database for medical records and

a NoSQL database for sensor-derived data like gait

dynamics and clinical biomarkers. This architecture

allows for efﬁcient querying and ﬂexible data man-

agement, facilitating real-time and historical analyses.

The core of EDAM is its Decision Support System

(DSS), which leverages pre-trained machine learn-

ing models to automatically diagnose speciﬁc dis-

eases based on collected data. EDAM currently in-

cludes models for detecting gait deviations, predict-

ing Parkinson’s disease, and calculating the Dynamic

Gait Index (DGI) (Shumway-Cook and Woollacott,

1995) (Balletti et al., 2024). Its ﬂexible design al-

lows for the easy integration of additional predictive

models through standardized APIs, ensuring scalabil-

ity for future enhancements.

A standout feature of EDAM is its Explainable AI

(XAI) module, which uses SHapley Additive exPla-

nations (SHAP) (Lundberg and Lee, 2017) to pro-

vide interpretable insights into the model’s predic-

tions. SHAP values highlight the individual and com-

bined contributions of features such as gait speed or

EMG signals to a diagnosis. This interpretability is

crucial in clinical settings, where understanding the

reasoning behind a diagnosis fosters trust and sup-

ports informed decision-making by specialists.

The XAI module also generates textual explana-

tions, clearly outlining the key factors that inﬂuenced

the prediction. These pre-reports are designed for

easy interpretation by medical professionals, offer-

ing transparency and aiding early diagnosis. Addi-

tionally, the system incorporates visual aids like force

plots and image analyses (e.g., Gait Energy Images

and Skeleton Energy Images (Han and Bhanu, 2005))

to further illustrate the inﬂuence of speciﬁc features

on the predictions.

The continuous learning capability of EDAM en-

sures it stays current with evolving clinical knowl-

edge. After each diagnosis, validated data from spe-

cialists are added to the knowledge base, and once

enough new data is gathered, the ML models are re-

trained to enhance accuracy, keeping EDAM up-to-

date with the latest advancements.

2.2 Data Acquisition and Analysis

The EDAM system facilitates gait data acquisition

and analysis by integrating various sensor types. The

process starts with the technical operator selecting a

patient from a list. Once the patient is chosen, the

operator initiates the data acquisition by selecting the

“Acquisition” option, which opens a window to spec-

ify the examination type. The operator then selects

the data sources, including posture (kinematics), EEG

(brain activity), EMG (muscle activity), and heart

rate. While posture data is mandatory, other sources

are optional depending on the exam’s focus. Appro-

priate devices are chosen for each source. If EMG

data is included, the system provides a conﬁguration

page where the operator assigns sensors to muscles

using a drag-and-drop interface (see Figure 2), with

visual indicators conﬁrming correct sensor-to-muscle

pairing.

After conﬁguration, the system presents both a

visual and textual guide for the patient, including a

video demonstration of the required movement and

a detailed text description. This ensures the patient

clearly understands the task to be performed.

During acquisition, the system captures real-time

data from the selected sources and displays it for the

operator. A 3D avatar visually represents the patient’s

movements, providing front, left, and right views of

the gait (see Figure 3). The operator can zoom in on

speciﬁc body parts for a closer examination of move-

ment details. Device status is shown using color-

coded icons: green for successful data transmission,

red for errors, and yellow for connection attempts.

After the exercise is completed, the operator selects

“End Acquisition” to conclude the session, at which

point a window appears for adding session notes,

which can be saved or discarded based on the oper-

ator’s evaluation.

After the data is acquired, it is analyzed through

a comprehensive dashboard. The operator can select

speciﬁc joints or muscles for detailed analysis, dis-

played in graphical form. If EMG data was collected,

the system allows the operator to choose speciﬁc mus-

cles for analysis; otherwise, it directly displays graphs

of the available data (see Figure 4). The dashboard is

Integrating Gait and Clinical Data with Explainable Artiﬁcial Intelligence for Parkinson’s Prediction: The EDAM System

131

EDAM

Driver 1 Driver n

…

Data Acquisition

Decision Support System

Model 1

…

Model 2 Model m

XAI Module

Database management

Gait Heart EEG EMG

Driver 1 Driver n

…

Driver 1 Driver n

…

Driver 1 Driver n

…

NoSQL DBRelational DB

Figure 1: The EDAM architecture.

Figure 2: Conﬁguration page for EMG data acquisition.

Figure 3: Real-time data acquisition interface of EDAM.

divided into two sections: on the left, a video of the

gait cycle is shown alongside predictions generated by

machine learning models and the Dynamic Gait In-

dex (DGI); on the right, synchronized graphs display

data from posture, EMG, EEG, and heart rate sensors.

These synchronized visuals provide a holistic view of

the data captured during the session.

The Explainable AI (XAI) module of EDAM pro-

vides detailed insights into how predictions are made,

offering visual aids such as force plots and image-

based analyses like Gait Energy Images (GEI) and

Skeleton Energy Images (SEI) to highlight the fea-

tures inﬂuencing predictions (see Figure 5). For in-

stance, the system can identify speciﬁc areas of the

body contributing to the classiﬁcation of walking pat-

terns or the diagnosis of conditions such as Parkin-

son’s disease (see Figure 5). This transparency fosters

trust by helping medical professionals understand the

reasoning behind the system’s predictions.

Additionally, the operator can generate a pre-

report that summarizes the collected data and anal-

ysis. This report, which includes system-generated

predictions and explanations (see Figure 5), as well

as any observations made by the operator, can be re-

viewed and reﬁned by specialists, such as physiatrists.

3 EVALUATION OF EDAM

PARKINSON’S DETECTION

CAPABILITIES

We conducted an empirical evaluation to assess the

effectiveness of the EDAM DSS in two key areas:

(i) automatically distinguishing between individuals

with Parkinson’s disease and healthy subjects, and (ii)

classifying gait patterns to support the early detec-

tion of Parkinson’s disease. This second study focuses

speciﬁcally on identifying subtle gait deviations that

may indicate the early stages of Parkinson’s disease.

HEALTHINF 2025 - 18th International Conference on Health Informatics

132

Figure 4: EDAM dashboard displaying gait cycle video, predictions, and synchronized sensor data.

Figure 5: Explainable AI (XAI) Interface Displaying Gait

Energy Images, Force Plots, and System-Generated Predic-

tions for Transparent Analysis.

3.1 Study Deﬁnition

The study is guided by the following interconnected

research questions:

• RQ

: To what extent can EDAM distinguish a

healthy subject from a subject with Parkinson’s

disease?

• RQ

: To what extent can EDAM automatically

classify a subject’s gait to support early Parkin-

son’s detection?

The ﬁrst research question assesses EDAM ac-

curacy in distinguishing between healthy individuals

and those with Parkinson’s disease. Building on this,

the second research question evaluates EDAM ability

to classify a subject’s gait into one of six categories:

Antalgic, Lurch, Normal, Steppage, Stiff-Legged, and

Trendelenburg. This link between the two research

questions is crucial, as certain gait types, such as Step-

page or Stiff-Legged, are known to be early indica-

tors of Parkinson’s disease. By accurately classifying

these gait patterns, EDAM can detect subtle devia-

tions in movement that may suggest the early onset

of Parkinson’s, thereby facilitating early diagnosis.

Thus, while the ﬁrst question focuses on distinguish-

ing known cases of Parkinson’s, the second question

explores EDAM potential for early detection through

gait analysis.

3.2 Context of the Study

To evaluate EDAM’s capabilities in detecting Parkin-

son’s disease and classifying gait types, we compiled

a comprehensive and standardized dataset by integrat-

ing several well-established datasets from the litera-

ture, including those referenced in the following stud-

ies:

• Mehrizi et al., 2019: This dataset includes gait

recordings from 23 patients with Parkinson’s dis-

ease, 22 with postural stroke, 25 with orthope-

dic issues, and 25 healthy controls. Participants

walked on a treadmill for about a minute, while

two digital cameras captured their movement, and

a motion capture system tracked reﬂective mark-

ers placed on key body joints. The dataset in-

cludes 24 time series representing the 3D position

of 8 body joints in three directions (x, y, z), col-

lected at a 100 Hz sampling rate. The goal was to

detect health problems related to gait using deep

Integrating Gait and Clinical Data with Explainable Artiﬁcial Intelligence for Parkinson’s Prediction: The EDAM System

133

neural networks for pose estimation.

• Schreiber and Moissenet, 2019: This dataset con-

tains gait data from 50 participants (24 women

and 26 men), acquired during a single session

where participants walked naturally on a 10-meter

walkway. Five walking speed conditions were

recorded, from slow to fast, using a 10-camera op-

toelectronic system sampling at 100 Hz. The sys-

tem tracked 3D trajectories of 52 reﬂective skin

markers placed on anatomical landmarks. The

dataset is designed for analyzing human gait at

different walking speeds.

• Jun et al., 2020: This dataset consists of gait data

from 10 healthy participants, each simulating ﬁve

different pathological gaits (antalgic, stiff-legged,

lurch, steppage, and Trendelenburg). The data

was captured using six Microsoft Kinect v2 sen-

sors, providing 3D coordinates for 25 body joints.

Each participant walked 20 times per gait type,

generating 3D skeletal data for the classiﬁcation

of different gait patterns. The simulation guide-

lines focused on replicating mechanical limita-

tions, such as restricted joint movement.

• Kour et al., 2020: This dataset includes gait

recordings from 96 subjects: 50 with knee os-

teoarthritis, 16 with Parkinson’s disease, and 30

healthy individuals. Each participant performed

two gait sequences (left to right and right to left) in

the frontal/sagittal plane. The data was collected

using a NIKON DSLR camera positioned 8 me-

ters away from the walking path, and six passive

reﬂective markers attached to the subjects’ joints.

The dataset is in video format (.mov), designed

for analyzing gait differences between healthy in-

dividuals and those with musculoskeletal or neu-

rological conditions.

These datasets were initially developed for train-

ing machine learning (ML) models in gait analysis

and Parkinson’s disease prediction. Still, they varied

in terms of the features they provided, leading to chal-

lenges in ensuring reliable and reproducible model

performance across different clinical and commercial

settings. To address this issue, we merged instances

from various datasets and enhanced the dataset by cal-

culating missing features, creating a uniﬁed resource

that covers all relevant domains.

The analysis of these datasets indicated that the

most commonly used gait features fall into three do-

mains: (i) 3D trajectories of body joints, (ii) step-

related features such as swing and stance phases, and

(iii) Gait Energy Image (GEI) and Skeleton Energy

Image (SEI) data. However, none of the original

datasets contained a complete set of features from all

domains. Therefore, we employed an extensive fea-

ture engineering process to create a cohesive dataset.

For datasets that contained only 3D body joint tra-

jectories, we derived additional step-related features

(e.g., swing and stance) using custom-built feature en-

gineering strategies (Amboni et al., 2021). Further-

more, using the 3D trajectory data, we animated a

mannequin and reconstructed GEI and SEI data, pro-

viding a more comprehensive feature set. In cases

where only exercise execution videos were available,

we utilized the Plask tool

to extract 3D trajectories,

allowing us to compute all necessary features using

the same procedures applied to the other datasets.

Through this process, we constructed the largest

dataset ever used in the literature for gait classiﬁca-

tion and Parkinson’s disease prediction, featuring a

total of 557 distinct features. These include 304 fea-

tures related to joint rotations, 228 features represent-

ing body trajectories, 17 features capturing pitch, and

6 features derived from GEI and SEI data. The dataset

encompasses 7,303 labelled instances, categorized as

follows: 20 instances labelled as Parkinson’s disease,

34 as healthy controls, and the remaining instances

divided among six different gait types (e.g., antalgic,

lurch).

3.3 Experimental Procedure

To address RQ

and RQ

, we adopted a Leave-

One-Subject-Out (L1SO) cross-validation approach

(Hastie et al., 2009). In this method, the dataset is

divided into n folds, each corresponding to a differ-

ent patient. For each iteration, one fold is used as the

test set, while the remaining n-1 folds are used for

training. This ensures that the data of each patient are

included in the training set n-1 times and in the test

set only once, preventing the model from being tested

on data from the same patient. This design simulates

a real-world scenario where predictions are made for

a patient being tested for the ﬁrst time.

As previously mentioned, the dataset used for the

experiments consists of 7,303 labeled instances. For

, we used the 20 instances labelled as Parkinson’s

disease and the 34 instances labelled as healthy, to-

talling 54 instances. For RQ

, the remaining 7,093 in-

stances, corresponding to the six different gait types,

were used to evaluate EDAM gait classiﬁcation accu-

racy.

We evaluated the proposed approach in two differ-

ent scenarios: Lower body, where only features from

the lower part of the body were used, and Full body,

where features from both the lower and upper body

were provided as input.

https://plask.ai/en-US

HEALTHINF 2025 - 18th International Conference on Health Informatics

134

During the experimentation, we applied two fea-

ture engineering techniques: (i) correlation analysis to

discard features with a correlation index higher than

0.95 (Guyon and Elisseeff, 2003), and (ii) automatic

feature selection to identify the most relevant descrip-

tors for classiﬁcation (Li et al., 2017).

We tested 13 different machine learning models

(Hastie et al., 2009): Random Forest (RF), Multi-

Layer Perceptron (MLP), Logistic Regression (LR),

K-Nearest Neighbors (KNN), Gaussian Naive Bayes

(GNB), Stochastic Gradient Descent (SGD), Deci-

sion Tree (DT), Bagging Classiﬁer (BC), Gradient

Boosting Classiﬁer (GBC), AdaBoost (AB), Passive

Aggressive Classiﬁer (PAC), Extra Trees Classiﬁer

(ETC), and Support Vector Machine (SVM).

3.4 Evaluation Metrics

The following metrics (Hastie et al., 2009) were used

to evaluate the performance of the EDAM DSS and

address our research questions:

• Accuracy: The ratio of correctly classiﬁed in-

stances to the total number of instances.

Accuracy =

T P + T N

T P + T N + FP + FN

• Precision: The ratio of correctly classiﬁed pos-

itive instances to the total number of instances

classiﬁed as positive.

Precision =

T P

T P + FP

• Recall: The ratio of correctly classiﬁed positive

instances to the sum of correctly classiﬁed posi-

tive instances and those incorrectly classiﬁed as

negative.

Recall =

T P

T P + FN

• F1-Score: The harmonic mean of precision and

recall.

F1-score = 2 ×

Precision × Recall

Precision + Recall

Before evaluating the machine learning models in

relation to our research questions, we performed a

feature analysis using Principal Component Analysis

(PCA) (Wold et al., 1987) to identify the most relevant

features for (i) predicting Parkinson’s disease and (ii)

classifying gait. This analysis was carried out for both

the Lower body and Full body scenarios.

4 ANALYSIS OF THE RESULTS

In this section, we present the analysis of results for

the two research questions (RQs) of our study.

4.1 RQ

: Parkinson’s Prediction

The PCA analysis conducted for both the Lower Body

and Full Body scenarios revealed important insights

regarding the relevance of various feature domains in

predicting Parkinson’s disease. In the Lower Body

scenario, PCA showed that 32 features explained

97.8% of the variance (see Figure 6), including 20

from the rotation domain, 7 from the pitch domain,

4 from trajectories, and 1 from the GEI/SEI domain.

This result is signiﬁcant as it highlights the contribu-

tion of each feature domain, particularly the rotation

domain, which was introduced in EDAM and has not

been extensively explored in prior studies.

Figure 6: Parkinson (Lower body): Variance Analysis.

In the Full Body scenario, PCA revealed that 33

features explained 97.5% of the variance (see Fig-

ure 7). Notably, none of the features from the

GEI/SEI domain were included in this set; the fea-

tures were instead distributed across the rotation (25

features), trajectory (6 features), and pitch (2 features)

domains. This suggests that GEI/SEI features may

be more valuable when only lower body information

is available. Importantly, the rotation features once

again proved to be crucial in identifying Parkinson’s,

reinforcing their relevance across both scenarios.

Focusing on the accuracy of detecting subjects

with Parkinson’s disease, experimental results indi-

cate that in the Lower Body scenario, 170 out of the

384 tested machine learning pipelines achieved 100%

accuracy. Similarly, in the Full Body scenario, 168

pipelines reached perfect accuracy. While these re-

sults are highly encouraging, further experimentation

is necessary to validate the generalizability of the

ﬁndings across more diverse datasets.

Nevertheless, it is worth noting that the combi-

nation of numerical features, such as 3D joint tra-

jectories and rotational data, with graphical features

like GEI and SEI, clearly enhances the overall predic-

Integrating Gait and Clinical Data with Explainable Artiﬁcial Intelligence for Parkinson’s Prediction: The EDAM System

135

Figure 7: Parkinson (Full body): Variance Analysis.

tion accuracy. As shown in Table 1, a system rely-

ing solely on GEI and SEI features achieved an accu-

racy of approximately 94%, underscoring their valu-

able contribution to the prediction process.

Table 1: Performance of EDAM Parkinson’s prediction

model based exclusively on GEI and SEI features.

Accuracy Precision Recall F1-Score

Mean 0.94 1.00 0.94 0.95

Median 1.00 1.00 1.00 1.00

Std.dev 0.21 0.00 0.21 0.18

To further analyze the contribution of individual

features, we examined one of the models with the

highest accuracy, which was based on a Decision

Tree. This model did not use synthetic oversampling

or feature correlation analysis but applied an auto-

matic feature selection algorithm based on Random

Forest. Decision Trees were chosen not only for their

accuracy but also for their interpretability, making

them well-suited for the EDAM system, which aims

to generate understandable preliminary reports (pre-

reports) based on predictions.

As seen in Table 2, the distribution of selected

features across the four domains (Rotations, Trajec-

tories, Step, and GEI/SEI) for both the Lower Body

and Full Body scenarios reinforces the ﬁndings from

the PCA analysis. This highlights the importance of

considering all feature domains in knowledge base of

EDAM, with rotation and trajectory features proving

especially inﬂuential.

Answer to RQ

. The evaluation of the EDAM DSS

demonstrates that the selected machine learning mod-

els, in both the Lower Body and Full Body scenarios,

achieved high accuracy in predicting Parkinson’s dis-

ease. The distribution of features across the domains

of rotations, trajectories, step, and GEI/SEI under-

scores the importance of each domain in the knowl-

Table 2: Feature distribution across Lower Body and Full

Body scenarios.

Domain Lower Body Full Body

Rotations 208 304

Trajectories 132 228

Step 13 17

GEI/SEI 6 6

Total Features 359 557

edge base of the system, further supporting the in-

sights gained from the PCA analysis.

4.2 RQ

: Gait Classiﬁcation

In the Lower Body scenario, PCA revealed that 113

features explained 98.4% of the variance (see Fig-

ure 8). Of these, 76 were from the rotation domain,

18 from the step domain, 18 from trajectories, and 1

from the GEI/SEI domain. This ﬁnding underscores

the relevance of all feature domains in the context of

gait classiﬁcation, particularly highlighting the rota-

tion domain, which, although rarely explored in the

literature, proves to be especially important in this

context.

Figure 8: Gait Classiﬁcation (Lower body): Variance Anal-

ysis.

A similar result was observed in the Full Body sce-

nario (see Figure 9), where PCA identiﬁed 174 fea-

tures that explained 98.8% of the variance. In con-

trast to the Lower Body scenario, no GEI/SEI features

were included in this set. Speciﬁcally, the 174 fea-

tures came from the rotation domain (124), the tra-

jectory domain (26), and the step domain (21). This

aligns with the ﬁndings from the Parkinson’s predic-

tion analysis, suggesting that GEI/SEI features may

be particularly useful when only lower body motion

data is available, while the rotation domain consis-

tently plays a crucial role in gait classiﬁcation.

HEALTHINF 2025 - 18th International Conference on Health Informatics

136

Table 3: Confusion matrix for the EDAM gait classiﬁcation model in the Lower Body scenario.

Actual Antalgic Lurch Normal Steppage Stiff-Legged Trendelenburg

Antalgic 2,122 16 32 38 62 108

Lurch 0 2,308 0 2 0 20

Normal 4 2 2,346 0 0 40

Steppage 44 28 8 2,246 2 42

Stiff-Legged 84 2 36 0 2,208 42

Trendelenburg 206 98 386 20 14 1,636

Table 4: Results obtained from the gait classiﬁcation model

in the Lower body scenario.

Class Precision Recall F1-score

Antalgic 0.86 0.89 0.88

Lurch 0.94 0.99 0.96

Normal 0.84 0.98 0.90

Steppage 0.97 0.95 0.96

Stiff-Legged 0.97 0.93 0.95

Trendelenburg 0.87 0.69 0.77

Global Accuracy 0.91

Figure 9: Gait Classiﬁcation (Full body): Variance Analy-

sis.

For gait classiﬁcation, the results in the Lower

Body scenario demonstrated that the best performance

was achieved using a machine learning pipeline based

on a linear Support Vector Machine (SVM) classiﬁer.

This pipeline did not use synthetic oversampling or

feature correlation analysis, but included automatic

feature selection via Logistic Regression. The model

was built using 124 features: 12 from the rotation do-

main, 94 from the trajectory domain, 15 from the step

domain, and 3 from the GEI/SEI domain. This out-

come aligns with the ﬁndings of the PCA analysis, re-

inforcing the critical importance of trajectory features

in the gait classiﬁcation process for the lower body

scenario. The model achieved an accuracy of 91%, in-

dicating strong performance in classifying gait types.

Table 3 presents the confusion matrix for the

gait classiﬁcation model in the Lower Body scenario,

while Table 4 reports the results obtained in terms of

precision, recall and F1-score for each class.

The analysis of the results shows that both the

Steppage and Stiff-legged gait classes, which are asso-

ciated with early neurological symptoms, performed

very well. The Steppage class achieved metrics ex-

ceeding 95%, while the Stiff-legged class, often linked

to early signs of Parkinson’s disease, reached an ac-

curacy of 97% and a recall of 93%. These ﬁndings

underscore the potential of EDAM in the early detec-

tion of Parkinson’s disease, demonstrating its ability

to effectively identify key gait deviations associated

with the onset of the condition.

In the Full Body scenario, similar to the Lower

Body scenario, the best performance was achieved us-

ing a linear Support Vector Machine (SVM) classi-

ﬁcation algorithm. This pipeline also excluded syn-

thetic oversampling and feature correlation analysis,

but employed an automatic feature selection algo-

rithm based on Extra Trees. The model was con-

structed with 250 features, including 131 from the ro-

tation domain, 63 from trajectories, 5 from the step

domain, and 6 from the GEI/SEI domain. Unlike the

Lower Body scenario, the most inﬂuential features for

gait classiﬁcation in this case were those from the ro-

tation domain, aligning with the ﬁndings from Parkin-

son’s disease prediction.

Regarding the accuracy metrics, Table 5 and Ta-

ble 6 present the precision, recall, and F1-score val-

ues for each class. In this scenario, the highest ac-

curacy was achieved for the Lurch class. However,

the Steppage and Stiff-legged classes continued to per-

form well, with recall values of 94% and 95%, re-

spectively, and precision values of 96% and 97%, re-

spectively. These results reinforce the robustness of

EDAM in classifying early symptoms of neurological

conditions such as Parkinson’s disease.

Answer to RQ

. The gait classiﬁcation models in

the Lower Body scenario achieved 91% accuracy,

Integrating Gait and Clinical Data with Explainable Artiﬁcial Intelligence for Parkinson’s Prediction: The EDAM System

137

Table 5: Confusion matrix of the EDAM gait classiﬁcation model in the Full body scenario.

Actual Antalgic Lurch Normal Steppage Stiff-Legged Trendelenburg

Antalgic 2,042 4 48 102 12 170

Lurch 8 2,262 2 12 28 18

Normal 48 0 2,312 10 0 22

Steppage 52 22 10 2,278 2 6

Stiff-Legged 30 6 10 8 2,308 16

Trendelenburg 134 18 186 2 10 2,010

Table 6: Results obtained from the gait classiﬁcation model

in the Full body scenario.

Class Precision Recall F1-score

Antalgic 0.88 0.86 0.87

Lurch 0.98 0.97 0.97

Normal 0.90 0.97 0.93

Steppage 0.94 0.96 0.95

Stiff-Legged 0.95 0.97 0.97

Trendelenburg 0.90 0.85 0.87

Global Accuracy 0.93

with trajectory features playing a key role, especially

for the Steppage class, which is associated with early

neurological symptoms. Similarly, in the Full Body

scenario, the models reached 93% accuracy, where

rotation features proved most inﬂuential. Notably,

in both scenarios, the Stiff-Legged class, which is

closely linked to the early signs of Parkinson’s dis-

ease, showed strong performance, achieving high pre-

cision and recall. These results underscore the po-

tential of EDAM to effectively identify gait patterns

that could be early indicators of Parkinson’s disease,

demonstrating its capability for early diagnosis.

5 RELATED WORK

Numerous studies in the literature have focused on

evaluating the effectiveness of machine learning (ML)

techniques in monitoring and predicting speciﬁc dis-

eases based on gait and posture data acquired through

motion analysis systems (e.g., (Abdulhay et al., 2018;

Costa et al., 2016; Cuzzolin et al., 2017; Man-

nini et al., 2016; Raknim and Lan, 2016)). Daliri,

2012 proposed a system that combines time series

data from foot signals with Support Vector Machine

(SVM) algorithms to predict diseases, including ALS.

Additionally, Ajay et al., 2018 introduced a system

for analyzing and classifying parkinsonian gait us-

ing videos captured by pervasive devices (e.g., smart-

phones, webcams, and surveillance cameras) through

a skeleton extraction model that directly detects joint

information from video frames. Vilas-Boas et al.,

2021 evaluated the use of ML techniques to build a

model capable of identifying the Val30Met mutation

based on gait characteristics. The study utilized the

Kinect v2 sensor to capture 24 gait parameters while

individuals walked toward the camera. Multiple ma-

chine learning algorithms were tested, including k-

nearest neighbors (KNN), decision trees, random for-

est, SVM, and Multilayer Perceptron. The authors

constructed a model with an average accuracy of 92%

in distinguishing healthy individuals from mutation

carriers (with or without symptoms), and 98% ac-

curacy in distinguishing between asymptomatic and

symptomatic carriers (both using SVM). Zhang and

Ma, 2019 investigated the application of supervised

machine learning algorithms in classifying sagittal

gait patterns in children with spastic diplegia. Gait

parameters were extracted from data obtained from

200 children, and the results demonstrated that an ar-

tiﬁcial neural network (ANN) achieved an accuracy

of 93.5%, proving to be a promising tool for auto-

matic interpretation of gait data. The literature high-

lights the beneﬁts of gait analysis in assessing mo-

tor deﬁcits, as gait is a fundamental, physiological,

and unforced form of locomotion with direct clini-

cal relevance. However, current systems that focus

on videographic gait analysis often produce variable

and non-repeatable results (Abada et al., 2013; Guil-

lot et al., 2008; Hampton and Amende, 2009; Mead

et al., 2011; Vinsant et al., 2013; Wooley et al.,

2005). This variability is not surprising, as many

key changes in limb positioning and movement dy-

namics are only visible from the lateral plane. Con-

sequently, recent efforts have focused on developing

systems that incorporate machine learning algorithms

(e.g., NeuroCube) and lateral view analysis (e.g., Mo-

toRater, Locomouse (Machado et al., 2015)) to ana-

lyze gait more comprehensively (Alexandrov et al.,

2015; Bellardita and Kiehn, 2015; de Bruin et al.,

2016; Talpalar et al., 2013). However, the full poten-

tial of lateral plane videography has yet to be realized,

as current analyses are often limited to a few func-

tional aspects and a small number of gait parameters

HEALTHINF 2025 - 18th International Conference on Health Informatics

138

(Preisig et al., 2016). A multi-camera system could

potentially provide better results by enabling detailed

joint analysis from different angles.

6 CONCLUSION AND FUTURE

WORK

In this paper, we addressed the challenges of gait

and posture analysis for the early diagnosis of neu-

rological disorders, particularly Parkinson’s disease,

through machine learning (ML)-based approaches.

Traditional methods often lack integration with clin-

ical biomarkers and fail to provide transparent, ex-

plainable predictions, limiting their clinical utility. To

address these limitations, we introduced EDAM (Ex-

plainable Diagnosis Recommender), a decision sup-

port system that integrates posture and gait data with

clinical biomarkers using Explainable AI (XAI) tech-

niques.

EDAM not only predicts the likelihood of neu-

rological disorders like Parkinson’s disease but also

explains its diagnostic decisions through visual and

natural-language outputs. This combination enhances

the trustworthiness and usability of predictions, sup-

porting specialists in making data-informed decisions

that incorporate their intuition, judgment, and experi-

ence. Furthermore, EDAM generates pre-reports that

assist clinicians in both qualitative and quantitative

evaluations of patient conditions.

To validate the effectiveness of EDAM, we con-

structed one of the largest dataset known in the lit-

erature by merging several established gait datasets.

This dataset contains 557 features and 7,303 labelled

instances, making it the most comprehensive resource

for evaluating machine learning models in the con-

text of Parkinson’s prediction and gait classiﬁcation.

EDAM achieved high accuracy in distinguishing be-

tween healthy individuals and those with Parkinson’s

disease, as well as in classifying abnormal gait pat-

terns linked to early-stage neurological disorders.

Future works will focus on expanding the range

of pathologies covered by the system, improving

model generalization across diverse populations, and

further reﬁning the interpretability of its predictions

through advanced XAI techniques. We also plan to

conduct studies to assess the contribution of addi-

tional biomarkers in identifying neurological disor-

ders when combined with posture data. Furthermore,

we intend to perform experiments with specialists to

evaluate the acceptability of the predictions of EDAM

and the clarity of their explanations.

ACKNOWLEDGEMENT

This research was supported by Italian General In-

spectorate of Military Medical Services (IGESAN)

under Military Health Research Plan - Project n.

9908115328 (CIG).

REFERENCES

Abada, Y.-s. K., Nguyen, H. P., Schreiber, R., and Ellen-

broek, B. (2013). Assessment of motor function, sen-

sory motor gating and recognition memory in a novel

bachd transgenic rat model for huntington disease.

PloS one, 8(7):e68584.

Abdulhay, E., Arunkumar, N., Narasimhan, K., Vellaiap-

pan, E., and Venkatraman, V. (2018). Gait and tremor

investigation using machine learning techniques for

the diagnosis of parkinson disease. Future Genera-

tion Computer Systems, 83:366–373.

Ajay, J., Song, C., Wang, A., Langan, J., Li, Z., and Xu,

W. (2018). A pervasive and sensor-free deep learn-

ing system for parkinsonian gait analysis. In 2018

IEEE EMBS International Conference on Biomedical

& Health Informatics (BHI), pages 108–111. IEEE.

Alexandrov, V., Brunner, D., Hanania, T., and Leahy, E.

(2015). High-throughput analysis of behavior for

drug discovery. European journal of pharmacology,

750:82–89.

Amboni, M. et al. (2021). Gait analysis may distinguish

progressive supranuclear palsy and parkinson disease

since the earliest stages. Scientiﬁc Reports, 11(1):1–9.

Balletti, N., Zinni, R., Russodivito, M., Laudato, G., Scal-

abrino, S., and Oliveto, R. (2024). A machine learn-

ing model for the qualitative assessment of human gait

based on video features. In Proceedings of the 17th In-

ternational Joint Conference on Biomedical Engineer-

ing Systems and Technologies – HEALTHINF, Rome,

Italy.

Bellardita, C. and Kiehn, O. (2015). Phenotypic charac-

terization of speed-associated gait changes in mice

reveals modular organization of locomotor networks.

Current Biology, 25(11):1426–1436.

Buckley, C., Alcock, L., McArdle, R., Rehman, R. Z. U.,

Del Din, S., Mazz

a, C., Yarnall, A. J., and Rochester,

L. (2019). The role of movement analysis in diagnos-

ing and monitoring neurodegenerative conditions: In-

sights from gait and postural control. Brain sciences,

9(2):34.

Costa, L., Gago, M. F., Yelshyna, D., Ferreira, J.,

David Silva, H., Rocha, L., Sousa, N., and Bi-

cho, E. (2016). Application of machine learning

in postural control kinematics for the diagnosis of

alzheimer’s disease. Computational intelligence and

neuroscience, 2016(1):3891253.

Cuzzolin, F., Sapienza, M., Esser, P., Saha, S., Franssen,

M. M., Collett, J., and Dawes, H. (2017). Metric learn-

ing for parkinsonian identiﬁcation from imu gait mea-

surements. Gait & posture, 54:127–132.

Integrating Gait and Clinical Data with Explainable Artiﬁcial Intelligence for Parkinson’s Prediction: The EDAM System

139

Daliri, M. R. (2012). Automatic diagnosis of neuro-

degenerative diseases using gait dynamics. Measure-

ment, 45(7):1729–1734.

de Bruin, N., Schmitz, K., Schiffmann, S., Tafferner, N.,

Schmidt, M., Jordan, H., H

außler, A., Tegeder, I.,

Geisslinger, G., and Parnham, M. (2016). Multiple

rodent models and behavioral measures reveal unex-

pected responses to fty720 and dmf in experimental

autoimmune encephalomyelitis. Behavioural brain

research, 300:160–174.

Edwards, L. and Veale, M. (2017). Slave to the algorithm?

why a’right to an explanation’is probably not the rem-

edy you are looking for. Duke L. & Tech. Rev., 16:18.

Guillot, T. S., Asress, S. A., Richardson, J. R., Glass, J. D.,

and Miller, G. W. (2008). Treadmill gait analysis does

not detect motor deﬁcits in animal models of parkin-

son’s disease or amyotrophic lateral sclerosis. Journal

of motor behavior, 40(6):568–577.

Guyon, I. and Elisseeff, A. (2003). An introduction to vari-

able and feature selection. Journal of machine learn-

ing research, 3:1157–1182.

Hampton, T. G. and Amende, I. (2009). Treadmill gait anal-

ysis characterizes gait alterations in parkinson’s dis-

ease and amyotrophic lateral sclerosis mouse models.

Journal of motor behavior, 42(1):1–4.

Han, J. and Bhanu, B. (2005). Individual recognition using

gait energy image. In 2005 IEEE Computer Society

Conference on Computer Vision and Pattern Recogni-

tion (CVPR’05), volume 2, pages i–i. IEEE.

Hastie, T., Tibshirani, R., and Friedman, J. (2009). The Ele-

ments of Statistical Learning: Data Mining, Inference,

and Prediction. Springer Science & Business Media.

Holzinger, A., Biemann, C., Pattichis, C. S., and Kell, D. B.

(2017a). What do we need to build explainable ai

systems for the medical domain? arXiv preprint

arXiv:1712.09923.

Holzinger, A., Malle, B., Kieseberg, P., Roth, P. M.,

uller, H., Reihs, R., and Zatloukal, K. (2017b).

Towards the augmented pathologist: Challenges of

explainable-ai in digital pathology. arXiv preprint

arXiv:1712.06657.

Jun, K., Lee, Y., Lee, S., Lee, D. W., and Kim, M. S.

(2020). Pathological gait classiﬁcation using kinect

v2 and gated recurrent neural networks. IEEE Access,

8:139881–139891.

Kour, N., Gupta, S., and Arora, S. (2020). Gait dataset

for knee osteoarthritis and parkinson’s disease anal-

ysis with severity levels.

Li, J., Cheng, K., Wang, S., Morstatter, F., Trevino, R. P.,

Tang, J., and Liu, H. (2017). Feature selection: A

data perspective. ACM Computing Surveys (CSUR),

50(6):1–45.

Lundberg, S. M. and Lee, S.-I. (2017). A uniﬁed approach

to interpreting model predictions. In Proceedings of

the 31st International Conference on Neural Infor-

mation Processing Systems (NIPS), pages 4765–4774.

Curran Associates Inc.

Machado, A. S., Darmohray, D. M., Fayad, J., Marques,

H. G., and Carey, M. R. (2015). A quantitative frame-

work for whole-body coordination reveals speciﬁc

deﬁcits in freely walking ataxic mice. elife, 4:e07892.

Mannini, A., Trojaniello, D., Cereatti, A., and Sabatini,

A. M. (2016). A machine learning framework for gait

classiﬁcation using inertial sensors: Application to el-

derly, post-stroke and huntington’s disease patients.

Sensors, 16(1):134.

Mead, R. J., Bennett, E. J., Kennerley, A. J., Sharp, P.,

Sunyach, C., Kasher, P., Berwick, J., Pettmann, B.,

Battaglia, G., Azzouz, M., et al. (2011). Optimised

and rapid pre-clinical screening in the sod1g93a trans-

genic mouse model of amyotrophic lateral sclerosis

(als). PloS one, 6(8):e23244.

Mehrizi, R., Peng, X., Zhang, S., Liao, R., and Li, K.

(2019). Automatic health problem detection from gait

videos using deep neural networks. arXiv preprint

arXiv:1906.01480.

Preisig, D. F., Kulic, L., Kr

uger, M., Wirth, F., McAfoose,

J., Sp

ani, C., Gantenbein, P., Derungs, R., Nitsch,

R. M., and Welt, T. (2016). High-speed video gait

analysis reveals early and characteristic locomotor

phenotypes in mouse models of neurodegenerative

movement disorders. Behavioural brain research,

311:340–353.

Raknim, P. and Lan, K.-c. (2016). Gait monitoring for

early neurological disorder detection using sensors in

a smartphone: Validation and a case study of parkin-

sonism. Telemedicine and e-Health, 22(1):75–81.

Schreiber, C. and Moissenet, F. (2019). A multimodal

dataset of human gait at different walking speeds es-

tablished on injury-free adult participants. Scientiﬁc

Data, 6(1):1–7.

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

dynamic gait index to quantify gait ability in patients

with vestibular and balance disorders. Physical ther-

apy, 75(6):538–548.

Talpalar, A. E., Bouvier, J., Borgius, L., Fortin, G., Pierani,

A., and Kiehn, O. (2013). Dual-mode operation of

neuronal networks involved in left–right alternation.

Nature, 500(7460):85–88.

Vilas-Boas, M. D. C., Rocha, A. P., Cardoso, M. N., Fer-

nandes, J. M., Coelho, T., and Cunha, J. P. S. (2021).

Supporting the assessment of hereditary transthyretin

amyloidosis patients based on 3-d gait analysis and

machine learning. IEEE Transactions on Neural Sys-

tems and Rehabilitation Engineering, 29:1350–1362.

Vinsant, S., Mansﬁeld, C., Jimenez-Moreno, R., Moore,

V. D. G., Yoshikawa, M., Hampton, T. G., Prevette,

D., Caress, J., Oppenheim, R. W., and Milligan, C.

(2013). Characterization of early pathogenesis in the

sod1 g93a mouse model of als: part i, background and

methods. Brain and behavior, 3(4):335–350.

Wold, S., Esbensen, K., and Geladi, P. (1987). Principal

component analysis. Chemometrics and intelligent

laboratory systems, 2(1-3):37–52.

Wooley, C. M., Sher, R. B., Kale, A., Frankel, W. N., Cox,

G. A., and Seburn, K. L. (2005). Gait analysis detects

early changes in transgenic sod1 (g93a) mice. Muscle

& Nerve: Ofﬁcial Journal of the American Associa-

tion of Electrodiagnostic Medicine, 32(1):43–50.

Zhang, Y. and Ma, Y. (2019). Application of supervised

machine learning algorithms in the classiﬁcation of

sagittal gait patterns of cerebral palsy children with

spastic diplegia. Computers in biology and medicine,

106:33–39.

HEALTHINF 2025 - 18th International Conference on Health Informatics

140