Personalized Asthma Recommendation System: Leveraging Predictive

Analysis and Semantic Ontology-Based Knowledge Graph

Ayan Chatterjee

Dept. of Digital Technology, STIFTELSEN NILU, 2007 Kjeller, Norway

Keywords:

Predictive Analysis, Personalized Recommendation System, Asthma Monitoring, Semantic Knowledge.

Abstract:

Personalized approaches are required for asthma management due to the variability in symptoms, triggers,

and patient characteristics. An innovative asthma recommendation system that integrates automatic predic-

tive analysis with semantic knowledge to provide personalized recommendations for asthma management is

proposed by this paper. Asthma exacerbation are predicted and the recommendations are enhanced by the

system, which leverages automatic Tree-based Pipeline Optimization Tool (TPOT) and semantic knowledge

represented in an OWL Ontology (AsthmaOnto). Furthermore, classiﬁcations are explained with Local Inter-

pretable Model-Agnostic Explanations (LIME) to identify feature importance. Tailored interventions based on

individual patient proﬁles are provided by this conceptual model, aiming to improve asthma management. The

proposed model has been veriﬁed using public asthma datasets, and a public weather air-quality dataset has

been utilized to support ontology development and veriﬁcation. In TPOT, the Gaussian Naive Bayes (Gaus-

sianNB) classiﬁer has outperformed other supervised machine learning models with an accuracy of 0.75, for

the used dataset. To implement and evaluate the proposed model in clinical settings, further development and

validation with more diverse and robust datasets with model calibration are required.

1 INTRODUCTION

Asthma is a signiﬁcant noncommunicable disease af-

fecting individuals of all ages, particularly children

(WHO, 2019). It is characterized by inﬂammation

and constriction of the lung’s small airways, leading

to coughing, wheezing, breathlessness, and chest con-

striction (WHO, 2019). In 2019, asthma affected an

estimated 262 million people, resulting in 455,000

deaths (WHO, 2019). Avoiding triggers is crucial

for symptom reduction. Most asthma-related deaths

occur in low- to lower-middle-income countries due

to under-diagnosis and under-treatment. Asthma is

classiﬁed as allergic or non-allergic, triggered by fac-

tors like allergens, pollution, weather, tobacco smoke,

and food allergens (Ajami et al., 2022). Symptoms

vary in frequency and severity based on individual

reactions to these triggers (Ajami et al., 2022). The

WHO is committed to improving asthma diagnosis,

treatment, and monitoring to reduce the global bur-

den of noncommunicable diseases and promote uni-

versal health coverage (WHO, 2019). Proper lifestyle

management and personalized recommendations can

https://orcid.org/0000-0002-1051-2814

help individuals maintain a normal, active life. Ef-

fective asthma management involves identifying trig-

gers, controlling symptoms, and preventing exacer-

bation. However, the diverse nature of asthma re-

quires tailored approaches, making its management

complex. Traditional monitoring often relies on over-

simpliﬁed systems that do not capture the condi-

tion’s complexity. Conventional methods classify pa-

tients as controlled or uncontrolled based on pre-

determined benchmarks. Personalized asthma man-

agement now includes patient-centered care and tai-

lored self-management support, requiring changes in

practice organization, consultation modes, and digi-

tal health options (Pinnock et al., 2023). Anantharam

et al. (Anantharam et al., 2015) developed kHealth,

a system that aggregates multisensory data from sen-

sors and questionnaires from asthma patients to help

doctors more precisely determine the cause, severity,

and control level of asthma, and to alert patients to

seek timely clinical assistance. Alharbi et al. (Al-

harbi et al., 2023) proposed a multi-layered eHealth

framework using telemonitoring, environmental sen-

sors, and machine learning to predict asthma attacks

and provide safe routes, updating recommendations

with real-time data. Another study by Alharbi et al.

Chatterjee, A.

Personalized Asthma Recommendation System: Leveraging Predictive Analysis and Semantic Ontology-Based Knowledge Graph.

DOI: 10.5220/0012894500003838

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 2: KEOD, pages 127-134

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

127

(Alharbi et al., 2021) presented a smart healthcare

framework that visualized asthma triggers and pro-

vided attack notiﬁcations, using deep learning to an-

alyze patient and environmental data, including a dy-

namic air pollution map and safe-route recommenda-

tions. Morita et al. (Morita et al., 2019) designed

Breathe, a web-based mHealth platform for manag-

ing asthma, which showed initial strong utilization but

later declined, with users reporting high satisfaction

and conﬁdence in its assessments. Bose et al. (Bose

et al., 2021) developed machine learning models to

predict persistent asthma in children, with the XG-

Boost model performing best, although direct com-

parisons were difﬁcult due to a lack of similar stud-

ies. Kadariya et al. (Kadariya et al., 2019) introduced

kBot, a personalized chatbot for pediatric asthma pa-

tients, monitoring medication adherence and tracking

health data, which showed high acceptance and us-

ability in preliminary evaluations.

Traditional asthma management often relies on

standardized guidelines, which, while informative,

may not fully account for individual variations in

symptoms, triggers, and treatment responses. Pre-

dicting and explaining asthma exacerbations remains

challenging due to factors like environmental trig-

gers, patient demographics, and medical history. This

research introduces a novel Asthma Recommenda-

tion System designed for personalized asthma mon-

itoring and evidence-based recommendations. The

system integrates an automatic prediction pipeline

with a semantic knowledge graph, leveraging pre-

dictive analysis and semantic reasoning to enhance

decision-making accuracy. By using OWL-based on-

tology and SPARQL querying, the system captures

complex data patterns, enriching personalized recom-

mendations. This approach aims to surpass tradi-

tional methods by offering more tailored and action-

able insights for optimal asthma management (Chat-

terjee et al., 2021b)(Chatterjee and Prinz, 2022)(Cima

et al., 2017)(Chatterjee et al., 2023)(Chatterjee et al.,

2022b). The research questions guiding this study are

as follows: a. How does integrating automatic pre-

dictive analysis with a semantic knowledge graph en-

hance the development of a personalized, evidence-

based asthma recommendation system? b. How

can the system’s predictive analysis be effectively ex-

plained? c. How can recommendations be system-

atically generated and clariﬁed within the proposed

asthma recommendation system?

This research aims to enhance the theoretical

understanding of hybrid recommendation generation

(automatic predictive analysis combined with seman-

tic rule-based methods) and to foster practical ap-

plications for proactive asthma management. The

study serves as a technical proof-of-concept, focus-

ing on the conceptual modeling of a recommenda-

tion system for asthma, validated with publicly avail-

able datasets. Future research will emphasize tech-

nical validation using real-world health monitoring

data. The paper is organized as follows: Section 2 dis-

cusses the proposed system architecture and modeling

approaches. Section 3 details the datasets used for

predictive model evaluation and semantic modeling

for personalized recommendations. Section 4 covers

the experimental setup, outcomes, and research ques-

tions. Finally, Section 5 presents conclusions and fu-

ture directions.

2 PROPOSED WORK

The design and development of the proposed asthma

recommendation system involve following aspects.

2.1 System Architecture

The proposed system architecture comprises multiple

layers distributed across system components, includ-

ing the Front-end User Interface (UI) and the Back-

end server with a database (see Fig. 1). These compo-

nents communicate via REST (Representational State

Transfer) architecture (Barbaglia et al., 2017), ensur-

ing a scalable, secure, and user-friendly framework,

using JSON (JavaScript Object Notation) (Barbaglia

et al., 2017) for data interchange. The system is

composed of the following layers: Data Acquisi-

tion Layer: Collects asthma data and demographics

from electronic health records, surveys, and monitor-

ing devices. Features dashboards, tools, and widgets

for recording data, viewing recommendations, and

monitoring progress. Data Integration Layer: Inte-

grates with healthcare IT systems like EHR, CDSS,

and telemedicine platforms, focusing on scalability

and security, with FHIR and HL7 standards support

(Chatterjee et al., 2022a). Data Processing Layer:

Applies preprocessing techniques to clean, transform,

and standardize data for predictive models. Predic-

tive Analysis Layer: Uses TPOT (Olson and Moore,

2016), an AutoML tool, to predict asthma exacerba-

tions and optimize machine learning pipelines with

scikit-learn compatibility. Knowledge Representation

Layer: Develops an asthma ontology (AsthmaOnto)

for domain-speciﬁc knowledge, including symptoms

and treatments. Recommendation Generation Layer:

Combines predictive results with ontology knowledge

to generate personalized asthma management recom-

mendations, mapped to the AsthmaReco ontology.

Automatic Machine Learning Pipeline: TPOT auto-

KEOD 2024 - 16th International Conference on Knowledge Engineering and Ontology Development

128

Figure 1: The architecture of the proposed recommendation system.

mates data preprocessing and pipeline construction,

reducing manual effort and time.

2.2 Predictive Analysis

TPOT automates the discovery of optimal data pre-

processing and machine learning model combina-

tions, improving predictive accuracy and robustness.

Using genetic programming, TPOT efﬁciently evalu-

ates thousands of potential pipelines to identify the

best model for a dataset, including hyperparame-

ter optimization. The process includes generating

random pipelines, evaluating performance, selecting

top performers, and applying crossover and mutation

across generations to reﬁne results, ultimately select-

ing the best pipeline after a set number of generations.

The TPOT conﬁguration customizes feature selec-

tion methods and multiple sklearn classiﬁers, includ-

ing RandomForest, Logistic Regression, SVC, Deci-

sionTree, KNN, and Naive Bayes (Olson and Moore,

2016). The TPOTClassiﬁer is conﬁgured with 100

generations and a population size of 100, allowing

extensive exploration of the feature space. A high

mutation rate of 0.9 and a crossover rate of 0.1 pri-

oritize new solutions over recombination. Models are

optimized for accuracy using 5-fold cross-validation.

TPOT uses following performance metrics to evalu-

ate and select the best pipelines − Accuracy, Preci-

sion, Recall, F1-score, and Matthews Correlation Co-

efﬁcient (MCC) (Chatterjee et al., 2023)(Chatterjee

et al., 2020)(Chatterjee et al., 2021a). Stratiﬁcation

technique, using a Stratiﬁed Split (80% training, 20%

test), ensures that class proportions in training and test

sets match the original dataset (Hutchinson, 1982).

This approach reduces bias and enhances reliability in

training and evaluation by maintaining consistent per-

formance metrics. While TPOT does not directly pro-

vide model explanations, the resulting pipelines has

been analyzed with tools like LIME. LIME approxi-

mates the model locally around speciﬁc predictions,

revealing which features signiﬁcantly inﬂuenced the

decision (Garreau and Luxburg, 2020). This approach

enhances transparency and offers actionable insights

to improve patient care and personalized recommen-

dations (Garreau and Luxburg, 2020).

2.3 AsthmaOnto OWL Ontology Model

While traditional databases are effective for struc-

tured data storage and retrieval, they fall short in se-

mantic reasoning and integrating diverse data sources.

Ontologies offer a more dynamic and interconnected

approach, enhancing personalized, context-aware rec-

ommendations. This paper presents an OWL on-

tology for asthma monitoring, encompassing domain

knowledge through classes, properties, relationships,

individuals, logical operators (AND, OR, NOT), in-

ference rules, and axioms using set theory or ﬁrst-

order predicate logic. The ontology can integrate

multiple relevant ontologies, including Asthma from

BioPortal, weather and environment from COPDol-

ogy, food allergens from FoodOn, and symptoms and

pollen concepts from SNOMED-CT (Ajami et al.,

2022).

Let G = (V, E) be a directed graph where V repre-

sents the set of all concepts and E represents the set of

all relationships (Chatterjee et al., 2021b)(Chatterjee

et al., 2023)(Chatterjee et al., 2022b).

V = {Symptom, Demographic, MedicalHistory, ...}

E = {(Patient, hasSymptom, Symptom), .....)}

SPARQL is a query language for handling RDF

data, representing web information in a graph form

(Chatterjee and Prinz, 2022). It is vital for extracting

Personalized Asthma Recommendation System: Leveraging Predictive Analysis and Semantic Ontology-Based Knowledge Graph

129

and using data from ontologies, particularly in retriev-

ing and managing information from ontology-based

knowledge graphs in asthma recommendations (Cima

et al., 2017). A basic graph pattern in SPARQL can

be represented as:

{?sP?o}

where ?s is a subject variable, P is a predicate, and

?o is an object variable.

A triple pattern is an RDF triple where each of the

subject, predicate, and object may be a variable.

Let T represent the set of all triple patterns:

T = {(?s, P,?o) |?s, ?o ∈ I ∪V ∪ V and P ∈ R}

where V is the set of SPARQL variables. The fur-

ther terminologies are depicted in Fig. 2.

2.4 Personalized Recommendation

Let X be the feature matrix and y be the label vector.

Train a model f : X → y, such as a logistic regres-

sion or random forest. For a new patient x, predict the

probability of an asthma exacerbation: ˆy = f (x). If

ˆy > θ (a chosen threshold), generate and recommend

preventive measures.

The proposed personalized recommendation gen-

eration algorithm (Algo. 1) leverages predictive

analysis and semantic knowledge to generate tai-

lored asthma management recommendations. It starts

by cleaning and normalizing raw asthma data, fol-

lowed by training a machine learning model to predict

asthma exacerbation probabilities. The system loads

an AsthmaOnto ontology, which provides domain-

speciﬁc semantic knowledge, to enhance the recom-

mendations. For each patient, the system predicts ex-

acerbation likelihood and queries the ontology based

on patient attributes, generating personalized asthma

management recommendations. The goal is to trigger

logical rules of the form (A IMPLIES B) or (NOT(A)

OR B), providing tailored advice when speciﬁc vari-

ables are inferred as true.

Time Complexity − Data Preprocessing: O(n ·

m), where n is the number of records and m is the

number of features. Predictive Analysis: O(t) to

O(t · n · m), with t as training time. Semantic Knowl-

edge: Ontology loading has O(1). Recommenda-

tions: O(k · q), where k is the number of patients and

q is query complexity. Overall: O(n · m + t + k · q).

Space Complexity − Data Preprocessing: O(s) for

data. Predictive Analysis: O(p) for the model. Se-

mantic Knowledge: O(o) for the ontology. Rec-

ommendations: O(r) for storing results. Overall:

O(s + p + o + r). The time and space complexity of

Algo. 1 depends on the dataset size, model complex-

ity, ontology size, and query efﬁciency.

Algorithm 1: Personalized Asthma Recommendation

Generation.

Require: • Asthma patient data: D = {(x

)}

i=1

• Asthma ontology

Ensure: Personalized recommendations for each pa-

tient

1: Preprocess data: Clean, engineer features, and

normalize D to obtain D

preprocessed

2: Train predictive model on D

preprocessed

to obtain

model M

3: Load asthma ontology as O

4: for each patient x

in D

preprocessed

5: Predict exacerbation probability for x

using M

6: Query O for personalized recommendations

based on x

7: Combine prediction and ontology knowledge

to generate recommendations for x

8: end for

9: return List of personalized recommendations for

all patients

Potential recommendation messages for personal-

ized asthma management can be represented using the

AsthmaOnto ontology and their potential types are

mentioned in Table 1.

3 DATA DESCRIPTION

The datasets have been used for predictive analysis

and semantic modeling. The proposed hypothesis

suggests that amalgamating these datasets could cre-

ate an asthma recommendation system for ongoing

surveillance of health metrics and weather patterns.

This system could anticipate potential asthma trig-

gers, generating timely alerts for patients during el-

evated pollen counts, drastic weather changes, or de-

viations in vital signs. The “Asthma Disease Predic-

tion” dataset (Dataset, 2024) on Kaggle offers key

data for predicting asthma, including demographic

details (age, gender), clinical symptoms (tiredness,

dry cough, breathing difﬁculties, sore throat, nasal

congestion, runny nose), and medical history (treat-

ment in different medical units, speciﬁc medical unit,

pneumonia history). In addition to features from the

used dataset, weather data from the OpenWeatherMap

API (OpenWeather, 2024) is utilized for semantic

modeling. The API provides essential weather-related

data relevant to asthma management, including tem-

perature (current, feels like, min/max), humidity, air

pressure, wind, precipitation, visibility, pollution lev-

els (Air Quality Index, PM2.5, PM10, ozone, nitrogen

dioxide, sulfur dioxide, carbon monoxide), and pollen

counts. These factors are critical for asthma manage-

KEOD 2024 - 16th International Conference on Knowledge Engineering and Ontology Development

130

The Terminological box (TBox) (Chatterjee et al., 2021b) contains the vocabulary of the ontology:

Patient(x) → ∀y (hasSymptom(x, y) → Symptom(y))

Patient(x) → ∀y (hasDemographic(x, y) → Demographic(y))

Patient(x) → ∀y (hasMedicalHistory(x, y) → MedicalHistory(y))

Patient(x) → ∀y (hasAllergyFactor(x, y) → ExternalAllergyFactor(y))

Patient(x) → ∀y (hasRecommendation(x, y) → Recommendation(y))

Symptom(x) → ∃y (Patient(y) ∧ hasSymptom(y, x))

Demographic(x) → ∃y (Patient(y) ∧ hasDemographic(y, x))

MedicalHistory(x) → ∃y (Patient(y) ∧ hasMedicalHistory(y, x))

ExternalAllergyFactor(x) → ∃y (Patient(y) ∧ hasAllergyFactor(y,x))

Recommendation(x) → ∃y (Patient(y) ∧ hasRecommendation(y, x))

The Assertional box (ABox) (Chatterjee et al., 2021b) contains assertions about “XYZ” individual:

Patient(XYZ), hasSymptom(XYZ,DryCough),hasDemographic(XYZ,Age 42), hasMedicalHistory(XYZ, Asthma)

Logical rules (Chatterjee et al., 2021b) to infer new knowledge as reasoning can be expressed as:

∀x(Patient(x) ∧ hasSymptom(x,y) → Symptom(y))

∀x(Patient(x) ∧ hasDemographic(x,y) → Demographic(y))

∀x(Patient(x) ∧ hasMedicalHistory(x,y) → MedicalHistory(y))

∀x(Patient(x) ∧ hasAllergyFactor(x, y) → ExternalAllergyFactor(y))

Figure 2: Ontology-Based Representation of Asthma Patients’ Data.

Table 1: Types of recommendations in Asthma management.

Category Key Actions

Preventive Measures Avoid triggers (pollen, dust, pets); use allergen-proof bedding, air puriﬁers,

and close windows during pollen seasons.

Medication Adherence Take prescribed medications consistently; keep rescue inhaler available.

Environmental Control Maintain humidity below 50%; clean home regularly; use allergen-proof cov-

ers.

Lifestyle Modiﬁcations Exercise regularly; eat a balanced diet; avoid tobacco smoke and pollutants.

Symptom Management Monitor symptoms; practice relaxation techniques.

Follow-Up Care Regular appointments; stay current with vaccinations.

ment, as extreme weather conditions, pollution, and

high pollen levels can trigger or worsen symptoms.

4 IMPLEMENTATION

The implementation has been divided into the follow-

ing subsections:

4.1 Experimental Setup

We used Python 3.9.15 with libraries like pandas (v.

1.5.2), NumPy (v. 1.22.4), SciPy (v. 1.7.3), Mat-

plotlib (v. 3.6.2), Seaborn (v. 0.12.0), and scikit-

learn (v. 1.1.3) for data processing and model de-

velopment. The environment was conﬁgured on Win-

dows 10 using Anaconda and Jupyter Notebook 6.5.2.

The system had 16 GB RAM and a 64-bit architec-

ture, running models on the CPU due to the dataset’s

small size. Additionally, we used OWL 2 Prot

e in

the Prot

e editor to design the ontology model, with

Hermit reasoner for consistency checking.

4.2 Predictive Analysis

The processed asthma dataset contained 3,16,800

records with 19 features, including 100,820 Class:0

and 33,606 Class:1 entries. Feature selection utilized

SelectKBest, which identiﬁes statistically signiﬁcant

features, and VarianceThreshold, which removes low-

variance features. Correlation analysis conﬁrmed that

features were not highly correlated. TPOT identiﬁed

the best pipeline, which exclusively used the Gaus-

sianNB classiﬁer, achieving an accuracy of 0.75 (Pre-

cision: 0.72, Recall: 0.75, F1-score: 0.75, MCC:

0.70). Despite its simplicity, GaussianNB proved

effective, likely due to the dataset’s alignment with

the model’s assumption of normally distributed fea-

tures. Its fewer parameters and simplicity helped pre-

vent overﬁtting, making it a robust choice for this

dataset. The advantages of GaussianNB stem from

its appropriate assumption of feature distribution, ef-

Personalized Asthma Recommendation System: Leveraging Predictive Analysis and Semantic Ontology-Based Knowledge Graph

131

(a) Top ten features to predict No asthma. (b) Top ten features to predict Asthma.

Figure 3: LIME explanations for predicting class: 0 (No asthma) and class: 1 (Asthma).

∀x, y(Patient(x)∧ hasDemographic(x, y) → Demographic(y))

∀x, y(Patient(x)∧ hasMedicalHistory(x, y) → MedicalHistory(y))

∀x, y(Patient(x)∧ hasAllergyFactor(x, y) → ExternalAllergyFactor(y))

∀x, y(Patient(x)∧ hasRecommendation(x, y) → Recommendation(y))

∀x(Symptom(x) → ∃y(Patient(y)∧ hasSymptom(y, x)))

∀x(Demographic(x) → ∃y(Patient(y)∧ hasDemographic(y, x)))

∀x(MedicalHistory(x) → ∃y(Patient(y)∧ hasMedicalHistory(y, x)))

∀x(ExternalAllergyFactor(x) → ∃y (Patient(y) ∧ hasAllergyFactor(y, x)))

∀x(Recommendation(x) → ∃y(Patient(y)∧ hasRecommendation(y, x)))

Figure 4: Logical Expressions Representing Relationships Between Patients and Associated Data.

fective feature selection, and its probabilistic nature,

which supports good generalization without overﬁt-

ting. This highlights the value of simple models like

GaussianNB when they align well with data charac-

teristics. GaussianNB was combined with LIME for

predictive explanation, enhancing model transparency

and trust. LIME’s visualizations use blue color to in-

dicate a push towards the negative class and orange

color towards the positive class. The MCC metric

provided a balanced evaluation of classiﬁer perfor-

mance on this imbalanced asthma dataset.

4.3 Semantic Recommendations

The key classes of OWL ontology are shown in Fig.

5 (Axiom: 143, Logical axiom: 82, Declaration ax-

iom: 61, Class: 10, object property: 9, data prop-

erty: 20, individual count: 23, SubClassOf: 2). Using

the Hermit reasoner in Prot

e, reasoning was com-

pleted in under 40 seconds with no inconsistencies.

When loaded in Jena (TTL format, OWL full), read-

ing time was 0.5-1.0 second, and queries for ontology

elements executed in under 1.0 second. Each ontol-

ogy model, as an RDF graph, links to a document

manager (default: OntDocumentManager”) for doc-

ument processing. In the ontology API, all ontology

classes inherit from OntResource,” sharing attributes

(versionInfo, comment, label, seeAlso, isDeﬁnedBy,

sameAs, differentFrom) and methods (add, set, list,

get, has, remove). These logical inferences, based on

the ontology, ensure data consistency and validity as

depicted in Fig. 4.

The SPARQL queries has been executed success-

fully to retrieve relevant health conditions from the

individual ontology instance to generate personal-

ized recommendations following the criteria: X ∈ R

be the feature vector representing an individual or

item. w ∈ R

be the weight vector of the predic-

tive model. ˆy ∈ R be the predicted score obtained

from the predictive model. θ ∈ R be the chosen

threshold for making recommendations. The predic-

tive score ˆy is calculated as: ˆy = f (X, w), where f

is the predictive function (e.g., a machine learning

model and here, GaussianNB). The decision rule for

making recommendations is given by: ˆy = f (X, w) >

θ =⇒ Recommend preventive measures. Here are

example of successfully executed SPARQL queries in

this purpose:

KEOD 2024 - 16th International Conference on Knowledge Engineering and Ontology Development

132

Figure 5: The structure of the proposed AsthmaOnto ontology in Prot

Query 1: Patients with speciﬁc symptoms,

age range, and gender

SELECT ?patient

WHERE {

?patient :hasSymptom :DryCough .

?patient :hasSymptom :DifﬁcultyInBreathing .

?patient :hasDemographic ?demographic .

?demographic :age ?age .

?demographic :gender “female” .

?patient :hasMedicalHistory ?medicalHistory .

?medicalHistory :pneumonia true .

?patient :hasSymptom :NasalCongestion .

FILTER (?age > 30 && ?age < 50)

}

Query 2: Patients with runny nose, asthma

prediction and symptom-based recommendations

SELECT ?patient ?recommendation

WHERE {

?patient :hasSymptom :RunnyNose .

?patient :hasDemographic ?demographic .

?demographic :gender “male” .

?patient :hasAsthmaPredicted ?asthmapredicted .

FILTER (?asthmapredicted = true)

?symptom rdf:type :SoreThroat .

?recommendation rdf:type :PreventiveMeasure .

?recommendation :relatedTo ?symptom .

}

For a 40-year-old asthma patient with symptoms

like dry cough and difﬁculty breathing, a history of

pneumonia, and exposure to high pollen levels, the

most relevant RecommendationType based on the

provided SPARQL example would be “Preventive

Measures”. This involves sending a Recommenda-

tionMessage that emphasizes proactive steps to min-

imize asthma triggers and improve indoor air quality,

helping the patient manage symptoms and reduce the

risk of exacerbations.

The paper integrates automatic predictive analy-

sis with a semantic knowledge graph to enhance per-

sonalized asthma monitoring, a novel approach ac-

cording to the existing literature. It preprocesses

a large asthma dataset using SelectKBest and Vari-

anceThreshold to build a robust GaussianNB model,

which proved highly accurate. To explain predic-

tions, GaussianNB is combined with LIME, offering

visual explanations that increase system transparency.

The paper also develops an OWL ontology and uses

SPARQL queries for personalized recommendations,

ensuring that both predictions and recommendations

are precise, personalized, and grounded in semantic

knowledge.

5 CONCLUSION

The proposed asthma recommendation system inte-

grates automatic predictive analysis with semantic

knowledge representation, offering personalized rec-

ommendations to improve asthma outcomes and pa-

tient care. TPOT automates machine learning pipeline

optimization, identifying the best model and hyper-

parameters for the asthma dataset, while LIME ex-

planations enhance the interpretability of the Gaus-

sianNB model. The use of predictive analysis and the

AsthmaOnto OWL ontology provides signiﬁcant ad-

vantages in efﬁciency, performance, interoperability,

and explainability. Future research will focus on re-

ﬁning models, expanding the ontology, and conduct-

ing clinical validation to further assess the system’s

effectiveness.

Personalized Asthma Recommendation System: Leveraging Predictive Analysis and Semantic Ontology-Based Knowledge Graph

133

REFERENCES

Ajami, H., Mcheick, H., and Laprise, C. (2022). First steps

of asthma management with a personalized ontology

model. Future Internet, 14(7):190.

Alharbi, E., Cherif, A., and Nadeem, F. (2023). Adaptive

smart ehealth framework for personalized asthma at-

tack prediction and safe route recommendation. Smart

Cities, 6(5):2910–2931.

Alharbi, E., Nadeem, F., and Cherif, A. (2021). Smart

healthcare framework for asthma attack prediction and

prevention. In 2021 National Computing Colleges

Conference (NCCC), pages 1–6. IEEE.

Anantharam, P., Banerjee, T., Sheth, A., Thirunarayan,

K., Marupudi, S., Sridharan, V., and Forbis, S. G.

(2015). Knowledge-driven personalized contextual

mhealth service for asthma management in children.

In 2015 IEEE international conference on mobile ser-

vices, pages 284–291. IEEE.

Barbaglia, G., Murzilli, S., and Cudini, S. (2017). Deﬁni-

tion of rest web services with json schema. Software:

Practice and Experience, 47(6):907–920.

Bose, S., Kenyon, C. C., and Masino, A. J. (2021). Per-

sonalized prediction of early childhood asthma per-

sistence: a machine learning approach. PloS one,

16(3):e0247784.

Chatterjee, A., Gerdes, M. W., and Martinez, S. G. (2020).

Identiﬁcation of risk factors associated with obesity

and overweight—a machine learning overview. Sen-

sors, 20(9):2734.

Chatterjee, A., Gerdes, M. W., Prinz, A., and Martinez,

S. G. (2021a). Comparing performance of ensemble-

based machine learning algorithms to identify poten-

tial obesity risk factors from public health datasets. In

Emerging Technologies in Data Mining and Informa-

tion Security: Proceedings of IEMIS 2020, Volume 1,

pages 253–269. Springer.

Chatterjee, A., Pahari, N., and Prinz, A. (2022a). Hl7

fhir with snomed-ct to achieve semantic and structural

interoperability in personal health data: a proof-of-

concept study. Sensors, 22(10):3756.

Chatterjee, A., Pahari, N., Prinz, A., and Riegler, M.

(2022b). Machine learning and ontology in ecoach-

ing for personalized activity level monitoring and

recommendation generation. Scientiﬁc Reports,

12(1):19825.

Chatterjee, A. and Prinz, A. (2022). Personalized recom-

mendations for physical activity e-coaching (ontore-

comodel): ontological modeling. JMIR Medical In-

formatics, 10(6):e33847.

Chatterjee, A., Prinz, A., Gerdes, M., and Martinez, S.

(2021b). An automatic ontology-based approach to

support logical representation of observable and mea-

surable data for healthy lifestyle management: Proof-

of-concept study. Journal of Medical Internet Re-

search, 23(4):e24656.

Chatterjee, A., Prinz, A., Riegler, M. A., and Meena, Y. K.

(2023). An automatic and personalized recommenda-

tion modelling in activity ecoaching with deep learn-

ing and ontology. Scientiﬁc Reports, 13(1):10182.

Cima, G., De Giacomo, G., Lenzerini, M., and Poggi, A.

(2017). On the sparql metamodeling semantics entail-

ment regime for owl 2 ql ontologies. In Proceedings of

the 7th International Conference on Web Intelligence,

Mining and Semantics, pages 1–6.

Dataset, A. D. P. (2024).

https://www.kaggle.com/datasets/deepayanthakur/

asthma-disease-prediction.

Garreau, D. and Luxburg, U. (2020). Explaining the ex-

plainer: A ﬁrst theoretical analysis of lime. In Interna-

tional conference on artiﬁcial intelligence and statis-

tics, pages 1287–1296. PMLR.

Hutchinson, C. F. (1982). Classiﬁcation improvement.

Photogrammetric Engineering and Remote Sensing,

44(1):123–130.

Kadariya, D., Venkataramanan, R., Yip, H. Y., Kalra, M.,

Thirunarayanan, K., and Sheth, A. (2019). kbot:

knowledge-enabled personalized chatbot for asthma

self-management. In 2019 IEEE International Con-

ference on Smart Computing (SMARTCOMP), pages

138–143. IEEE.

Morita, P. P., Yeung, M. S., Ferrone, M., Taite, A. K., Made-

ley, C., Lavigne, A. S., To, T., Lougheed, M. D.,

Gupta, S., Day, A. G., et al. (2019). A patient-

centered mobile health system that supports asthma

self-management (breathe): design, development, and

utilization. JMIR mHealth and uHealth, 7(1):e10956.

Olson, R. S. and Moore, J. H. (2016). Tpot: A tree-

based pipeline optimization tool for automating ma-

chine learning. In Workshop on automatic machine

learning, pages 66–74. PMLR.

OpenWeather (2024). https://openweathermap.org/api.

Pinnock, H., Noble, M., Lo, D., McClatchey, K., Marsh,

V., and Hui, C. Y. (2023). Personalised management

and supporting individuals to live with their asthma in

a primary care setting. Expert review of respiratory

medicine, 17(7):577–596.

WHO (2019). https://www.who.int/news-room/fact-

sheets/detail/asthma.

KEOD 2024 - 16th International Conference on Knowledge Engineering and Ontology Development

134