Novel graph-based machine-learning technique for viral infectious diseases: application to influenza and hepatitis diseases

Abstract

Background

Most infectious diseases are caused by viruses, fungi, bacteria and parasites. Their ability to easily infect humans and trigger large-scale epidemics makes them a public health concern. Methods for early detection of these diseases have been developed; however, they are hindered by the absence of a unified, interoperable and reusable model. This study seeks to create a holistic and real-time model for swift, preliminary detection of infectious diseases using symptoms and additional clinical data.

Materials and methods

In this study, we present a medical knowledge graph (MKG) that leverages multiple data sources to analyse connections between different nodes. Medical ontologies were used to enhance the MKG. We applied various graph algorithms to extract key features. The performance of multiple machine-learning (ML) techniques for influenza and hepatitis detection was assessed, selecting multi-layer perceptron (MLP) and random forest (RF) models due to their superior outcomes. The hyperparameters of both graph-based ML models were automatically fine-tuned.

Results

Both the graph-based MLP and RF models showcased the least loss and error rates, along with the most specific, accurate recall, precision and F1 scores. Their Matthews correlation coefficients were also optimal. When compared with existing ML techniques and findings from the literature, these graph-based ML models manifested superior detection accuracy.

Conclusions

The graph-based MLP and RF models effectively diagnosed influenza and hepatitis, respectively. This underlines the potential of graph data science in enhancing ML model performance and uncovering concealed relationships in the MKG.

Keywords:

• Automatic tuning
• graph algorithms
• graph machine-learning
• viral infectious disease
• medical knowledge graph
• influenza
• hepatitis

Introduction

The World Health Organization (WHO) [1] identifies infectious diseases as significant public health challenges. Diseases like lower respiratory tract infections rank among the top 10 global causes of death. Fatalities often occur from contagious respiratory diseases like COVID-19 and influenza, which stem from viral infections. Especially vulnerable to complications from these diseases are the elderly and infants with chronic illnesses [2].

The Coronaviridae family can lead to severe respiratory issues, including COVID-19, SARS and MERS [3]. Influenza epidemics happen yearly due to the virus’ high mutation rate and infectiousness. Its symptoms can vary from mild to severe. There are four types of influenza viruses, A, B, C and D, but only A and B cause the yearly flu epidemics [4].

Viral hepatitis, a liver infection, often causes symptoms such as yellow eyes and skin fatigue. The liver performs various functions, such as protein creation, toxin elimination and bile production [5]. Because of this, hepatitis-induced inflammation can lead to liver dysfunction and health decline in infected individuals.

Early disease identification plays a vital role in enhancing patient care, controlling disease spread and preventing future outbreaks. Several issues have been identified in previous studies focused on diagnosing influenza [6]. For instance, one problem is that the interpretability of the models has not been thoroughly examined [7]. There is also a reliance on Twitter data, which may not accurately represent the entire population and does not support health conditions or diseases beyond influenza.

Studies may be limited by a small sample size, lack of diversity, and a narrow feature set, and findings may not be compared to other methods [7,8]. Similarly, a study with only 116 participants was conducted in a controlled environment, limiting its applicability to broader scenarios [9]. Another case involves just 20 healthy adults, raising questions about the generalizability of its results [10]. In one instance, a study used a wearable device to monitor heart rates and activity levels but did not compare its performance to other diagnostic tests for influenza [11]. An issue was also found in another machine-learning (ML) model that applied synthetic patient data, rather than real patient data [12] and necessitating further validation. Furthermore, another study requires further validation to confirm its accuracy and applicability to other populations [13].

A study used graph neural networks to mimic human behaviour and identify symptoms of influenza using mobile sensing data [14]. However, it relied on Bluetooth interactions as a measure of social proximity, which could introduce bias. The lack of a laboratory-confirmed influenza diagnosis was another downside.

Fuzzy logic provides the ability to represent uncertainty and imprecision in data, although it can become computationally complex, particularly with large and intricate datasets [15]. An amalgamated model of linear support vector machine (SVM) and AdaBoost was developed using imbalanced datasets, yet it required further accuracy and sensitivity enhancements [16]. Enhanced ensemble learning attained more dependable and consistent diagnostic results [17]. However, its comparison to other cutting-edge methods and external validation were lacking. The random forest (RF) algorithm successfully predicted diseases with high accuracy, sensitivity and specificity [18]. Despite this, its exclusive reliance on the hepatitis C dataset from the UCI library may limit its representativeness and completeness. The use of electronic health records (EHRs) and ML classifiers for predicting hepatitis C and cirrhosis diagnoses was found to be effective, but the study was limited by a small sample size of only 75 patients [19]. Additionally, excluding other relevant features may undermine the model’s predictive accuracy. Although the research provided useful insights, potential biases or missing data could distort the findings [20]. The main challenges include the need for substantial computing power and the difficulty of implementing and maintaining tailor-made models in real-world healthcare environments [21]. Enhanced diagnostic capabilities for hepatitis based on a few features and effective class balancing with the synthetic minority over-sampling technique (SMOTE) strategy were among the reported benefits [22]. Nevertheless, the study only investigated a limited assortment of classifiers without exploring alternate ML algorithms, leaving its general applicability on other datasets questionable. Lastly, the relatively small dataset size may restrict the findings’ generalizability and potentially affecting the ML algorithms’ performance [23].

Researchers used artificial intelligence, ML and pathway enrichment analysis to identify virus-targeted cellular pathways. The study in [24] proposed a two-step ML process for evaluating COVID-19 pathways: gene selection, network development, pathway ranking, and a thorough analysis of top pathways. Rapid identification of effective medications is crucial in preventing worldwide disease transmission. In [25], researchers aimed to find human proteins that can bind to approved drugs. This approach was supported by human-virus protein–protein interactions (PPIs) and the biology of the host cell. They also identify topological and statistical features of proteins within the PPI network.

An entity-relation knowledge graph (KG) is a multi-dimensional graph where relationships (edges) and entities (nodes) illustrate the connections across various domains. Semantic Web technology provides a framework allowing for semantic interlinking, where data are annotated in a machine-readable format. This usually involves implementing ontologies, which provide a clear, formal representation of the domain’s knowledge.

This research applies a novel use of an existing methodology to generate MKG through intelligent health information retrieval [26–28]. Previous studies have utilized MKG for disease diagnosis [29–35] and drug analysis [36,37]. Additionally, the Graph-Guided Multi-Task Sparse Learning (GG-MTSL) model was employed to identify variations in the antigenic properties of the influenza A (H3N2) virus [38]. A separate study constructed biological and PPI networks using graphs [39]. These graphs are crucial for calculating substantial topological features. Ultimately, these features are incorporated into unsupervised learning algorithms to select genes relevant to SARS-CoV-2.

We developed an innovative method for creating the proposed medical knowledge graph (MKG) for viral infections, aiming to aid in early disease detection. This involved consolidating data from multiple sources into a unified graphical database. This integration culminated in a comprehensive MKG for COVID-19, influenza and hepatitis, simplifying the understanding of the complex relationships between the MKG nodes and boosting the proposed detection model’s performance. Our research reaffirms the utility of graph algorithms in extracting unique features from specific datasets. Furthermore, we highlight how graph-based ML can enhance detection model performance and unveil concealed connections.

Methods

Building medical knowledge graph

We created a comprehensive MKG for COVID-19, influenza and hepatitis following the methods outlined by Gao et al. [40] and Chen et al. [41]. The study integrates data from various sources, according to the graph model depicted in (Figure 1) and summarized in Table 1. Our goal is to identify the critical nodes and connections within this model. The import processes are illustrated in Table 2.

Figure 1. Schema of the proposed graph model.

Table 1. Heterogeneous data implemented in the proposed MKG.

Ref.	Name	Brief definition
[42]	HDO, influenza and hepatitis ontologies	It contains a comprehensive hierarchical representation of human diseases. It covers infectious diseases, transmission methods, microorganisms and symptoms in 18,019 classes. It contains symptoms of influenza and liver inflammation caused by a viral infection but also by toxic chemicals.
[43]
[44]
[45]	KEGG	It contains valuable and extensive knowledge about infectious diseases, their causes, pathogens and drugs.
[46]	Wikidata	It mentions infection types, symptoms, transmission, hosts, diagnostic methods and treatments.
[47]	LitCovid	It is a unique online system for keeping track of the most recent articles and material on the 2019 coronavirus.
[48]	CORD-19	It is a free resource for the global academic community that contains several thousand scholarly publications concerning COVID-19, SARS-CoV-2 and other related coronaviruses.
[49]	CoV-AbDab	It includes data on antibodies and nanobodies that have been shown to bind to at least one beta-coronavirus.
[50]	STRING	It provides information on known and expected human interactions between proteins.
[51]	UniProtKB	It is the principal repository for functional data on proteins with accurate, consistent and comprehensive annotation.
[52]	DrugBank	It is a global source of structured medication data and patient insight solutions that help speed drug development and enhance healthcare delivery.
[53]	PubTator	It is a web-based system that automatically annotates biomedical concepts like genes and mutations in PMC papers and PubMed abstracts.
[54]	iTextMine	RLIMS-P’s abstract mining of the protein phosphorylation of LitCovid.
[55]	SemRep	It extracts semantic predictions from biological text using the Unified Medical Language System (UMLS).
[56]	iPTMnet	In systems biology, this bioinformatics tool explains protein post-translational modifications (PTMs).

Table 2. Importing procedures for heterogeneous data into a graph model.

Data source	Importing procedure	Focus	Output format
Medical Ontologies	Importing through n10s.rdf.import.fetch procedure. Importing classes, class hierarchies, properties and property hierarchies.	HDO, influenza and hepatitis	RDF/XML
KEGG	The KEGG data were extracted with its API. Parsing text files into CSV files to generate nodes and edges for diseases, drugs, taxons and pathogens.	Genome-based classification of infectious diseases (viral infections).	CSV files
Wikidata	Taxonomies were extracted through several SPARQL queries that return triples. Importing triples to the graph model through n10s.rdf.import.fetch procedure.	Virus taxonomies for influenza and hepatitis diseases.	RDF
LitCovid	iTextMine, PubTator and SemRep can retrieve COVID-19 information from these databases. This is made by using semantic web technologies, specifically the resource description framework (RDF). Applying application programming interface (API) supported in [41] Submitting a POST request and get the response in the form of cypher queries for the graph model to the following endpoint: http://<host>:<port>/db/<databaseName>/tx/commit	The focus is to turn the new information into a standard and accessible COVID-19 knowledge graph (KG).	The RDF dumps have been made available for download.
CORD-19
CoV-AbDab
STRING
UniProtKB
DrugBank
iPTMnet

The graph model consists of 13 nodes and 14 edges, with one of the nodes representing a patient, as extrapolated from the dataset. This model utilizes a graph ML technique to sort patients with particular infectious diseases. The model provides comprehensive data relevant to the treatment and management of such diseases.

Each infectious disease illustrated in the graph has its causes, symptoms and methods of transmission. Infectious diseases are primarily transmitted via pathogens and reservoirs, which could be of animal, human or environmental origin.

Pathogens are microorganisms that cause diseases in living cells. Genes, the hereditary units of living cells, also significantly play a role in infection susceptibility. Certain pathogenic entities, known as viruses, contain a viral genome, granting them the power to infiltrate and infect living cells. A ‘viral genome,’ to elaborate, is a virus’s genetic content. Viruses are grouped based on shared traits under taxonomic classifications called ‘viral families.’

The relationship between drugs and pathogens involves drugs’ unique targeting of specific biological aspects of the pathogens and their effectiveness in treating the pathogens. The drugs are used to treat, prevent and control infectious diseases, depending on the specific pathogenic agent.

Figure 2 illustrates the method where a healthcare provider enters patient symptoms and clinical data into the MKG through an EHR at the point of care. Each node within the MKG represents a vertex and is connected to multiple graphs. We utilized this model to classify infectious diseases and fetch all related information from the connected graphs.

Figure 2. General schema of the proposed method.

Medical providers can gain crucial insights such as disease complications and suitable therapies from the proposed MKG by using similarity algorithms like cosine similarity. After building the desired MKG, we created and applied a ML model for node classification on graphs of influenza and hepatitis datasets. We store the optimal prediction models in our model catalogue. For future utilization and training on new datasets, we preserve the model in our database. This entire process aids in decision-making and enhances the MKG.

The accuracy of our ML model’s predictions was enhanced by employing graph algorithms, including scale properties and Fast Random Projection (FastRP). Furthermore, automated hyperparameter tuning guaranteed optimal values for each hyperparameter, consequently maximizing the model’s predictive accuracy.

Influenza disease dataset

The dataset used in this study, which aimed to distinguish influenza from influenza-like illness (ILI), was based on records collected from both inpatients and outpatients over a span of five days [57]. It consisted of 4569 entries and 16 features capturing symptoms and laboratory tests associated with influenza, along with a target feature for identifying infection.

The cross-sectional study included demographic, clinical and exposure data. Subjects were tested for influenza types A and B. Table 3 provides a list of the 17 features, brief descriptions and their respective values.

Table 3. Features of the influenza dataset.

Feature	Description	Value
Age	How old is the patient?	Frequency from 1 to more than 60 years old
Gender	Male or female	1: male, 2: female
Vaccine	Did the patient take the influenza vaccine in the last 12 months?	0 or 1
Cough	It is a protective reflex to remove irritates from the throat or airways.	0 or 1
Malaise	Called fatigue.	0 or 1
Sore throat	It is irritation or scratchiness of the throat that often worsens when the patient swallows.	0 or 1
Runny nose	Mucus being discharged from the nostrils.	0 or 1
Chill	Feeling cold.	0 or 1
Fever	Rise in body temperature.	0 or 1
Diarrhoea	Loose and watery stool during a bowel.	0 or 1
Breathing	Uncomfortable feeling of not being able to breathe well enough.	0 or 1
Headache	Pain in any region of the head.	0 or 1
RIDTs	Identifies the presence of influenza A and B and displays the result qualitatively (positive vs. negative).	0 or 1
RT-PCR	It is a real-time detection test for influenza type in respiratory specimens.	0: negative. Otherwise, 1
Injected_pharynx	Determines whether pharynx appeared irritated or red.	0 or 1
Injected_tympanic	Determines whether inner tympanic is infected	0 or 1
Flu Result	Target result whether it is negative or positive	0 or 1

Datasets preprocessing

In the influenza dataset, a value of 0 denotes no infection, while a value of 1 signifies an influenza infection. The number of infected males, at 2487, was higher than the number of infected females, at 2082. Patients’ ages ranged from 1 to 60 years. The dataset included 3076 cases (67.32%) that were not infected and 1493 cases (32.68%) that were infected. The synthetic minority oversampling technique (SMOTE) was employed to balance the data in the influenza dataset due to the underrepresentation of the minority class. Table 4 shows that five factors had a stronger association with the final flu result: age, malaise, chills, a runny nose, and a cough. To process this data, we converted all ‘yes’ values to 1 and all ‘no’ values to 0.

Table 4. Phi-values and p values for categorical features in the influenza dataset.

Feature	Phi-value	p Value
Gender	−0.020921	.1573706
Vaccine	−0.118448	<.00001
Fever	0.095431	<.00001
Cough	0.126149	<.00001
Malaise	0.169940	.000002
Sore throat	−0.043981	.2684052
Runny nose	0.165930	.000008
Chill	0.209078	<.00001
Diarrhoea	−0.038443	.0058791
Breathing	−0.081991	<.00001
Headache	0.037192	.0119308

The calculated Pearson correlation coefficient (PCC) for age yields a value of 0.214051, suggesting a correlation with the target class, Flu Result. Table 4 presents phi coefficient values and p values for other categorical features within the influenza dataset.

From Table 4, our findings indicated that gender, vaccination status, fever, diarrhoea, breathing issues, sore throat and headaches have no relation to the flu result. Contrastingly, features such as a cough, malaise, runny nose and chills share an association with flu results. This study exclusively focuses on signs and symptoms, therefore disregarding unrelated features such as injections and RT-PCR tests.

Hepatitis disease dataset

The hepatitis dataset, procured from the UC Irvine ML repository [58], consists of 155 observations with 19 features and a target feature. This dataset is imbalanced, as 74% of patients survive while 26% die, posing a significant challenge to the classification process.

Datasets preprocessing

Just as we did with the influenza dataset, we utilized SMOTE for data balancing in the hepatitis dataset although the datasets do have a few missing data points, the most considerable absence is found in the prothrombin time (PROTIME) feature, where 67 data points are missing. This presents a significant 43% deficiency in the dataset. Other missing data points constitute less than 10% of the total dataset. Training a ML model using these incomplete datasets might significantly impact its quality.

Categorical features originally reflecting ‘yes’ or ‘no’ were changed to numeric values ‘1’ and ‘2’. Table 5 displays hepatitis dataset features.

Table 5. Features of the hepatitis dataset.

Feature	Description	Value
Age	How old is the patient?	From 10 to 80
Gender	What is the sex of the patient?	1 or 2
Steroid	It is a response to corticosteroid treatment.	1 or 2
Antivirals	It is a response to antiviral treatment.	1 or 2
Fatigue	It is the main symptom of a patient with liver disease.	1 or 2
Malaise	It is a general feeling of discomfort.	1 or 2
Anorexia	It represents an eating disorder.	1 or 2
Liver big	It denotes an enlarged liver.	1 or 2
Liver firm	It indicates scarring or liver cirrhosis.	1 or 2
Spleen palpable	It is an external examination that detects whether a spleen is palpable.	1 or 2
Varices	It is a large blood vessel in the oesophagus.	1 or 2
Spiders	It is a type of telangiectasis (swollen blood vessels) that appears beneath the surface of the skin.	1 or 2
Ascites	It is a severe liver disease that causes a build-up of fluid in the belly.	1 or 2
Bilirubin test	It measures the level of bilirubin in the blood.	0.39, 0.80, 1.20, 2.00, 3.00, 4.0
Alk phosphate	This test measures the level of alkaline phosphatase enzyme.	33, 80, 120, 160, 200, 250
SGOT test	This test measures the level of liver enzymes.	13, 100, 200, 300, 400, 500
Albumin	It measures the level of protein produced by the liver.	2.1, 3.0, 3.8, 4.5, 5.0
PROTIME test	Assesses blood clotting in seconds.	10, 20, 30, 40, 50, 60, 70, 80, 90
Histology	It is the microscopic anatomy.	1 or 2
Class	It is the outcome of the patient’s class.	1 or 2

We selected all continuous features correlated to the target feature, excluding the PROTIME feature due to its substantial missing values. The correlated features were determined by calculating PCCs. These correlated features are shown in Figure 3.

Figure 3. Correlated features with class target feature.

Table 6 reveals that bilirubin and albumin are the features most closely correlated with the class category. Furthermore, we calculated the phi coefficient and p value (at p < .05) for other categorical features in the dataset, as shown in Table 6, to identify features associated with the target class. Our analysis indicated no significant correlations between gender, steroid usage, antivirals, liver size, liver firmness, histology and the target class. However, fatigue, malaise, loss of appetite (anorexia), palpable spleen, varices, spider nevi and ascites were found to have an association with the target class. These latter features were, therefore, chosen to train ML models throughout this study.

Table 6. Phi-values and p values for categorical features in the hepatitis dataset.

Feature	Phi-value	p Value
Gender	0.1627052	.053035577
Steroid	0.086206	.307688566
Antivirals	−0.109289	.195423715
Fatigue	0.278513	.000790218
Malaise	0.341806	.000031382
Anorexia	0.166677	.047420057
Liver_big	−0.075417	.372379243
Liver_firm	0.081184	.336822536
Spleen_palable	0.188224	.024882765
Spiders	0.402216	<.00001
Ascites	0.498498	<.00001
Varices	0.386316	<.00001
Histology	−0.310245	.00017164

To mitigate the missing data problem in both datasets: for continuous variables (as we had non normal distribution) and for categorical variables, we substituted with the most common value (mode).

Infectious disease detection through node classification method

The graph ML method has been employed in numerous studies. For instance, it was utilized for the classification of diabetes [59] and the investigation of Alzheimer’s disease [60]. In this study, we conducted a binary node classification on a training graph. The goal was to classify unlabelled nodes for influenza and hepatitis based on the features of their fellow nodes.

The training pipeline started by enhancing the graph with newly extracted features, as well as the dataset’s existing features. This augmented graph was then employed to train our proposed model. Essentially, selected graph algorithms – FastRP and scaled properties – were used to create new node features.

Following this, a selection of the node features was used, with nodes split into training, testing and validation phases for the ML model. The process involved dividing the entire graph into two sections – training and testing. We also implemented a stratified 10-fold cross-validation on the training graph. This ensured that each partition contained training and validation sets and provided the necessary performance data. To conclude, the proposed model was employed to classify the unlabelled nodes.

The training process of the node classification pipeline implemented in [61] was applied to detect influenza and hepatitis diseases in this study as indicated in Algorithm 1.

Algorithm 1

Training node classification pipeline

Require: Infectious disease dataset

Ensure: Split all dataset graphs into training and testing graphs

1: for All training graphs do

2: Divide into validation folds = k

3:  for Each cross-validation evaluation do

4:   Strati ed training graphs

5:   for All train and validation parts ∈Validation folds do

6:    Train on training sets

7:    Evaluate on validation sets

8:    if Performance metrics are the best for a model  candidate then

9:     Select the winning model hyperparameters

10:    else

11:      Select other values for hyperparameters

12:    end if

13:   end for

14:  end for

15: end for

16: for Winning model candidate do

17:  Retrain on entire training graph

18:   Evaluate on testing graph

19:   Retrain on entire original graph

20:   if Performance metrics are the best for a model candidate then

21:    Return winning model registration

22:    Return performance metrics

23:    Save models in a model catalog

24:    Apply the model for new prediction

25:   else

26:   Select other values for k and for hyperparameters

27:  end if

28: end for

Results

Applying the graph-based model for influenza and hepatitis detection

The multi-layer perceptron (MLP) model was used to train 4569 nodes of an influenza dataset. We used the hold-out method for node classification, dividing the graph into training and testing subgraphs, comprising 80% and 20%, respectively. The training graph was then subjected to a stratified 10-fold cross-validation.

Features extracted from the FastRP and scale property algorithms were included in the training, along with node properties – a subset of the influenza dataset features. We sought to identify a target feature indicating whether a patient was infected with influenza.

As part of the training process, the MLP model underwent automatic tuning to determine optimal hyperparameter values. Table 7 sets out the results from this tuning process, detailing the best maximum and minimum epochs, patience and tolerance, learning rate, batch size, penalty, hidden layer sizes and class weights.

Table 7. Values for hyperparameters automatic tuning results.

Proposed model	Disease	Hyperparameter settings
Graph MLP	Influenza	max epochs = 100, min epochs = 1, class weights = [0.67, 0.33], penalty = 0.5, patience = 2, focus weight = 0.0904839039446324, learning rate = 0.001, hidden layer size = [5], tolerance = 0.001, batch size = 100
Graph RF	Hepatitis	Max depth: 224,387, Criterion: Gini, Min leaf size: 1, Min split size: 4, Max features ratio: 1.0, Decision trees: 10, Samples Ratio: 1.0

An average training accuracy of 98.62% and a weighted average F1 score of 0.975 were achieved with the MLP model in training. Additionally, the trained model was used to detect unclassified patient nodes in the testing graph. This was achieved using a sigmoid activation and a cross-entropy loss function, the latter also known as logarithmic loss. Figure 4 presents the confusion matrix for the testing graph.

Figure 4. Influenza testing graph confusion matrix.

The RF model, trained on 155 nodes from the hepatitis dataset, employed a hold-out method that split each dataset’s graph into subgraphs for training (80%) and testing (20%). The training graph underwent a 10-fold stratified cross-validation. The training incorporated features were extracted by the FastRP and scale property algorithms, commencing with node properties, a subset of the dataset’s characteristics. Through training, the method conducts automatic tuning on the ML model to identify ideal hyperparameter values, with Table 7 detailing the optimal hyperparameters determined. The RF model achieved a perfect accuracy of 100% and a weighted average F1 score of 0.999 in average training. Furthermore, the trained ML model was applied to predict infection and classify previously unidentified patient nodes in the testing graphs, depicted in Figure 5.

Figure 5. Hepatitis testing graph confusion matrix.

Table 8 presents the statistical parameters for the graph MLP and RF classifiers when applied to influenza and hepatitis datasets, as derived from their respective confusion matrices. The MLP model demonstrated a latency of 6001 ms, while the RF model indicated a lower latency of 4001 ms. Their performance was further evaluated using the Matthews correlation coefficient (MCC), with values approaching or equal to one suggesting successful classification. Another testament to the efficacy of these models is the log loss, which was calculated to be near zero, as well as the error rate, which was precisely zero, further establishing their accuracy and reliability.

Table 8. Performance values for the graph ML models.

Measure	Graph MLP	Graph RF	Formula
Recall	0.9444	1.0000	$\frac{\mathrm{\text{TruePositive}}}{\mathrm{\text{TruePositive}}+\mathrm{\text{FalseNegative}}}$
Specificity	0.9898	1.0000	$\frac{\text{TrueNegative}}{\mathrm{\text{FalsePositive}}+\mathrm{\text{TrueNegative}}}$
Precision	0.9808	1.0000	$\frac{\mathrm{\text{TruePositive}}}{\mathrm{\text{TruePositive}}+\mathrm{\text{FalsePositive}})}$
Negative predictive value	0.9701	1.0000	$\frac{\mathrm{\text{TrueNegative}}}{\mathrm{\text{TrueNegative}}+\mathrm{\text{FalseNegative}}}$
False positive rate	0.0102	0.0000	$\frac{\mathrm{\text{FalsePositive}}}{\mathrm{\text{FalsePositive}}+\mathrm{\text{TrueNegative}}}$
False discovery rate	0.0192	0.0000	$\frac{\text{FalsePositive}}{\mathrm{\text{FalsePositive}}+\mathrm{\text{TruePositive}}}$
False negative rate	0.0556	0.0000	$\frac{\mathrm{\text{FalseNegative}}}{\mathrm{\text{FalseNegative}}+\mathrm{\text{TruePositive}}}$
Accuracy	0.9737	1.0000	$\frac{\mathrm{\text{TruePositives}}+\mathrm{\text{TrueNegatives}}}{\mathrm{\text{TruePositives}}+\mathrm{\text{TrueNegatives}}+\mathrm{\text{FalsePositives}}+\mathrm{\text{FalseNegatives}}}$
F1 score	0.9623	1.0000	$2\times \frac{(\mathit{\text{Precision}}\times \mathit{\text{Recall}})}{(\mathit{\text{Precision}}+\mathit{\text{Recall}})}$
MCC	0.9425	1.0000	$\frac{\mathit{\text{TP}}\times \mathit{\text{TN}}-\mathit{\text{FP}}\times \mathit{\text{FN}}}{\sqrt{(\mathit{\text{TP}}+\mathit{\text{FP}})\times (\mathit{\text{TP}}+\mathit{\text{FN}})\times (\mathit{\text{TN}}+\mathit{\text{FP}})\times (\mathit{\text{TN}}+\mathit{\text{FN}})}}$
Error rate	–	0.0000	$\frac{\mathit{\text{FalsePositive}}+\mathit{\text{FalseNegative}}}{ \mathit{\text{TruePositives}}+\mathit{\text{TrueNegatives}}+\mathit{\text{FalsePositives}}+\mathit{\text{FalseNegatives}}}$
Log loss	0.026	–	$\mathrm{\text{Log}} \mathit{\text{Loss}}\left(x,y\right)=-\frac{1}{N}\times {\sum }_{i=1}^n(y_i\times \mathrm{\text{Log}}(\widehat{y})+(1-y_i)\times \mathrm{\text{Log}}(1-\widehat{y}$))

Efforts have been extensive in detecting influenza and hepatitis. This study enhances such efforts by automatically fine-tuning MLP and RF models. We leveraged the feature permutation importance method to identify the model’s most critical features in Figures 6 and 7 that highlight the key features affecting the proposed graph ML models’ accuracy. FastRP is the most influential feature, followed by scaled features. Not including these features significantly diminishes the model’s accuracy, as they provide crucial information.

Figure 6. Most significant features of the graph MLP model.

Figure 7. Most significant features of the graph RF model.

Furthermore, Figure 8 displays the results of the receiver operating curve (ROC) analysis for the proposed graph MLP and RF models. Both models achieved nearly perfect scores, with their areas under the curve being close to one.

Figure 8. Testing graph ROC curve analysis for both classes in the datasets.

Exploring various relations in the proposed MKG

The cosine similarity algorithm was employed, revealing numerous links between nodes in the complete MKG. Figure 9 displays a notable connection between the lung cancer node and the graphs for COVID-19 and influenza. This suggests an increased likelihood of lung cancer due to influenza exposure, which could be worsened by recurrent exposure, as described in [62]. However, unreliable data on smoking may introduce confounding factors, necessitating more research to definitively establish the link between influenza and lung cancer. Moreover, it is possible that SARS-CoV-2 shares pathophysiological traits similar to an oncovirus found in the lung. The authors of [63] summarized three potential cancer-causing mechanisms of SARS-CoV-2 that could contribute to the occurrence of lung cancer.

Figure 9. Exploring the possibility of lung cancer in the MKG.

Figure 10 illustrates a newly discovered correlation between renal failure and chronic kidney disease. This profound connection was unearthed from the hepatitis B node in the hepatitis graph with the disease in the anatomical entity graph. Furthermore, this connection has been substantiated [64], showing renal failure as a possible outcome of advanced chronic kidney disease stages.

Figure 10. Chronic kidney disease and hepatitis B relation.

Figure 11 illustrates a newly established correlation between hepatitis C and type 2 diabetes mellitus, as proven in the MKG [65]. The liver plays an indispensable role in glucose metabolism and insulin regulation. Nonetheless, a hepatitis C infection can interfere with these metabolic processes, triggering insulin resistance and flawed glucose management.

Figure 11. Type 2 diabetes mellitus disease and hepatitis C relation.

Figure 12 displays an additional segment of the influenza graph. While most individuals recover from influenza within days, some may develop serious complications like bacterial pneumonia. Certain complications, including myocarditis (heart inflammation), myositis and rhabdomyolysis (muscle tissue inflammation), encephalitis (brain inflammation) and multiorgan failure, such as respiratory and kidney failures, can be fatal. Medications such as Xofluza (baloxavir marboxil), Tamiflu (oseltamivir phosphate), Relenza (zanamivir) and Rapivab (peramivir) are used for treating types A and B of influenza. Figure 13 shows classified patient nodes for both influenza and hepatitis diseases.

Figure 12. Part of influenza graph for different treatment medicine.

Figure 13. Classified nodes infected with influenza and hepatitis diseases.

Discussion

Comparison between machine-learning algorithms

We evaluated nine ML algorithms: K-nearest neighbors (KNN), decision tree (DT), SVM, stochastic gradient descent (SGD), RF, MLP, naive Bayes, logistic regression (LR) and AdaBoost. The comparison of these ML algorithms’ performance on the influenza and hepatitis datasets is presented in Tables 9 and 10. Because the MLP and RF models outperformed the others, we chose them.

Table 9. Comparison between detection performances of various ML algorithms for the influenza dataset.

Algorithm	Accuracy	Precision	F1 score	Recall
kNN	0.90047	0.92164	0.89643	0.94830
DT	0.83421	0.90699	0.89493	0.91937
SVM	0.80066	0.60169	0.82456	0.47366
SGD	0.83933	0.90574	0.88468	0.92782
RF	0.91410	0.91396	0.89777	0.93075
MLP	0.92586	0.92465	0.88381	0.96944
Naive Bayes	0.91615	0.92250	0.87514	0.97529
LR	0.87541	0.90308	0.88315	0.92392
AdaBoost	0.82013	0.88095	0.88383	0.87808

Table 10. Comparison between detection performances of various ML algorithms for the hepatitis dataset.

Algorithm	Accuracy	Precision	F1 score	Recall
kNN	0.77477	0.79527	0.80794	0.82580
DT	0.95312	0.98694	0.98730	0.98709
SVM	0.96443	0.96714	0.96768	0.96774
SGD	0.9375	0.97353	0.97500	0.97419
RF	0.98653	0.98694	0.98730	0.98709
MLP	0.85086	0.85016	0.84892	00.85806
Naive Bayes	0.98475	00.93086	0.93518	0.92903
LR	0.98678	0.96030	0.96138	0.96129
AdaBoost	0.95655	0.96792	0.96822	0.96774

Feature extraction process

In this study, we analysed features selected for the influenza and hepatitis datasets. Additionally, we included crucial features obtained through the FastRP and scale properties algorithms.

Fast Random Projection algorithm

The performance of graph ML can be improved using the FastRP method. This method efficiently reduces the graph data’s dimensions while preserving the essential structural information. Even with sparse datasets such as those of influenza and hepatitis, the FastRP algorithm proves to be highly efficient. It excellently reduces network data dimensions while retaining key structural details. Thus, the use of the FastRP algorithm enables enhanced performance in graph ML.

To apply the FastRP algorithm [66], we first calculated node similarity. This algorithm assesses a group of nodes in reference to the connected nodes within a graph. We measured pairwise similarities using the Jaccard metric, which calculates the similarity by dividing the intersection size of two node sets by the size of their union, as per the given formula: (1) $$J\left(A,B\right)=\frac{\left|A\cap B\right|}{\left|A\cup B\right|}=\frac{\left|A\cap B\right|}{\left|A\right|+\left|B\right|-\left|A\cap \mathrm{B}\right|}$$(1)

The FastRP was obtained through the following steps:

• The Jaccard similarity results were used to obtain node similarity matrix $M^{n\times n}$

• Dimensionality reduction (random projection matrix) was performed through:

(2) $$N^{d\times n}=M^{n\times n}\times R^{d\times n}$$(2) where n is the number of nodes in the graph and d is the dimensionality, $M^{n\times n}$ is the transition matrix, and $R^{d\times n}$ is the dimensionality reduction matrix. In the end, the FastRP algorithm is calculated through the following equation: (3) $$\mathit{\text{FastRP}}=\hspace{0.17em}\left\{\begin{array}{c}{N}_1=A\times L\times R\\ N_i=A\times N_{i-1}\\ N={\alpha }_1\times N_1+\dots +{\alpha }_k\times N_k\end{array}\right.$$(3)

The final matrix was obtained by multiplying the transition, normalization and random projection matrices. The end result was a combination of random projection matrices from all stages, each enhanced by a specific factor to determine its individual strength.

Scale properties algorithm

We used a scale properties algorithm to preprocess node features prior to model training. This determined the range and distribution of the processed data. This is particularly relevant when dealing with datasets featuring a wide range of values, such as the ‘age’ feature in the influenza dataset, or those skewed towards specific values, like the ‘albumin’ and ‘bilirubin’ features in the hepatitis dataset used for our study. These features are scaled using the MinMaxScaler. (4) $$p_{\mathit{\text{scaled}}\hspace{0.17em}}=\hspace{0.17em}\frac{p-\mathit{\text{min}}\left(p\right)}{\mathit{\text{max}}\left(p\right)-\mathit{\text{min}}\left(p\right)}$$(4)

The vector ‘p’ holds all the feature values for a single feature across all nodes in a graph. This method normalizes the data to a specific range, typically 0–1, by subtracting the minimum value from each data point and dividing the result by the range. The MinMaxScaler is particularly useful for normalizing features on different scales. It ensures uniform minimum and maximum values across all features, keeping the data tidy and well-organized. The MinMaxScaler offers several advantages; it maintains consistent feature values and handles outliers more effectively than standard methods. By setting the lowest and highest data points as 0 and 1, it diminishes the impact of extreme values on scaling. Furthermore, the MinMaxScaler preserves the original data range, making the scaled features easier to understand.

In this study, we implemented a MinMaxScaler for our proposed graph-based ML model. This is to retain the structure of the original distribution and interpret the scaled features’ impact on the target variable. It is also necessary for eliminating bias, achieving balanced, normalized presentations and enhancing the model’s performance. These scaled features influence the model’s efficacy based on dataset attributes and activation functions.

In our graph-based ML model, the magnitude of the input features affects the network’s weights and biases. If features vary in size, larger ones may disrupt the learning process, leading to subpar performance. By normalizing features to a consistent range, we ensure equal learning impact across the board.

Multi-layer perception model

The MLP model is a feedforward neural network employed for classifying or labelling nodes in a graph. It consists of multiple interconnected layers that modify incoming data. The first layer, known as the input layer, integrates the features from the graph. The following hidden layers perform non-linear transformations to extract significant features from the input data. The final layer, or the output layer, creates a probability distribution for each node across all possible labels. The MLP is a supervised learning method that learns the following functions: (5) $$F\left(x\right)={\sum }_{i=1}^n{x}_{i\mathrm{ }}{w}_{i\mathrm{ }}+b\mathrm{ }$$(5) where x is the input neuron, w is the weight and b is the bias. The sigmoid function is an activation function and is expressed as: (6) $$\sigma \mathrm{ }\left(z\right)=\frac{1}{1-e^{-z}}$$(6)

Figure 14 illustrates the structure of our uncomplicated graph-based MLP model implemented in this research. For classifying nodes, the model uses seven node features (age, malaise, chills, runny nose and cough) and two additional features (FastRP and scaled age) as inputs. It has one hidden layer with five neurons. The model identifies each class – infected or uninfected – via an output layer with a single neuron and utilizes a sigmoid activation function.

Figure 14. The architecture of the proposed graph MLP model for binary classification.

Random forest model

The Ensemble Learning RF model is a widely used, supervised ML method. This method integrates multiple predictors, deploying numerous DTs that are individually trained on unique sections of the training data. Each tree’s prediction is then collected into a single, unified forecast. Importantly, training each DT independently aids in preventing overfitting. Classification using the RF model is presented in Equation (7)(7) $${\widehat{C}}_{\mathit{\text{rf}}}^B\left(x\right)=\hspace{0.17em}\mathit{\text{majority vote}}\hspace{0.17em}\left\{{\widehat{C}}_{b\hspace{0.17em}}{\left.\left(x\right)\right\}}_1^B\right.$$(7) . (7) $${\widehat{C}}_{\mathit{\text{rf}}}^B\left(x\right)=\hspace{0.17em}\mathit{\text{majority vote}}\hspace{0.17em}\left\{{\widehat{C}}_{b\hspace{0.17em}}{\left.\left(x\right)\right\}}_1^B\right.$$(7)

Equation (7)(7) $${\widehat{C}}_{\mathit{\text{rf}}}^B\left(x\right)=\hspace{0.17em}\mathit{\text{majority vote}}\hspace{0.17em}\left\{{\widehat{C}}_{b\hspace{0.17em}}{\left.\left(x\right)\right\}}_1^B\right.$$(7) begins with a random selection of a bootstrap sample from the training dataset. These bootstrap data are then analysed using an RF tree. Each tree’s final node is assigned a random variable. Subsequently, the ideal variable for splitting is chosen, dividing the node into two. This procedure is repeated for all terminal nodes in the tree until their size is minimized. Finally, the ensemble of trees and the predicted class of the RF tree are generated as outputs. In the RF model for graphs, we have several hyperparameters that need adjustment to balance training speed and memory utilization, as well as bias and variance.

Predicting relationships in the proposed MKG

After importing all the data, we use a KNN graph algorithm to identify key features for each node in the graph, and then we write these values back into the proposed MKG.

K-Nearest neighbors algorithm

The similarity algorithm proposed by Dong et al. [67] calculates distance values for all paired nodes in a graph, establishing new connections between each node and its closest neighbors. In our study, we used the FastRP feature to determine the similarity measure for the KNN algorithm.

Cosine similarity

We used cosine similarity to identify similar nodes in the MKG due to the FastRP feature being a floating-point number. Calculated by determining the cosine of the angle between two vectors, it establishes whether the vectors point in the same direction, as shown in the following equation: (8) $$\mathit{\text{Cosine}}\left(p_s,p_t\right)=\frac{\sum _i{p}_s\left(i\right).p_t\left(i\right)}{\sqrt{\sum _i{p}_s{\left(i\right)}^2.}\mathrm{ }\sqrt{\sum _i{p}_t({i)}^2}}$$(8)

Equation (8)(8) $$\mathit{\text{Cosine}}\left(p_s,p_t\right)=\frac{\sum _i{p}_s\left(i\right).p_t\left(i\right)}{\sqrt{\sum _i{p}_s{\left(i\right)}^2.}\mathrm{ }\sqrt{\sum _i{p}_t({i)}^2}}$$(8) shows the vectors’ dot product divided by the product of their lengths. The above algorithm yields a score between [–1]. Score = (score + 1)/2 is used to standardize the score into the range [0, 1].

The flowchart in Figure 15 illustrates the research process used to mitigate overfitting and demonstrate the generalizability of the graph-based ML model by handling the dataset effectively. The node classification process begins with pipeline creation. We first import and preprocess an infectious diseases dataset, then manually apply graph algorithms. FastRP is used in this study because our dataset graphs are sparse, and we use scaled properties to prevent biased results from certain features. Next, we add node features and split the nodes into training and testing graphs. The training graph is further divided into training and validation graphs, on which we perform stratified k-fold cross-validation. For testing graphs, we utilize a confusion matrix. Automatic hyperparameter tuning and regularization techniques like L1, L2, and the dropout method are implemented. The necessary performance metrics and feature importance analysis are generated for the optimal graph ML model. Finally, we compare our results with other related studies.

Figure 15. Flowchart of the proposed method.

Figure 16 illustrates that our proposed graph MLP and RF models surpassed other models from related studies in accuracy. It is evident from Figure 16 that we achieved better accuracy in influenza detection compared to previous best model [11]. Moreover, we achieved superior accuracy for hepatitis detection than [18], as they achieved 99.9% with the RF model and the same dataset.

Figure 16. Comparison between the graph-based ML models and other related studies.

Strengths and limitations

Our study boasts several advantages. First, it uses graph algorithms for node embedding (FastRP) and scale properties to obtain additional features. These extracted features aptly represent the nodes as vectors while preserving the graph’s structure. The use of vertex embeddings simplifies the task of classifying similar patients. The graph ML model achieves convergence considerably quicker with feature scaling than without it, preventing premature saturation. Moreover, the study incorporates an automatic hyperparameter tuning procedure, requiring mere milliseconds to present all results. The employment of graph-based ML allows the fusion of heterogeneous data, enhancing the performance of ML models. We ensure the inclusion of all crucial features in the datasets.

However, our research does have certain limitations. First, medical practitioners could further this field by adopting our graph model in conjunction with large and live datasets. Second, although the primary features in our datasets are signs and symptoms, incorporating more features from diverse sources could yield higher effectiveness. Notably, laboratory data and radiology reports comprise valuable data in EHRs that can enhance the reliability of diagnoses in clinical settings. Furthermore, our research is confined to viral infectious diseases, warranting trials for bacterial, fungal and parasitic diseases. Lastly, handling large graphs presents a challenge but can be overcome by using a large-scale distributed graph data method.

Conclusions

This study enhanced the detection of influenza and hepatitis diseases using graph ML models. The proposed MKG was built using a variety of data sources, including HDO, KEGG, Wikidata, COVID-19, influenza ontology and hepatitis ontology. These diverse data sources aid in understanding various relationships within the MKG, such as infectious agents, hosts, treatments, complications and other key aspects. To extract additional features from the influenza and hepatitis datasets, we employed graph algorithms. The graph ML models we proposed outperformed other ML algorithms tested on the same dataset. We used cosine similarity algorithms to discover relationships.

Future studies could explore different graph algorithms and data collection techniques. We aim to create a link prediction ML model that will identify key relationships from the MKG using diverse graph algorithms. Furthermore, we plan on developing a graph-based ML model that can differentiate between infectious diseases with similar symptoms. This study enhances research that combines graph theory and artificial intelligence for early detection of infectious diseases.

Author contributions

The conceptualization and design were done by Eman Alqaissi, Fahd Alotaibi and Muhammad Sher Ramzan. The manuscript was initially drafted by Eman Alqaissi and afterward revised by Fahd Alotaibi, Muhammad Sher Ramzan and Abdulmohsen Algarni to ensure intellectual content. All authors of this article have provided approval for the final version to be published and agreed to take accountability for all parts of the work.

Disclosure statement

No potential conflict of interest was reported by the author(s).

Data availability statement

Influenza Dataset is available at Dryad: https://datadryad.org/stash/dataset/doi:10.5061/dryad.t7n48

Hepatitis dataset is available at UC Irvine: https://archive.ics.uci.edu/dataset/46/hepatitis

Code data available at: https://github.com/EMANaLQAISSI/MKG.git

Additional information

Funding

This work was supported by the Deanship of Scientific Research at King Khalid University through Large Groups (Project under Grant Number (RGP.2/549/44)).