Global Shapley Explanations and LIME on HLA-B27 Positivity in Ankylosing Spondylitis Patients

doi:10.21203/rs.3.rs-3556482/v1

Download PDF

Research Article

Global Shapley Explanations and LIME on HLA-B27 Positivity in Ankylosing Spondylitis Patients

https://doi.org/10.21203/rs.3.rs-3556482/v1

This work is licensed under a CC BY 4.0 License

Version 1

posted

You are reading this latest preprint version

Ankylosing spondylitis (AS), an autoimmune disease, has the HLA-B27 gene in more than 90% of its patients. This study investigated the ability of health parameters to predict the presence of the HLA-B-27 gene and clinical and demographic data used in diagnosing AS. For this purpose, various classification models were evaluated, and the best-performing RFC model was selected. In addition, the model's predictions are understood and explained using XAI techniques such as SHAP and LIME. The model development results show that the RFC model performs best (Accuracy:0.75, F1 Score:0.74, Recall:0.75, Precision:0.75, Brier Score:0.25, AUC: 0.76), and XAI techniques provide the ability to explain the decisions of this model. Among the health parameters, WBC, Hematocrit, uric acid, and gender were found to show the strongest association with HLA-B-27. This study aims to understand the genetic predisposition of AS and to illuminate the potential of XAI techniques in medical diagnosis. The study's strengths include comprehensive model evaluation, explainability of model decisions, and revealing the relationship between health parameters and HLA-B-27. In addition, this study considered ethical dimensions like the confidentiality of personal health data and the privacy of patients.

ankylosing spondylitis

Shapley explanations

Local Interpretable Model-agnostic Explanations

prediction

HLA B-27

Ankylosing spondylitis (AS) is a disease that affects the spine and joints, causing fusion of the axial skeleton of the body, and is included in the disease spectrum of axial spondyloarthritis (axSpA) [1]. AS usually occurs in the mid-20s and mostly in men. While it causes significant problems in physical and mental health and a decrease in work efficiency, the social and individual costs of AS are very high [2, 3]. The incidence and prevalence of AS varies from country to country. The incidence of AS is estimated to be between 0.05 and 1.4/10 000 person-years, and its prevalence is estimated between 0.1% and 1.4% [4, 5]. Early treatment of AS increases the likelihood of better clinical outcomes and remission [6]. However, timely diagnosis of AS is difficult because the disease has no noticeable symptoms or signs [6, 7].

For this reason, there can be a long delay between the first symptoms, such as inflammatory low back pain (IBP) and the diagnosis. Hospital-based studies may overlook that the first symptoms, such as IBP, are often first seen in primary care settings and that diagnostic delay studies do not distinguish between AS and other axSpA concepts [8–10]. With the availability of effective treatment, efforts have been made to improve the recognition and diagnosis of AS in the early and non-radiographic stages. Classification criteria published in 2009 have increased sensitivity for earlier diagnosis of AS disease, and MRI has become an essential tool for identifying inflammatory lesions before sacroiliitis can be detected radiographically. However, it is unclear whether this increased knowledge contributes to reducing diagnostic delays [9]. The exact cause of AS is unknown, but genetic factors are thought to play an important role [11]. The most important genetic factor associated with AS is the HLA B-27 antigen. The HLA B-27 antigen is a protein complex that helps the immune system recognise foreign substances [12]. People with the HLA B-27 antigen have a higher risk of developing AS [13–15]. However, the HLA B-27 antigen alone is not the cause of AS. Other genetic and environmental factors also play a role in the development of AS [16]. Early diagnosis of AS is crucial to improve a patient's quality of life and prevent progressive complications. The treatment of AS disease should be planned individually. It should be taken into account that patients may respond differently to different treatment options.

Artificial intelligence (AI) and machine learning (ML) methods can be used to diagnose and treat AS. AI and ML methods can help identify risk factors, symptoms, biomarkers and treatment responses associated with AS disease by analysing large volumes of data [17]. AI and ML methods can be essential in the early diagnosis, prognosis and personalised treatment of AS[18–20].

However, very few studies on AS disease use AI and ML methods. Most of these studies did not use explainable artificial intelligence (XAI) methods. XAI methods are methods used to explain the decisions of AI and ML models [21–23]. XAI methods increase the intelligibility and reliability of AI and ML models by showing how and why they make a particular decision. XAI methods can help both doctors and patients be better informed and achieve better outcomes by explaining the decisions of AI and ML models regarding AS disease [22–27].

For this reason, it is necessary to increase the number and quality of studies on AS using AI and ML methods. It is also essential to explain the decisions of AI and ML models by using XAI methods in these studies. In this way, the potential of XAI and ML methods can be better evaluated in diagnosing and treating AS. The primary purpose of this study was to determine the presence of the HLA-B27 gene associated with AS disease, demographic characteristics and health parameters using XAI methods (Fig. 1). For the study, the hypothesis that the presence of the HLA-B27 gene in AS patients can be predicted by XAI methods using health parameters and demographic characteristics was started. The research question at the centre of this study asks how effective XAI techniques are for indicating the presence of the HLA-B27 gene in individuals with AS based on demographic data and health parameters.

In the literature, there are not enough studies on predicting the HLA-B27 gene associated with AS disease by demographic characteristics and health parameters. Our study offers a new approach to explain this relationship by combining real patient data with XAI techniques. This study can be used as a potential tool in the early diagnosis of the disease and contribute to understanding the relationship between AS and the HLA-B27 gene. This study contributes to the literature on how data analysis and XAI methods can be used in medicine and may help us better understand the relationship between AS disease and the HLA-B27 gene. For this purpose, the model with the best performance was determined among the models created using the AS disease dataset, and the factors that were effective in predicting the HLA-B27 gene were revealed in more detail by applying XAI to this model.

This study aimed to predict the presence of the HLA-B27 gene using the data set containing the parameters of AS disease. In the second part, the related studies of the methods used on other diseases are summarised, and the context in the literature is indicated. In the third chapter, the model development phase on the AS data set is explained in detail, and the model with the best performance is selected. The model's predictions are explained in the same section by applying the XAI techniques SHAP and LIME. The results are presented in the third section, and the estimations and XAI results are detailed. In the fourth chapter, the study's results are discussed with other literature and suggestions for future research are presented. Finally, in the fifth chapter, a general evaluation of the study was made, and the importance of the results was emphasised.

In this study, HLA B-27 estimation was performed based on the data of AS patients. Shapley explanations and LIME methods were used in the estimation process. While Shapley's explanations determine the variables contributing to the model's predictions, the LIME method explains the model's predictions understandably. This study was conducted to understand the relationship between AS disease and HLA B-27 antigen and determine the effect of the variables in predicting the disease. In this section, a connection is made with other studies in the literature by drawing attention to the related studies on different diseases using a similar method.

Glueria et al. developed an explainable artificial intelligence (XAI) framework using classification techniques to predict cardiovascular diseases (CVD). For this purpose, they used various ML algorithms and interpreted the decisions of these algorithms with XAI methods such as SHAP. The authors evaluated the performance of the XAI-based classification models with different evaluation metrics such as curve under field (AUC), receiver operating characteristic (ROC), sensitivity, specificity, and F1-score. The authors stated that the XAI-supported SVM, LR and Naive Bayes models provide an accuracy of 89%, and these models can play an essential role in the early detection of CVD. This study highlights the importance of XAI for health analytics. [28].

Shad et al. investigated the status of patients at different stages of the disease using various artificial neural network models and XAI to predict Alzheimer's disease (AD). The authors applied classification techniques using MRI imaging data to detect AD early. It has been observed that multimodal biological, imaging and other available neuropsychological data can distinguish AD patients from the cognitively normal elderly by a high percentage. They tried classification techniques to detect AD at an early stage using existing artificial neural network (ANN) models. In this process, they used a hybrid dataset from Kaggle to make four separate classifications. They used the LIME-based XAI framework to analyse T1-weighted MRI scans from the hybrid dataset. They used Resnet50, VGG16 and Inception v3 for Convolutional Neural Network Models and obtained categorical accuracy of 82.56%, 86.82% and 82.04%, respectively. This study highlights the importance of XAI in predicting Alzheimer's disease [29].

Kırboğa et al. is an article examining the relationship between troponin and D-Dimer, mortality and CK-MB in COVID-19 patients using XAI. Troponin is a protein that is a marker of heart muscle damage. For this purpose, the authors created a data set that includes troponin, D-Dimer, CK-MB and mortality values by collecting the data of 1000 COVID-19 patients. Then, they analysed this dataset with one of the XAI algorithms, SHAP. Based on their results with SHAP, the authors found that troponin was positively correlated with D-Dimer and mortality but negatively correlated with CK-MB. They also suggested that troponin is an essential marker of heart damage and death risk in COVID-19 patients. This study shows that XAI has an important application area in the health field [30]. Thanks to XAI, we can better understand and explain the changes caused by the disease and its consequences. This may offer us better treatment options.

Aldughayfiq et al. aimed to interpret DL models using XAI to diagnose retinoblastoma. In this study, the authors used XAI methods such as LIME and SHAP to interpret DL models in diagnosing retinoblastoma. For this purpose, the authors used a dataset consisting of 1000 fundus images. The authors analysed the predictions of the models with LIME and SHAP. Their results with LIME and SHAP showed how DL models perceive the clinical features of retinoblastoma and which features are more effective in prediction. For example, in group E, the most severe form of retinoblastoma features such as tumour size, shape, colour, and location was more predictive [31]. This is a study showing that XAI can make an essential contribution to the diagnosis of retinoblastoma. Thanks to XAI, we can better understand and explain the decisions of DL models. This can provide us with a more accurate and reliable diagnosis.

Huang et al. aimed to demonstrate best practices for the explainable discovery of disease markers in ovarian cancer using ML and Shapley analysis. For this purpose, the authors conducted a case study using the CDAS PLCO Ovarian Biomarkers dataset. This dataset includes data from 1000 ovarian cancer patients. The authors trained this dataset using a game-theoretic approach based on classification tasks and Shapley values. The authors used the SHAP algorithm to construct and evaluate the models. The case study results demonstrated the ML pipeline's efficiency, consistency and advantages over traditional statistical approaches. The resulting guidelines provide a general framework for applying XAI in medical research. This framework can inform clinicians and validate and explain ovarian cancer markers [32].

This study used a Dell Gaming G15 5520 model portable computer. Systemic infrastructure, this system with Intel Core i7 12700H processor, 16GB memory and 512GB NVMe SSD storage; It is powered by a 15.6-inch 120Hz FHD display and RTX 3050 Ti graphics. Within the scope of the work carried out, necessary calculations and data processing processes were successfully carried out through the hardware features of the computer in question. This study used various Python libraries for data analysis, model training and interpretation. Pandas [33] and Numpy [34, 35] were used for data processing and manipulation. The Scikit-learn [36] library was used to train and evaluate various ML models such as RandomForestClassifier (RFC), KNeighborsClassifier(KNN), SVC, and DecisionTreeClassifier (DTC). XGBoost was chosen to implement gradient-boosting algorithms and improve model performance. Matplotlib was used to perform graph plotting. Model interpretation and feature importance rankings were achieved with the ELI5 and SHAP libraries. The scikit-plot library was chosen for plotting evaluation plots like ROC curves. The LIME library was used to generate the native annotations. As of the start of the study, the latest versions of all used libraries were used.

3.1. Data Collection and Preprocessing

117 Ankylosing Spondylitis (AS) patient data used in this study were collected from Erzurum City Hospital for approximately 3 months with the ethics committee's approval (ethical decision number: B.30.2.ATA.0.01.00/314). Using the demographic characteristics and health parameters of the data set presented in this study, HLA-B27 estimation was performed by XAI studies. The mean age of the participants was 39.67, and the standard deviation was 12.85. Gender distribution was balanced as 50% male and 50% female. The results obtained regarding the presence of the HLA B-27 gene, a vital feature examined within the scope of the research, are stated. Among the health parameters, the mean white blood cell (WBC) is 309.19 ± 71.33, the mean haemoglobin level (HGB) is 14.34 ± 1.81, and the mean platelet count (PLT) is 299.91 ± 63.53. Data mining and ML studies are disciplines in which complex and diverse data sets are usually carried out. Therefore, the data preprocessing phase forms a fundamental part of these processes. Data preprocessing consists of steps that enable raw data to be analysed and made available. These stages are carried out to increase the data set's quality, improve the model performance and ensure the reliability of the results.

In this study, it was necessary to express categorical features such as gender and positive/negative values numerically. Coding the gender as male and female and positive and negative values as 0 and 1 made the data set more suitable for the model. The data preprocessing phase also includes addressing missing values in the dataset, correcting or considering outliers, feature selection, and scaling. These steps help the model produce more reliable and generalisable results. Data preprocessing is a stage that should be carried out with care and attention to increase the performance of ML models, minimise false predictions, and increase the model's accuracy. Data preprocessing is one of the cornerstones of ML and data mining projects. Correct data preprocessing methods should be determined according to the nature of the data set and the project's objectives. As seen in this study, coding and fitting the data is necessary to ensure the model produces accurate and reliable results.

3.2. Model Development

This study used various classification models to predict HLA-B27 in AS patients. The models used are Random Forest Classifier (RFC) [37–39], XGBoost Classifier (XGB) [40], Decision Tree Classifier (DTC) [41], Support Vector Machine (SVM) [42], Gaussian Naive Bayes [43], Logistic Regression [44] and K Nearest Neighbors Classifier (KNN) [45]. These models were chosen because tree-based models effectively capture complex relationships and provide high accuracy. XGBoost is an algorithm developed to improve the performance of tree-based models, DTC is a simple and straightforward model, and KNN is a distance-based model. The models were trained and tested by dividing the data into 80% training and 20% testing sections. The performance of the models was evaluated using various metrics. Accuracy, F1 score, recall, precision, Brier score and AUC (Area Under the Curve) metrics show how well the models categorise, the rate of false positives or false negatives, how consistent the probability estimates are, and the size of the area under the ROC curve [46] .

3.3. Computational Complexity

Computational complexity is a field of computer science that measures an algorithm's resource demand (time, memory, energy, etc.) of an algorithm or computational problem. This complexity determination is of great importance for the development of efficient algorithms, for drawing computability limits, and classifying the degree of difficulty of computational problems. Computational complexity (Table 1) should be presented in academic reviews for several reasons. In this study, we give the computational complexity,

• Computational complexity provides a standard criterion for evaluating the performance of an algorithm or problem solution. This type of presentation increases the credibility of the study and ensures comparability.

• Computational complexity helps show how practical an algorithm or problem is. Complexity analysis reveals potential applications and limitations of the study.

• Computational complexity is necessary to highlight the theoretical significance of an algorithm or problem. This type of analysis reflects the work's contribution to computability theory and its ability to point out open issues.

Table 1

Computational complexity calculations of model algorithms used in the study.
Model	Training Time Complexity	Test Time Complexity	Ref.
RFC	O(npk)	O(kp)	[47]
XGB	O(npk)	O(kp)	[48]
DTC	O(nplogn)	O(plogn)	[49]
SVM	O(n²p)	O(np)	[50]
GNB	O(np)	O(cp)	[51]
LR	O(np)	O(cp)	[52]
KNN	O(1)	O(nkp)	[53]

[Where n is the number of training data; p is the size of the data; k is the number of trees (for RFC and XGB), the number of neighbours (for KNN), or the number of classes (for GNB and LR); c, number of classes; and logn is the average number of comparisons required to rank the training data.]

3.4. SHAP (SHapley Additive exPlanations)

This study adopted an ML-based approach to predict the HLA-B27 gene, which is critical in diagnosing AS. In this context, it was essential to understand the features in the dataset and the relationship between the HLA-B27 gene. RFC was used as the ML model. However, understanding the internal decisions of complex models such as the RFC and interpreting the model's predictions is often challenging. Therefore, the model's predictions were made more understandable using the SHAP (SHapley Additive exPlanations) method [54–57]. The SHAP method describes how each feature contributes to the model's prediction. Treeshap [58] method was preferred to facilitate the understanding of the internal structure of the RFC. Treeshap calculates the effect of each decision tree on each feature and combines them to produce SHAP values. This study demonstrates how data mining and ML techniques can be used to predict the HLA-B27 gene, which is influential in diagnosing AS. The model's predictions became more understandable thanks to the SHAP method, while the Treeshap method helped illuminate the effects of the RFC's decision trees on the features. This approach is valuable for understanding complex diseases' genetic basis and supporting clinical decision-making processes.

3.5. LIME (Local Interpretable Model-agnostic Explanations)

This study applied an annotation method for the model using the LIME library. LIME is a method used to produce fast and understandable annotations to the model [59, 60]. LimeTabularExplainer, LIME's tabular annotation module, was chosen to help us understand the predictions of classification models. The LimeTabularExplainer class is initialised with parameters such as training data, feature names, class names, and classification modes. Then, it was aimed to reason the predictions of a particular data sample with the explain_instance function. During this process, it was understandably explained how the predictions affected the features and why they turned out the way they did. The explain_instance function was called for each example, and the explanations produced by LIME were visualised with the show_in_notebook function. This method provides insights into the model, allowing us to understand how the model makes decisions and which features contribute more to the prediction results.

4.1. Model Development

Table 2

Performance Comparison of Different Classification Models.
Model	Accuracy	F1 Score	Recall	Precision	Brier Score	AUC
Random Forest Classifier	0.75	0.74	0.75	0.75	0.25	0.76
XGBoost Classifier	0.625	0.63	0.62	0.66	0.375	0.67
Decision Tree Classifier	0.58	0.59	0.58	0.60	0.41	0.57
K Nearest Neighbors Classifier	0.78	0.57	0.58	0.57	0.41	0.53
Support Vector Machine (SVM)	0.625	0.39	0.62	0.48	0.37	0.27
Gaussian Naive Bayes	0.54	0.55	0.54	0.60	0.45	0.70
Logistic Regression	0.68	0.68	0.65	0.72	0.39	0.72

Within the scope of this study, a series of model evaluations were performed to estimate the HLA-B27 values of AS patients using various classification models. The results show that the RFC model performs best when evaluated according to the table. RFC has come to the fore as the preferred and chosen model. This model appears to be preferred because of its high accuracy, F1 score, sensitivity, insight and AUC values, ability to capture complex relationships, and tendency to achieve generally good results. The performance of other models has been evaluated in light of these results, and the prominence of the RFC supports the correctness of the choice made.

ROC (Receiver Operating Characteristic) curves were used to evaluate the performance of the models in Table 2. The ROC curve is a graph that visualises the relationship between the sensitivity and specificity of classification models (Fig. 2). The curve shows how the valid positive rate (sensitivity) changes against the false positive rate (1 - specificity) at different probability thresholds. The area under the ROC curve (AUC) measures the model's classification performance.

According to Table 2, the AUC value of the RFC model is 0.76. This refers to the size of the area under the ROC curve and indicates good overall classification performance. The RFC model has a higher AUC value than other models, meaning that the positive and negative classes are better differentiated, and the model generally makes more accurate predictions.

When looking at the other models, the AUC value of the XGB is 0.67, the AUC of the DTC is 0.57, the AUC of the KNN is 0.53, the AUC of the SVM model is 0.27, the AUC of the GaussianNB model is 0.70, and the AUC of the Logistic Regression model is 0.72. These values help in ranking the classification performance of each model. High AUC values indicate that the model distinguishes positive and negative classes well, and its predictions are generally correct, while low AUC values indicate poorer performance [61]. The RFC model occupies the most significant area under the ROC curve and exhibits better classification performance than other models. Therefore, it can be said that the RFC model was preferred and used in this study.

4.2. SHAP (SHapley Additive exPlanations)

In this study, the SHAP method was applied to explain the decisions of the RFC model used in estimating HLA B-27 value in AS patients. With the SHAP method, the contribution of each feature to the model output was calculated. Bar graph (A) and beeswarm (B) graph were created to visualise SHAP values. The bar graph shows the average absolute SHAP values of the features, expressing their average contribution to the model output. According to the bar graph analysis, the most prominent features were WBC, haematocrit, uric acid and gender (Fig. 3). These features' mean absolute SHAP values were calculated as [values]. These features were found to have similar importance in HLA B-27 positive and negative classification. On the other hand, the Beeswarm graph shows the features' contributions, showing each test sample's SHAP values. According to the Beeswarm plot analysis, the traits with the highest SHAP values were WBC, haematocrit and uric acid. The SHAP values of these traits varied between them. It was observed that these features were most effective in HLA B-27 positive and negative classification.

4.3. LIME (Local Interpretable Model-agnostic Explanations)

The LIME chart contains a bar chart for each attribute (Fig. 4). The length of the bars indicates the severity of the attribute. The colour of the bars indicates the effect of the attribute. Blue bars indicate attributes that increase the probability of prediction, while orange bars indicate attributes that decrease the probability of prediction. As seen in Fig. 3, our model predicted a 91% probability of negative HLA B-27 for this data point. CA, Creatinine, and Age reduced this negative while other features increased this value.

This research aims to explore the ability of health parameters to predict the presence of the HLA-B27 gene, as well as clinical and demographic data used in diagnosing AS. In addition to the model developed for this purpose, using SHAP and LIME applications, which are XAI techniques, is considered an essential step in understanding and explaining the model's predictions. The results aim to present a new approach to understanding the genetic susceptibility to AS and illuminate the potential of XAI techniques in medical diagnostics. The model development results of the study showed that the RFC model performed the best, and the XAI methods could explain the decisions of this model. These results are essential, considering that very few studies on AS disease use AI and ML methods. This study reveals the potential of AI and ML methods in diagnosing and treating AS and increases these methods' transparency, reliability and intelligibility. The strengths of this study are that it comprehensively evaluates the performance of different classification models, explains XAI methods and model decisions, and contributes to understanding the relationship between AS disease and the HLA-B27 antigen. The weaknesses of this study are that it raises questions about the generalizability and validity of the results obtained on only one data set. The performance of classification models may vary depending on the data set's characteristics, algorithm parameters, evaluation metrics and XAI methods.

The areas where this study can guide or inspire future research are to compare the results by testing different classification models and XAI methods on different datasets, to investigate further the effectiveness of AI and ML methods in the diagnosis and treatment of AS disease, and to identify other genetic or environmental factors associated with AS disease. WBC, Hematocrit, uric acid and gender were significant in estimating HLA B-27 in SHAP results. WBC stands for white blood cells and is an important immune system component. HLA B-27 is a genetic trait found on chromosome 6 in humans and plays a role in antigen presentation. The relationship between WBC and the HLA B-27 gene has not yet been fully elucidated. However, some studies have suggested different hypotheses to explain this relationship. One of these hypotheses is that HLA B-27 affects the function or structure of WBCs, causing the immune system to overreact or attack its tissues [61]. According to this hypothesis, in HLA B-27 positive individuals, HLA B-27 molecules on the surface of WBCs recognise some normally harmless bacterial or viral antigens, triggering an inflammatory response. This response leads to inflammation and damage to joints or other tissues [62].

In HLA B-27 positive individuals, abnormal folding or accumulation of HLA B-27 molecules on the surface of WBCs has been observed. This indicates that WBCs are under stress and impair cellular functions [61]. Changes in the number or activity of specific subtypes of WBCs have been detected in positive individuals. For example, the number or function of cells involved in antigen recognition and presentation of WBCs, such as NK, T, and dendritic cells, differ in HLA B-27 positive individuals [63]. In these individuals, changes in the level or profile of cytokines secreted by WBCs were detected. Cytokines are molecules that allow the immune system to communicate. In addition, while the level of inflammatory cytokines increases, the level of anti-inflammatory cytokines decreases. This indicates that the balance of the immune system is disturbed. Hwang et al. reported that in the method they used to reduce the risk of developing acute graft-versus-host disease (GVHD) in patients undergoing stem cell transplantation, HLA-A, HLA-B and HLA-DR compatibility was achieved between the recipient and the donor. This study supports our results, emphasising that the relationship between WBC and HLA B-27 is vital for adapting the immune system [64]. Ramsbottom et al., on the other hand, associated the relationship between WBC and HLA B-27 with agranulocytosis caused by some drugs used in the treatment of thyroid diseases. Agranulocytosis is a condition in which the number of WBCs is abnormally reduced. In this study, the authors examined the genetic profiles of patients who developed agranulocytosis and found that different HLA-B alleles were associated with this condition. This study demonstrates that the relationship between WBC and HLA B-27 influences susceptibility to drug side effects [65]. In light of these findings, it can be said that the relationship between WBC and HLA B-27 gene plays an essential role in the pathogenesis of AS. However, the exact mechanism of this relationship is not yet clear. Therefore, more research is needed on this subject.

Hematocrit is the ratio of the red blood cell portion of the blood to the total blood volume. The haematocrit value indicates the oxygen-carrying capacity of the blood, blood diseases such as anaemia or polycythemia, and the body's hydration status. The haematocrit value may vary depending on gender, age, race, height, nutrition and genetic factors [66].

Aboud et al. investigated the effects of biological and non-biological treatment on haematological indices in patients with AS and psoriatic arthritis (PsA). Erythrocyte distribution width was found to be lower. As a result of this study, it can be said that biological therapy positively affects haematological indices in patients with AS and PsA. This situation is thought to be explained by the strong anti-inflammatory effect of biological therapy [67, 68]. This study reveals that haematological indices are important in diagnosing and treating AS and PsA. About uric acid, Zhang et al. investigated the effect of HLA-B27 status on kidney function in patients with AS and secondary IgA nephropathy (IgAN). They have reported [69]. Lee et al., on the other hand, supported the statistically significant difference in uric acid levels between AS patients with and without proteinuria. While the mean uric acid level was 6.8 mg/dL in patients with proteinuria, it was 5.9 mg/dL in patients without proteinuria. (p = 0.001) and while the mean uric acid level was 6.2 mg/dL in patients with hematuria/AS, it was 6.0 mg/dL (p = 0.624) in patients without hematuria. This indicates that uric acid levels are associated with proteinuria in AS patients but not with hematuria, decreased renal function, or renal pathology. Therefore, uric acid levels have shown that it can be used as a marker to predict renal involvement in AS patients [70]. Therefore, the absence of additional diseases in our data affected our study's supporting or not supporting these results. More accurate analyses can be made on real patient data with AS and other known diseases.

Xiong et al. reviewed the medical records of 386 patients with AS admitted to the Rheumatology Department of the First Affiliated Hospital of Wenzhou Medical University between January 2007 and January 2013. They used a two-way classification analysis of variance (ANOVA) to examine the effect of HLA-B27 status and gender on age at disease onset. The ANOVA test showed that HLA-B27-positive and male patients had the disease earlier than HLA-B27-negative and female patients, respectively [71]. Similar to the location of our data, 100 acute anterior uveitis (AAU) patients (50 HLA-B27-positive, 50 HLA-B27-negative) admitted to Istanbul University Eye Hospital between 2000–2004 were retrospectively analysed. The results showed no significant difference between HLA-B27-positive and negative patients regarding AAU complications, treatment and visual acuity [72]. There was no apparent effect on the absence of AS but the effect of gender on the gene. Therefore, more studies are needed on the impact of age on AS and HLA B-27.

The LIME results show which health parameters our model attaches more or less importance to in predicting the presence of the HLA B-27 gene. In this way, it allows us to understand the logic of our model and the decision process. The bar graph shows the probabilities of our model predicting whether the HLA B-27 gene is negative or positive. In our study, our model predicted the HLA B-27 gene with a 91% probability of being negative and a probability of 9% of being positive. The colours in the bar graph show the effect of health parameters on the forecast. Blue bars indicate parameters that increase the probability of prediction; the orange bars show the parameters that reduce the prediction probability. The attribute value table shows the values of the health parameters of each sample. In our study, the value of gender indicates 0.00 and male. Gender is a parameter that reduces the probability of the HLA B-27 gene being negative. The value of CA is 8.80, which increases the probability of the HLA B-27 gene being negative. The creatinine value is 0.63, a parameter that increases the probability of the HLA B-27 gene being negative. The value of PLT is 288, a parameter that reduces the probability of the HLA B-27 gene being negative. Korniluk et al. [73] compared 50 HLA-B27-positive AS patients with 50 HLA-B27-negative healthy controls. It has been reported that MPV is significantly higher in the serum of AS patients with a positive HLA-B27 gene. This result indicates that there is platelet activation and inflammation in AS patients. The article's authors suggested that MPV is an inflammatory marker that can be used in diagnosing and treating AS. Ünal et al., on the other hand, stated that MPV was significantly higher in the serum of AS patients with positive HLA-B27 gene, supporting our results [74].

Within the scope of research, XAI is a concept used to understand and explain the decisions of AI models. XAI aims to increase AI models' transparency, accountability and reliability, especially those that make high-risk and complex decisions. The importance of using XAI in healthcare has great potential, especially in making medical decisions that can directly affect patients' health become more understandable and traceable. The advantages of using XAI in the health field, especially with techniques such as SHAP and LIME, become apparent at this point. Thanks to these methods, it can be understood which features an AI model gives more or less weight to so that the model's logic can be better understood. SHAP and LIME can also be used to evaluate and improve model performance. By revealing the model's strengths and weaknesses, errors and shortcomings, the accuracy and reliability of the model can be increased. However, besides such advantages, it should not be forgotten that using XAI in health brings ethical problems.

In particular, the ethical dimensions of using XAI in the healthcare field include the confidentiality of personal health data and the privacy of patients [17, 75, 76]. XAI techniques can lead to privacy breaches when disclosing patients' health information, and sensitive data needs to be protected. In addition, it must be ensured that the explanations are accurate and reliable. Otherwise, misleading results may be obtained. These ethical problems highlight the limitations of XAI use in the healthcare industry and areas that require attention [77]. While using XAI techniques throughout our research, we paid great attention to the privacy and protection of personal data. In this context, we considered obtaining the patient's consent an essential requirement and actively applied encryption, anonymisation and masking methods to process data securely. While working on the datasets, we made sure that the explanations met the criteria of consistency, comparability and reliability by comparing the explanations with different models and scenarios. However, we have also clearly addressed liability issues. We have developed mechanisms to prevent or compensate for possible errors, with a detailed allocation of responsibility to determine who is responsible for the results of the forecasts. In this way, we emphasised that our work complies with ethical standards and that using XAI techniques in the health field is carried out responsibly. Thus, this study discussed the advantages and ethical problems of using XAI techniques, SHAP and LIME, in the health field. Considering the ethical issues and the benefits of XAI in the health field will ensure that research and practices are more responsible and sustainable.

This study aimed to explore the ability of health parameters to predict the presence of the HLA-B27 gene, as well as clinical and demographic data used to diagnose ankylosing spondylitis (AS). For this purpose, various classification models were evaluated, and the best-performing RFC model was selected. In addition, the model's predictions are understood and explained using XAI techniques such as SHAP and LIME. This study aims to understand the genetic predisposition of AS and to illuminate the potential of XAI techniques in medical diagnosis. The study's strengths include comprehensive model evaluation, explainability of model decisions, and revealing the relationship between health parameters and HLA-B27. The weaknesses of the study are the generalizability of the results obtained on only one data set and the limitations of the XAI methods. In addition, this study considered ethical dimensions such as the confidentiality of personal health data and the privacy of patients. This study contributes to the literature by emphasising the effectiveness of AI and ML methods in diagnosing and treating AS. Future studies should compare results with different datasets, models and XAI methods and focus on identifying other genetic or environmental factors associated with AS disease.

Availability of data and materials

The datasets used and analysed during the current study can be available from the corresponding author at reasonable request.

Funding statement

This work was self-funded by the first and corresponding author.

Declaration of Competing Interest

The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.

Acknowledgement

We are grateful to Turkey Erzurum Training and Research Hospital, which provided Ankilozan Spondilit data with the ethics committee's approval (ethical decision number: B.30.2.ATA.0.01.00/314) within the scope of this study.

Zhu, W., et al., Ankylosing spondylitis: etiology, pathogenesis, and treatments. Bone Res, 2019. 7: p. 22.
Martindale, J., R. Shukla, and J. Goodacre, The impact of ankylosing spondylitis/axial spondyloarthritis on work productivity. Best Practice & Research Clinical Rheumatology, 2015. 29(3): p. 512-523.
Park, J.-S., et al., Impact of ankylosing spondylitis on depression: a nationwide cohort study. Scientific reports, 2019. 9(1): p. 6736.
Braun, J. and J. Sieper, Ankylosing spondylitis. The Lancet, 2007. 369(9570): p. 1379-1390.
McLeod, C., et al., Adalimumab, etanercept and infliximab for the treatment of ankylosing spondylitis: a systematic review and economic evaluation. Health technology assessment (Winchester, England), 2007. 11(28): p. 1-158, iii.
van den Berg, R., et al., ASAS modification of the Berlin algorithm for diagnosing axial spondyloarthritis: results from the SPondyloArthritis Caught Early (SPACE)-cohort and from the Assessment of SpondyloArthritis international Society (ASAS)-cohort. Annals of the rheumatic diseases, 2013. 72(10): p. 1646-1653.
Linden, S.V.D., H.A. Valkenburg, and A. Cats, Evaluation of diagnostic criteria for ankylosing spondylitis. Arthritis & Rheumatism, 1984. 27(4): p. 361-368.
Keat, A., et al., BSR guidelines for prescribing TNF-α blockers in adults with ankylosing spondylitis. Report of a working party of the British Society for Rheumatology. Rheumatology, 2005. 44(7): p. 939-947.
Rudwaleit, M., et al., The development of Assessment of SpondyloArthritis international Society classification criteria for axial spondyloarthritis (part II): validation and final selection. Annals of the rheumatic diseases, 2009. 68(6): p. 777-783.
Elm, E.v., The Strengthening the Reporting of Observational Studies in Epidemiology (STROBE) statement: guidelines for reporting observational studies. Ann Inter Med, 2007. 147: p. 573-577.
Li, Z., et al., Genotype by sex interactions in ankylosing spondylitis. Nature Genetics, 2023. 55(1): p. 14-16.
Parameswaran P, L.M., HLA-B27 Syndromes. 2023: StatPearls Publishing.
Khan, M., et al., HLA B27 in ankylosing spondylitis: differences in frequency and relative risk in American Blacks and Caucasians. The Journal of Rheumatology, 2023. 50(1): p. 39-43.
Bezuglov, E., et al., HLA B27-positive ankylosing spondylitis professional soccer player with a successful return to sports. The Egyptian Rheumatologist, 2023. 45(4): p. 267-272.
Lin, A., et al., Development of ankylosing spondylitis in patients with ulcerative colitis: A systematic meta-analysis. Plos one, 2023. 18(8): p. e0289021.
Zhou, S.-Z., et al., Exploring the common diagnostic gene KCNJ15 and shared pathway of ankylosing spondylitis and ulcerative colitis through integrated bioinformatics. Frontiers in Physiology, 2023. 14: p. 624.
Eurlings, C.G.M.J., et al., Use of artificial intelligence to assess the risk of coronary artery disease without additional (non-invasive) testing: validation in a low-risk to intermediate-risk outpatient clinic cohort. BMJ Open, 2022. 12(9): p. e055170.
Vadapalli, S., et al., Artificial intelligence and machine learning approaches using gene expression and variant data for personalized medicine. Briefings in Bioinformatics, 2022. 23(5).
Blasiak, A., J. Khong, and T. Kee, CURATE. AI: optimizing personalized medicine with artificial intelligence. SLAS TECHNOLOGY: Translating Life Sciences Innovation, 2020. 25(2): p. 95-105.
Eskofier, B.M. and J. Klucken, Predictive Models for Health Deterioration: Understanding Disease Pathways for Personalized Medicine. Annual Review of Biomedical Engineering, 2023. 25: p. 131-156.
Palacio, S., et al., XAI Handbook: Towards a Unified Framework for Explainable AI. 2021.
Holzinger, A., et al., Explainable AI Methods - A Brief Overview. 2022, Springer International Publishing. p. 13-38.
Manresa-Yee, C., et al., Advances in XAI: Explanation Interfaces in Healthcare, in Handbook of Artificial Intelligence in Healthcare: Vol 2: Practicalities and Prospects, C.-P. Lim, et al., Editors. 2022, Springer International Publishing: Cham. p. 357-369.
Heberle, H., et al., XSMILES: interactive visualization for molecules, SMILES and XAI attribution scores. Journal of Cheminformatics, 2023. 15.
Belle, V. and I. Papantonis, Principles and Practice of Explainable Machine Learning. Frontiers in Big Data, 2021. 4.
Das, A. and P. Rad, Opportunities and Challenges in Explainable Artificial Intelligence (XAI): A Survey. 2020.
Gunning, D., Explainable artificial intelligence (xai). Defense advanced research projects agency (DARPA), nd Web, 2017. 2(2): p. 1.
Guleria, P., et al. XAI Framework for Cardiovascular Disease Prediction Using Classification Techniques. Electronics, 2022. 11, DOI: 10.3390/electronics11244086.
Shad, H.A., et al. Exploring Alzheimer's Disease Prediction with XAI in various Neural Network Models. in TENCON 2021 - 2021 IEEE Region 10 Conference (TENCON). 2021.
Kırboğa, K.K., et al., CVD22: Explainable Artificial Intelligence Determination of the Relationship of Troponin to D-Dimer, Mortality, and CK-MB in COVID-19 Patients. Computer Methods and Programs in Biomedicine, 2023: p. 107492.
Aldughayfiq, B., et al. Explainable AI for Retinoblastoma Diagnosis: Interpreting Deep Learning Models with LIME and SHAP. Diagnostics, 2023. 13, DOI: 10.3390/diagnostics13111932.
Huang, W., et al., Explainable discovery of disease biomarkers: The case of ovarian cancer to illustrate the best practice in machine learning and Shapley analysis. J Biomed Inform, 2023. 141: p. 104365.
McKinney, W., pandas: a Foundational Python Library for Data Analysis and Statistics. Python High Performance Science Computer, 2011.
Harris, C.R., et al., Array programming with NumPy. Nature, 2020. 585(7825): p. 357-362.
van der Walt, S., S. Colbert, and G. Varoquaux, The NumPy Array: A Structure for Efficient Numerical Computation. Computing in Science & Engineering, 2011. 13: p. 22-30.
Pedregosa, F., et al., Scikit-learn: Machine Learning in Python. J. Mach. Learn. Res., 2011. 12(null): p. 2825–2830.
Hung, T.N.K., et al., An AI-based Prediction Model for Drug-drug Interactions in Osteoporosis and Paget's Diseases from SMILES. Mol Inform, 2022. 41(6): p. e2100264.
Heo, J., et al., Machine Learning-Based Model for Prediction of Outcomes in Acute Stroke. Stroke, 2019. 50(5): p. 1263-1265.
Breiman, L., Random forests. Machine learning, 2001. 45(1): p. 5-32.
Chen, T. and C. Guestrin, XGBoost: A Scalable Tree Boosting System. 2016. 785-794.
Song, Y.Y. and Y. Lu, Decision tree methods: applications for classification and prediction. Shanghai Arch Psychiatry, 2015. 27(2): p. 130-5.
Chang, C.-C. and C.-J. Lin, LIBSVM: a library for support vector machines. ACM transactions on intelligent systems and technology (TIST), 2011. 2(3): p. 1-27.
Bustamante, C., L. Garrido, and R. Soto, Comparing Fuzzy Naive Bayes and Gaussian Naive Bayes for Decision Making in RoboCup 3D. 2006. 237-247.
Sperandei, S., Understanding logistic regression analysis. Biochem Med (Zagreb), 2014. 24(1): p. 12-8.
Zhang, Z., Introduction to machine learning: k-nearest neighbors. Ann Transl Med, 2016. 4(11): p. 218.
scikit-learn. 3.3. Metrics and scoring: quantifying the quality of predictions. 2007-2013; Available from: https://scikit-learn.org/stable/modules/model_evaluation.html.
Kumar, P. Computational Complexity of ML Models. 2019; Available from: https://medium.com/analytics-vidhya/time-complexity-of-ml-models-4ec39fad2770.
Hu, X., et al., Model Complexity of Deep Learning: A Survey. 2021.
Virgolin, M. Time complexity for different machine learning algorithms. 2021; Available from: https://marcovirgolin.github.io/extras/details_time_complexity_machine_learning_algorithms/.
Stackoverflow. How to compute the complexity of machine learning models. 2019; Available from: https://stackoverflow.com/questions/53384906/how-to-compute-the-complexity-of-machine-learning-models.
Pérez, A., P. Larrañaga, and I. Inza, Supervised classification with conditional Gaussian networks: Increasing the structure complexity from naive Bayes. International Journal of Approximate Reasoning, 2006. 43(1): p. 1-25.
Braverman, M., et al., The gradient complexity of linear regression. 2019.
Adamczyk, J., k nearest neighbors computational complexity. 2020.
Roth, A.E., Introduction to the Shapley value. The Shapley value, 1988: p. 1-27.
Winter, E., The shapley value. Handbook of game theory with economic applications, 2002. 3: p. 2025-2054.
Derks, J. and H. Peters, A Shapley Value for Games with Restricted Coalitions. International Journal of Game Theory, 1993. 21(4): p. 351-60.
Shapley, L., A Value for n-Person Games. Princeton University Press, Princeton, 1953: p. 307-317.
Mitchell, R., E. Frank, and G. Holmes, GPUTreeShap: massively parallel exact calculation of SHAP scores for tree ensembles. PeerJ Comput Sci, 2022. 8: p. e880.
Kırboğa, K.K., S. Abbasi, and E.U. Küçüksille, Explainability and white box in drug discovery. Chemical Biology & Drug Design. n/a(n/a).
Palatnik de Sousa, I., M. Maria Bernardes Rebuzzi Vellasco, and E. Costa da Silva, Local Interpretable Model-Agnostic Explanations for Classification of Lymph Node Metastases. Sensors, 2019. 19(13): p. 2969.
Crux, N.B. and S. Elahi, Human Leukocyte Antigen (HLA) and Immune Regulation: How Do Classical and Non-Classical HLA Alleles Modulate Immune Response to Human Immunodeficiency Virus and Hepatitis C Virus Infections? Frontiers in Immunology, 2017. 8.
Khan, M.A., An Update on the Genetic Polymorphism of HLA-B*27 With 213 Alleles Encompassing 160 Subtypes (and Still Counting). Current Rheumatology Reports, 2017. 19(2): p. 9.
Kavadichanda, C.G., et al., Spondyloarthritis and the Human Leukocyte Antigen (HLA)-B(*)27 Connection. Front Immunol, 2021. 12: p. 601518.
Hwang, M.C., L. Ridley, and J.D. Reveille, Ankylosing spondylitis risk factors: a systematic literature review. Clinical Rheumatology, 2021. 40(8): p. 3079-3093.
Ramsbottom, K.A., et al., Informatics investigations into anti-thyroid drug induced agranulocytosis associated with multiple HLA-B alleles. PLOS ONE, 2020. 15(2): p. e0220754.
Mondal H, L.S., Hematocrit. 2023: StatPearls Publishing.
Aboud, F.M., et al., Impact of biological and non-biological treatment on hematological indices in patients with ankylosing spondylitis and psoriatic arthritis. Egyptian Rheumatology and Rehabilitation, 2023. 50(1): p. 14.
Bengtsson, K., et al., Are ankylosing spondylitis, psoriatic arthritis and undifferentiated spondyloarthritis associated with an increased risk of cardiovascular events? A prospective nationwide population-based cohort study. Arthritis Research & Therapy, 2017. 19(1): p. 102.
Zhang, T., et al., HLA-B27 Negativity Is Associated With Renal Function Decline in Patients With Ankylosing Spondylitis and Secondary IgA Nephropathy. Frontiers in Medicine, 2020. 7.
Lee, S.-H., et al., Renal involvement in ankylosing spondylitis: prevalence, pathology, response to TNF-a blocker. Rheumatology International, 2013. 33(7): p. 1689-1692.
Xiong, J., et al., Association of HLA-B27 status and gender with sacroiliitis in patients with ankylosing spondylitis. Pak J Med Sci, 2014. 30(1): p. 22-7.
Tuncer, S., et al., Clinical features and outcomes of HLA-B27-positive and HLA-B27-negative acute anterior uveitis in a Turkish patient population. Ocular Immunology and Inflammation, 2005. 13(5): p. 367-373.
Korniluk, A., et al., Mean Platelet Volume (MPV): New Perspectives for an Old Marker in the Course and Prognosis of Inflammatory Conditions. Mediators of Inflammation, 2019. 2019: p. 9213074.
Ünal, K., et al., THE RELATIONSHIP OF MEAN PLATELET VOLUME LEVELS WITH MEFV AND HLA-B27 GENE ANALYSIS. 2016. 8: p. 2016.
Okolo, C.T., Towards a Praxis for Intercultural Ethics in Explainable AI. arXiv pre-print server, 2023.
Izumo, T. and Y.-H. Weng, Coarse ethics: how to ethically assess explainable artificial intelligence. AI and Ethics, 2022. 2(3): p. 449-461.
Ahmed, A., AI Ethics and Practicing XAI: Explainable Artificial Intelligence. 2021.

No competing interests reported.

Download PDF

Version 1

posted

You are reading this latest preprint version

Global Shapley Explanations and LIME on HLA-B27 Positivity in Ankylosing Spondylitis Patients

Status:

Version 1

Abstract

Figures

1. Introduction

2. Related Works

3. Experiment

3.1. Data Collection and Preprocessing

3.2. Model Development

3.3. Computational Complexity

3.4. SHAP (SHapley Additive exPlanations)

3.5. LIME (Local Interpretable Model-agnostic Explanations)

4. Results

4.1. Model Development

4.2. SHAP (SHapley Additive exPlanations)

4.3. LIME (Local Interpretable Model-agnostic Explanations)

5. Discussion

Conclusion

Declarations

References

Additional Declarations

Status:

Version 1