Towards XAI agnostic explainability to assess differential diagnosis for Meningitis diseases

CDSSs are intelligent systems that assist medical professionals in facilitating decision-making at different stages of the diagnosis and treatment of diseases using specific recommendations [10].

CDSSs are classified into knowledge-based or non-knowledge-based, with the latter leveraging ML and AI or statistical pattern recognition. However, non-knowledge-based systems face challenges, such as understanding AI's logic (black boxes) and obtaining high-quality data due to fragmentation, inconsistent formats, and privacy concerns [11]. Several studies have been proven effective in diagnosing different pathologies.

Regarding Meningitis diagnosis, the work presented by D'Angelo et al [12] aims to improve the discrimination between bacterial and viral Meningitis etiologies through the use of ML-based methodologies. Two cases were considered: one in which both blood and cerebrospinal parameters were taken into account and another in which only blood data were used. The results showed that a combination of clinical parameters is necessary to distinguish between the two Meningitis etiologies properly. The study used four classifiers: Naive Bayes, multilayer perceptron (MLP), Decision tree-J48, and genetic programming (GP). The GP classifier achieved the best performance. It obtained 100%

Zaccari and Marujo [13] focused on developing a quantitative measure to help healthcare professionals decide whether or not patients need to undergo a cerebrospinal fluid (CSF) exam to diagnose Meningitis. Their approach involves using ML techniques to analyze data from blood and urine exams and patient chief complaint reports to identify patterns that could indicate the presence of Meningitis. The study used seven classifiers: Adaptive Boosting (AdaBoost), Decision Tree, Gradient Boosting, K-nearest neighbors, Logistic Regression, Random Forest and Support Vector Machines. Their analysis found that the Decision Tree model had the best performance, with an accuracy of 96.18%

Authors in [14] aimed to develop a system to classify subjects with Meningitis using a feedforward artificial neural network (ANN). They employed two learning algorithms to develop the ANN: Levenberg-Marquardt training algorithm, suitable for pattern recognition, and particle swarm optimization to adjust the decision threshold. The goal was to achieve better performance by optimizing the decision threshold using a database that included several parameters such as temperature, CSF/blood glucose ratio, proteins, CSF leukocytes, glucose, lactates, erythrocyte sedimentation rate (ESR), and C-reactive protein (CRP).

This study [15] aimed to identify the best classification model to assist in diagnosing Meningitis. The researchers examined the performance of seven classification techniques applied to nine clinical symptoms of a patient, as well as their age, sex, and geographical location. They found that all models could predict Meningitis even before the completion of laboratory tests, indicating the possibility of a non-invasive and early diagnosis. The best classification technique was determined to be the J48 decision tree.

The researchers expanded their research on Meningitis [16] by creating a computerized decision support system (CDSS) that can help doctors identify the illness. They developed three decision models: DM1 determines if a patient has Meningitis based on observable symptoms, DM2 predicts the probability of meningococcal Meningitis using the same symptoms, and DM3 explores the disease's cause using chemical and cytological test data. The decision models achieved a high classification accuracy of 94.3% for Meningococcal Disease Meningitis. Evaluation of the system with real patient data showed that diagnosing Meningitis based solely on observable symptoms is challenging, but the CDSS correctly diagnosed 88% of Meningitis cases from the database.

The same researchers [17] explored data-driven techniques to differentiate between viral and bacterial Meningitis using a dataset of 26 228 patients and 19 attributes. They experimented with various sampling, feature selection, and classification models, finding that combining ensemble methods with decision trees achieved the best performance. The best classifiers had precision, recall, and f-measure of 89% and an AUC value of 95%. Their results suggest that this approach outperforms previous work using only decision trees.

Despite the promising results of previous studies in accurately predicting diagnosis, the black-box nature of these models poses a challenge for their adoption in clinical settings [18], as it can be challenging to comprehend the reasoning behind the model's predictions. This transparency is essential as it involves both the acknowledgment of AI usage and understanding how AI arrives at its conclusions or classifications [19]. Ensuring transparency in AI usage is essential, and it involves both the acknowledgment of AI usage and the understanding of how AI arrives at its conclusions [19]. Applying rigorous controls and testing from the medical field to AI deployment in healthcare reinforces this transparency, providing clear explanations of AI decision-making to ensure safety, accountability, and responsible use in medical settings [20].

Choi et al [5] conducted a study on meningitis and encephalitis classification in patients hospitalized within the initial 24 h. Various ML models were applied, including XGBoost, Random Forest, Light Gradient Boosting Machine, KNN, Gaussian Naive Bayes, and TabNet. An ensemble model (80% XGBoost, 20% TabNet) achieved the highest performance, with accuracy, precision, recall, and F1 score of 0.89 and AUROC of 0.91. Classifiers were applied to baseline characteristics, medical history, vital signs, and diagnostic results (CT, CXR, EEG). Laboratory findings from CSF, blood, and urine were also considered. Model-agnostic techniques (PIMP, LIME, SHAP) provided explainability. AI models slightly outperformed human clinicians due to the absence of certain factors considered in actual clinical practice. However, the researchers highlighted that AI still performed very well, suggesting it could be helpful for neurologists in making quick treatment decisions.

Yang et al [21] conducted an insightful retrospective study on febrile infants aged $\unicode{x2A7D}$60 days, which involved using a deep neural network to develop a predictive model of invasive bacterial infection (IBI). The model's performance was then compared to that of the IBI score. The SHAP technique was used to explain the model's predictions at different levels. Five influential predictive variables (absolute neutrophil count, body temperature, heart rate, age, and C-reactive protein) were identified using SHAP. The study resulted in developing an explainable deep learning model that performs better than previous scoring systems and provides insight into how it arrives at its predictions through individual features and cases.

This study [22] aimed to predict the severity of COVID-19 by using ML and Deep Learning algorithms that consider various clinical markers, vital signs, and critical factors. The researchers evaluated five data-balancing techniques and twelve classifiers to find the most effective method. They discovered that Random Forest trained on Borderline SMOTE balanced data was the best-performing method, achieving an 83% recall rate in predicting COVID-19 severity. To better understand the models, the team deployed Explainable AI tools such as SHAP, Local Interpretable model-agnostic explanations (LIME), ELI5, Qlattice, Anchor, and Feature Importance to determine the importance of critical features in predicting COVID-19 severity. Their findings showed that respiratory rate, blood pressure, lactate, and calcium values were the primary contributors to the increase in severity of a COVID-19 patient. Ultimately, this architecture aims to serve as an explainable decision-support triaging system for medical professionals in countries lacking advanced medical technology and infrastructure to reduce COVID-19 fatalities.

Laatifi et al [23] investigates the impact of cytokines on the severity of SARS-CoV-2 infection. Plasma levels of 48 cytokines were measured in 87 participants from the COVID-19 study. Five models (Random Forest, XGBoost, Bagging Classifier, Decision Tree, and Gradient Boosting Classifier) were trained on synthetic data, with the Gradient Boosting Classifier showing superior performance. The interpretations of the Gradient Boosting model by SHAP and the LIME (Local Interpretable Model-agnostic Explanations) provide detailed insights into the cytokine dataset. The results revealed significant variations in cytokine levels among COVID-19-infected patients, with VEGF-A, MIP-1b, and IL-17A showing elevated levels in severe cases. At the same time, M-CSF, IL-27, IL-9, IL-12p40, RANTES, and TNF were associated with non-severe cases and healthy individuals. These findings suggest the involvement of these cytokines in disease promotion and offer new possibilities for prevention and treatment.

In contrast, Mercaldo et al [24] adopted a different approach by utilizing medical images to detect coronavirus disease. They introduced a deep learning method that categorized computed tomography (CT) images into healthy patients, patients with pulmonary disease, and patients affected by Coronavirus 19. To provide explainability in their model, they employed the Gradient-weighted class activation mapping (Grad-CAM) algorithm, which automatically highlighted the symptomatic areas of infection within CT images. This technique enhanced the diagnostic process by offering visual insights into the regions contributing to disease detection. Integrating Grad-CAM with deep learning improved the efficiency and accuracy of disease detection, providing valuable information for medical professionals. Shi et al [25] employed ML techniques to diagnose tuberculous Meningitis, offering a potential solution to enhance diagnostic accuracy.

Building on previous works, our current work determines which signs or indicators have the highest predictive value by analyzing the Meningitis disease, providing valuable insights for accurate disease diagnosis. Our method derives the underlying factors influencing the IA-based algorithms' outcomes, allowing clinicians to trust and effectively incorporate AI-based recommendations into their clinical practice of Meningitis.

We conducted the experiments on a total of 6729 notified Meningitis cases of individuals over 18, retrieved from SINAN [30], the Information System on Notifiable Diseases of the Brazilian Government's Health Department from 2003 to 2022. SINAN is a database that encompasses compulsory disease notifications throughout Brazil. Notifications and investigations are stored in the SINAN NET database, with dedicated tables for each specific disease, including meningitis.

Initially, the dataset consists of 123 attributes. In our work, we focused on 34 specific attributes that are centered around clinical signs and biological examinations. These attributes were carefully chosen, in consultation with the infectious disease experts at Setif Hospital in Algeria, as they are the most informative and relevant for Meningitis diagnosis. Tables 1 and 2 summarizes the description and the possible values of the selected attributes from the dataset.

We collected a dataset from Setif's Hospital in Algeria comprising cases notified as Meningitis. Notable disparities were observed in comparison to the dataset sourced from the Brazilian SINAN database. This dataset diversification aims to enhance our samples for evaluations, offering a broader perspective on the performance and generalizability of our explainable AI model. Within Setif's hospital dataset, we identified cases suitable for testing our model. However, these cases only encompass two types of Meningitis, constituting a subset of the classes found in the training dataset of the SINAN database.

In this phase, a series of data preparation steps have been performed on the data set, including addressing duplicate data and missing values and handling outliers. Additionally, a crucial step involved balancing the data to mitigate any biases and enhance the overall learning of the model. These rigorous data preparation steps have been undertaken to minimize distortions and improve the accuracy of predictions:

• The data type transformation: We converted data type on selected features to ensure compatibility with the model's requirements and enhance prediction accuracy.
• The missing Data: To preserve the integrity of the analysis and considering the substantial volume of data available, all observations with missing values were omitted from the analysis.
• The structural errors: Fixing structural errors in the preprocessing step, including eliminating typos and inaccurate information, ensures consistent features. This step enhances accuracy, reduces biases, and improves subsequent analysis and modeling.
• The categorical features encoding: We identified several categorical attributes pre-encoded from the original data in the dataset. We selectively converted the necessary attributes into integer form for model compatibility and prediction generation, enabling their effective utilization within our analysis. It is worth noting that certain ML algorithms, such as decision trees and rule-induction methods (e.g. CART, C5.0, etc), possess inherent capabilities to handle high-cardinality categorical attributes without the need for external preprocessing steps [31].
• The outliers: In medical research, outliers hold valuable insights into rare or atypical cases, contributing to a comprehensive analysis. These extreme values may or may not represent aspects of data intrinsic variability and may have a legitimate place in the dataset [32]. We applied the Interquartile Range (IQR) method to handle outliers specifically to numerical attributes, including laboratory biomarkers. The IQR method is a statistical technique employed to detect and manage outliers within the dataset. We systematically deal with outliers by utilizing the IQR method, leading to a more comprehensive data variability. Following this process, the dataset size was reduced to 5072 samples.
• The data balancing: we employed the SMOTE-NC (Synthetic Minority Over-sampling Technique for Nominal and Continuous) method to deal with the class imbalance. SMOTE-NC extends SMOTE to handle categorical and numerical features, generating synthetic samples. Choosing SMOTE-NC over SMOTE allows us to balance the dataset while maintaining feature integrity, which ensures fair representation and reduces bias towards the majority class [33]. Figure 2 shows the class distribution pre-application and post-application of SMOTE-NC.

Zoom In Zoom Out Reset image size

In this section, we justify the choice of three tree-based decision models: the decision tree classifier, random forest classifier, and XGBoost classifier. Among the several attractive properties of tree-based methods is their ability to capture complex interactions between predictors [34]. Moreover, they are considered less prone to outliers, require no distributional assumptions or data transformations, and are intuitive [35].

Various metrics assess distinct characteristics of the classifier generated by the classification algorithm [36]. This study utilized a diverse set of performance metrics to evaluate the developed classifiers. The chosen metrics encompass accuracy, recall, precision, F1-score, and area under the receiver operating characteristic curve (AUROC). The dataset was divided into two subsets to ensure an appropriate evaluation of the classifiers' performance: 70% for training the models and 30% for testing their generalization capabilities. This division allowed us to train the classifiers on a significant portion of the data effectively while reserving a separate portion for independent assessment, serving as a benchmark to evaluate the classifiers' predictive accuracy on unseen instances.

The SHAP method is a model-agnostic interpretation technique used to interpret the results of a predictive model. It employs the Shapley value, a cooperative game theory concept, to quantify each feature's contribution to the model's prediction. The Shapley value formula ($\phi_j(val)$) considers all possible feature combinations and calculates the marginal contribution of each feature. It satisfies desirable properties like efficiency, symmetry, dummy, and additivity.

The formula for the Shapley value ($\phi_j(val)$) is:

$$\begin{align} \phi_j\left(val\right) = \sum_{{S \subseteq \left\{1,\ldots,p\right\} \backslash \left\{j\right\}}} \frac{|S|!\left(p-|S|-1\right)!}{p!} \left(val\left(S \cup \left\{j\right\}\right) - val\left(S\right)\right) .\end{align} \tag{ 1 }$$

In the formula (1), S represents a subset of features used in the model, x is the vector of feature values for the instance being explained, and p is the total number of features. The function val(x) represents the prediction for the feature values x, marginalized over the features not included in the set S. It is obtained by integrating the model's predictions over the ranges of the excluded features, weighted by their respective probabilities. The Shapley value for feature j is calculated by summing the contributions of all subsets S that do not contain feature j and weighing each contribution by the number of possible orderings of the features in the subset [37, 38].

The Shapley value satisfies several desirable properties, such as:

• (i)  
  Completeness/Efficiency: The sum of the Shapley values for all features equals the difference between the model's prediction for the instance x and the average prediction for all possible instances:

  $$\begin{align} \sum_{i = 1}^{p} \phi_j = \hat{f}\left(x\right)-E_X\left(\hat{f}\left(X\right)\right) .\end{align} \tag{ 2 }$$
• (ii)  
  Symmetry: If two features values j and k contribute equally to all possible coalitions, their Shapley values should be the same:

  $$\begin{align} & \mbox{if} \qquad\quad val\left(S \cup \left\{j\right\}\right) = val\left(S \cup \left\{k\right\}\right) \nonumber\\ & \mbox{for all} \quad\, S\subseteq\left\{1,\ldots,p\right\}\backslash\left\{\,j,k\right\} \nonumber\\ & \mbox{then} \,\,\,\,\,\,\quad \phi_j\left(val\right) = \phi_k\left(val\right) .\end{align} \tag{ 3 }$$
• (iii)  
  Dummy: If a feature j has no impact on the model's prediction, regardless of its inclusion in coalitions, its Shapley value should be zero:

  $$\begin{align} & \mbox{if} \quad\quad\,\,\, val\left(S \cup \left\{j\right\}\right) = val\left(S\right) \nonumber\\ & \mbox{for all}\quad S\subseteq\left\{1,\ldots,p\right\} \nonumber\\ & \mbox{then} \quad \phi_j\left(val\right) = 0 .\end{align} \tag{ 4 }$$
• (iv)  
  Linearity: When combining two models, represented by val and val', the overall prediction should correspond to the sum of the contributions from each model

  $$\begin{align} \phi_j\left(val + val^{^{\prime}}\right) = \phi_j\left(val\right)+\phi_j\left(val^{^{\prime}}\right) .\end{align} \tag{ 5 }$$

The linearity property is beneficial when using ensemble models like XGBoost with TreeSHAP. It enables the computation of the Shapley value for a feature by averaging the contributions of each tree in the ensemble. This simplifies the calculation process and enhances the interpretation of individual feature contributions.

In our study, we employed the TreeSHAP method to interpret the results of the XGBoost model. We chose SHAP due to its model-agnostic nature, theoretical grounding, and ability to provide local and global interpretations. This approach improves the transparency, interpretability, and identification of potential issues or biases in the model.

The performances of the classifiers are provided in table 3. It includes key metrics such as precision, recall, F1-score, and support for each Meningitis class within the dataset.

Precision represents the proportion of correctly predicted positive instances out of all instances predicted as positive, indicating the model's ability to avoid false positives. Recall, also known as sensitivity, measures the proportion of correctly predicted positive instances out of all actual positive instances, indicating the model's ability to capture all relevant positives. The F1-score is the harmonic mean of precision and recall, providing a balanced evaluation of the classifier's performance by considering both metrics. Lastly, the support column in the classification report indicates the number of instances belonging to each class, providing context and understanding of the distribution and representation of classes in the dataset.

Table 3 provides valuable insights into the model's ability to classify different Meningitis classes accurately. Random Forest (RF) and XGBoost (XGB) achieved consistently high precision scores, ranging from 0.83 to 1.00 and 0.78 to 1.00, respectively, indicating their effectiveness in minimizing false positives. The Decision Tree (DT) classifier also performed well, although slightly lower than RF and XGB, with precision scores ranging from 0.50 to 0.99.

Regarding recall, XGB demonstrated superior performance, consistently achieving scores above 0.73 for all classes, indicating its capability to capture the majority of actual positive instances. Decision Tree and Random Forest classifiers also exhibited lower recall scores, ranging from 0.42 to 0.99.

The F1 score combines precision and recall, offering a balanced evaluation of the model's overall performance for each class. XGB achieved the highest F1 scores across all classes, ranging from 0.76 to 1.00, followed closely by RF with scores ranging from 0.57 to 1.00. DT showed slightly lower F1 scores, ranging from 0.45 to 0.99, indicating a trade-off between precision and recall.

The support metric ensures a fair evaluation and comparison of the classifiers' performance across classes. The findings demonstrate that XGBoost and Random Forest consistently perform strongly across all metrics. These classifiers are reliable choices for accurate Meningitis classification, providing high precision in identifying positive cases and effectively capturing most positive instances. Their ensemble nature, robustness to overfitting, and ability to capture complex relationships contribute to their superior performance compared to the Decision Tree classifier.

To provide an overview of the classifiers' outcomes, table 4 presents a summary of the obtained results. It can be concluded that the XGBoost model outperformed the other classifiers. With the highest accuracy, precision (Macro Avg), recall (Macro Avg), F1-Score (Macro Avg) scores, and One-vs-Rest AUROC (Area Under the Receiver Operating Characteristic), the XGBoost model demonstrated superior performance in accurately classifying the data and capturing the overall patterns in the dataset.

Furthermore, an in-depth analysis of the distinct test set originating from Setif's Hospital, comprising instances of pneumococcal Meningitis and tuberculous Meningitis, revealed significant performance metrics for the XGBoost model: (Accuracy: 0.7143, Precision: 1.0, Recall: 0.7143, F1-Score: 0.7857). These results underscore the efficacy of the XGBoost model in precisely categorizing cases within this subset, further affirming its robust performance across specific meningitis types.

Figure 3 shows ROC curves and AUC calculations for meningitis classes to compare the classification performance for each class. Therefore, we selected the XGBoost model as the optimal choice for our classification task due to its superior performance. To provide an in-depth understanding of the XGBoost model and the factors driving its excellent performance, we further delve into its interpretability in section 3.4. By utilizing Shapley values, we explore the contributions of individual features towards the model's predictions, unraveling the key insights and highlighting the factors that significantly influence the classification outcomes. This interpretability analysis enhances our understanding of the inner workings of the XGBoost model, shedding light on its decision-making process and reinforcing our confidence in selecting it as the optimal choice for our classification task.

Zoom In Zoom Out Reset image size

In our study, we primarily focused on specific classes of Meningitis, namely meningococcaemia, meningococcal Meningitis, Tuberculous Meningitis, Aseptic Meningitis, H. influenzae Meningitis, and Pneumococcal Meningitis. To gain insights into the predictive performance of our XGBoost model, we generated a summary variable importance plot (figure 4).

Zoom In Zoom Out Reset image size

Our analysis reveals that several features, including Neutrophils and Lymphocytes level, WCC, Protein and Glucose ratio, the presence of Petechiae/Haemorrhagic suffusion, along with signs of vomiting and neck stiffness, as well as the identification of Gram-negative diplococci in CSF bacterioscopy and S. pneumoniae in CSF culture, are among the most influential factors affecting the model's predictions.

Furthermore, we observed that the level of Neutrophils hardly influences the classification of H. influenzae meningitis, Aseptic Meningitis, and Meningitis by other bacteria. WCC has the most significant influence on Meningococcal and Tuberculosis meningitis. We found that the presence of Petechiae/Haemorrhagic suffusion signs hardly influences the classification of Meningococcaemia. These findings align with the existing literature, which recognizes petechial purpuric exanthema as a classic sign of meningococcemia, present in approximately 40%–80% of cases [39].

Glucose ratio substantially impacts Aseptic Meningitis, while Protein ratio has a greater influence on Tuberculosis meningitis. Lymphocytes have a more significant impact on Pneumococcal and Aseptic meningitis cases. The figure shows that Neisseria meningitidis, and Gram-negative diplococcus in CSF are associated with invasive meningococcal disease [40] along with neck stiffness.

The Murky aspect of the CSF has the most impact on H. influenzae meningitis and Pneumococcal cases. Additionally, the presence of S. pneumoniae in CSF culture has the most impact on Pneumococcal meningitis cases. Furthermore, vomiting has a more pronounced effect on H. influenzae Meningitis than on the other meningitis types.

The Global Interpretability results (section 4.2) demonstrate that our diagnosis outcomes typically align with expert knowledge. We provide additional diagrams constructed to depict the feature's importance and effect on each diagnosis outcome of Meningitis type (see figure 5).

Zoom In Zoom Out Reset image size

Notably, as shown in figure 5(a), our findings emphasize the significance of petechiae/hemorrhagic suffusion as the most influential feature, increasing the likelihood of meningococcemia. This observation aligns with clinical reports, where this symptom is detected in approximately 50%–60% of patients, strongly associating it with the disease. In cases of meningococcemia, confirmation of the diagnosis involves detecting the presence of Neisseria meningitidis in blood cultures [41]. The strong positive impact of finding this organism in blood cultures significantly increases the probability of meningococcemia for our model.

Moreover, our analysis reveals that clear CSF appearance has a comparatively modest yet positive influence on the model's predictions. We also find that a low WCC, low protein ratio, low neutrophil level, and elevated glucose ratio positively impact the diagnosis of meningococcemia. However, these results deviate from the typical diagnostic characteristics of the condition, characterized by an elevated WCC count in CSF, increased protein levels, low glucose levels, and gram-negative diplococcus. Further research is essential to reconcile these disparities and gain a comprehensive understanding of the diagnostic indicators for meningococcemia.

The CSF analysis is instrumental in diagnosing meningococcal Meningitis, encompassing vital parameters like Gram stain, culture, glucose and protein levels, and cell count. Notably, CSF findings indicative of bacterial Meningitis frequently include low glucose and elevated protein levels. In some instances, Gram stains may reveal the presence of Gram-negative diplococci [42]. The presence of neutrophils in CSF is a crucial indicator of a bacterial origin of the Meningitis.

Our analysis, represented in the beeswarm plot for Meningococcal Meningitis (figure 5(b)), highlights noteworthy patterns. Specifically, higher values of Gram-negative diplococci, signifying their presence, result in positive SHAP values. This implies that the absence of this bacterial type in CSF bacterioscopy corresponds to a lower predicted Meningococcal meningitis class. In essence, detecting these bacteria in CSF bacterioscopy is a robust indicator of a meningococcal meningitis diagnosis [42]. A similar trend is observed for detecting Neisseria meningitidis in blood culture. Furthermore, 'no agent' in the PCR test aligns with the absence of viral agents as the causative factor in meningococcal Meningitis. Additionally, our findings demonstrate that higher Glucose ratio values yield negative SHAP values, while lower values correspond to positive SHAP values. This pattern also extends to the age attribute. In contrast, the Protein ratio and Neutrophils level exhibit the opposite effect.

Our findings are consistent with these clinical observations and diagnostic standards. Common symptoms of meningococcal Meningitis often involve a stiff neck, reduced cognitive function, and other signs of meningeal inflammation.

Protein ratio, WCC, and Neutrophil levels significantly influence the diagnosis of tuberculous Meningitis. Elevated WCC, Neutrophils, and glucose levels negatively impact the prediction of this class. In contrast, high Lymphocyte level and protein ratio contribute positively to diagnosing tuberculous Meningitis. Shap values also indicate that the presence of the BCG vaccine feature affects the model's predictions for tuberculosis meningitis. This vaccine consistently protects against the most severe forms of TB, including TB meningitis in children [43]. However, its effectiveness in preventing tuberculosis in adults is comparatively lower.

Figure 5(c) demonstrates that vomiting and neck stiffness signs negatively affect the diagnosis of tuberculous Meningitis. Tuberculous Meningitis is frequently observed in patients with tuberculosis and/or HIV/AIDS, either as a new occurrence of the disease or as a consequence of a prior tuberculosis infection. In cases of co-occurring tuberculosis and HIV, patients may currently have tuberculosis, indicating a co-infection, or have a history of tuberculosis. Although our dataset lacks specific information on the causative agent, these observations align with the well-established understanding that tuberculous Meningitis is primarily caused by Mycobacterium tuberculosis. It is essential to consider these factors when diagnosing patients with suspected tuberculous Meningitis.

Aseptic Meningitis is a condition characterized by negative bacterial cultures of CSF and can be caused by various aetiologies [44]. It is commonly, associated with viral Meningitis or prior antibiotic usage. Our analysis (figure 5(d)) has revealed factors that positively correlate with diagnosing aseptic Meningitis. These include elevated levels of glucose and lymphocytes in the CSF, along with negative results in culture and latex tests for bacterial agents. Conversely, a diagnosis of aseptic Meningitis has been associated with lower levels of neutrophils and reduced protein levels in the CSF. The presence of seizures and neck stiffness positively predicts this particular outcome.

The diagnosis of H. influenzae meningitis relies on a combination of clinical manifestations and specific diagnostic investigations, including laboratory testing and CSF analysis. In cases of H. influenzae meningitis, the CSF often appears cloudy or turbid (murky) and contains an increased number of white blood cells, primarily neutrophils. Our Shapley value analysis (figure 5(e)) reinforces these findings, highlighting that high neutrophil values, relatively elevated WCC, and a cloudy CSF appearance positively influence the prediction of the target class. Confirmation of the presence of H. influenzae type b (Hib), the most common causative organism, can be achieved through Gram staining and bacterial culture of the CSF. H. influenzae infections are most prevalent at the extremes of age, affecting infants, young children, and older adults. However, due to specific vaccines, the incidence of H. influenzae infections has significantly decreased in the general population. Nevertheless, it is still observed in patients aged 65 and older with underlying conditions. The dataset used for this analysis consists of medical records and clinical data from individuals aged 18 years and older. This age range was chosen to focus on the adult population and provide insights into CSF profiles in cases of Haemophilus meningitis. Our analysis further confirms that the presence of H. influenzae, as indicated by culture results, is a positive factor in diagnosing Haemophilus meningitis.

Pneumococcal Meningitis is often marked by a cloudy appearance of the CSF, elevated white blood cell count with a predominance of neutrophils, a substantial reduction in glucose levels, occasionally reaching near zero, and high protein levels within the CSF. Moreover, CSF bacterioscopy and LATEX tests can confirm the presence of S. pneumoniae, the bacteria responsible for pneumococcal Meningitis. Our analysis (figure 5(f)) reaffirms the typical diagnostic indicators of pneumococcal Meningitis. This includes the positive influence of a cloudy/Murky appearance of the CSF and decreased glucose levels on predicting the disease. Furthermore, the presence of Gram-positive cocci and Gram-positive diplococci in CSF bacterioscopy strongly contributes to diagnosing pneumococcal Meningitis. Consistent with these findings, identifying S. pneumoniae through culture and LATEX tests further supports the diagnosis of Pneumococcal Meningitis.

Figure 6 illustrates the SHAP dependence plot, offering insights into the impact of Neutrophil and Lymphocyte levels on the model's predictions. It provides a clear visualization of how changes in these specific biomarkers impact the predictive outcome of the model. We focused our analysis on cases of meningococcal Meningitis, tuberculosis meningitis, aseptic Meningitis, and H. influenzae meningitis, as these categories displayed distinct interaction patterns with these features.

Zoom In Zoom Out Reset image size

Analysis shows a common trend in the dependence plots, characterized by a monotonically increasing curve for meningococcal and H. influenzae Meningitis. This suggests that changes in these features significantly affect the model's predictions, as evidenced by the positive Shapley values.

We further identified specific ranges where variations in neutrophil and lymphocyte levels substantially impacted the predictions. This range typically falls within the $\gt$40%

Similarly, for H. influenzae meningitis, we observed a monotonic increasing pattern, particularly with a strong predominance of neutrophils. However, a significant threshold effect becomes evident when neutrophil levels reach approximately 80%, resulting in a noticeable and substantial increase in Shapley values. This highlights the impact of neutrophils on the model's predictions for both types of Meningitis.

Conversely, Tuberculous and Aseptic meningitis exhibited a typical pattern characterized by a monotonically decreasing curve. Positive Shapley values in Tuberculous Meningitis were associated with neutrophil levels below 60% and high lymphocyte levels above 60%, signifying their influence on the model's predictions. In contrast, Shapley values consistently remained below 0 for Aseptic Meningitis, primarily for neutrophil levels around 40%