1.1 Objective of the Workshop:

The objective of this two-day workshop is to develop capacity on Malaria Risk Prediction in Kenya using Artificial Intelligence and Machine Learning for improvement of early detection and intervention strategies.

1.2 Specific Objectives:

• Objective: Develop capacity on Malaria Risk Prediction using interpretable ML/AI models
• Significance: Improving early detection and intervention strategies
• Approach: Utilizing R for transparent and explainable predictive modeling

1.3 Target audience:

Malaria Modelers, Statisticians, Data Scientists, Public health researchers, Policy makers, Health workers

1.4 AMMnet’s mission of connecting people & exchanging knowledge

The workshop supports AMMnet’s mission by fostering collaborative environment for sharing knowledge among researchers, health professionals, and policy makers. By engaging in practical training in AI/ML and R, participants will exchange knowledge, create clear predictive models, and utilize these competencies to enhance malaria control methods, driving innovation and tangible solutions. New AMMnet members will gain hands-on experience with R, a widely used programming language in epidemiological modeling.

2.1 Summary of WHO world Malaria 2024 Report

3.1 Key Data Components

1. Patient Information

   • Age (Betweewn 15-49)
   • Gender (Male or Female)
   • Symptoms (e.g., fever, chills, Headache, Nausea, Fatigue, Muscle_Aches)
   • Malaria test results(Outcome:positive/negative)

2. Environmental Predictors

   Temperature

   • Warm temperatures accelerate mosquito development and parasite replication.

   • In highland areas like Kericho and Nyamira, rising temperatures have been linked to increased malaria risk

   Rainfall

   • Rain creates breeding sites for Anopheles mosquitoes.

   • In Kisumu (Lake Basin), malaria incidence rises with rainfall up to 160 mm, peaking after a 15-week lag.

   • In Ukunda (Coastal zone), rainfall up to 150 mm also correlates with increased cases

   • Though temperature has a negative correlation in coastal zone

   Humidity

   • High humidity boosts mosquito survival and biting rates.

   • High humidity increases vector survival rates, especially for mosquitoes

   • Coastal and lake regions, with consistently high humidity, show persistent transmission.

   Endemic Zones

   • Lake Basin & Coastal zones: High transmission due to favorable climate and water bodies.

   • Highlands: Historically low risk, but warming trends are changing that.

   • Seasonal zones: Outbreaks align with rainy seasons.

   • Low-risk zones: Dry or elevated areas with minimal vector activity

   NB:Malaria in Kenya is a climate-sensitive disease, and understanding these environmental links helps tailor interventions.

   NB: Christiansen-Juchtet et al., 2014, reported that temperature, Rainfall and humidity influence the mosquito’s life cycle for their growth and development. The vector larvae survive only when the environmental condition is conducive and the rate at which mosquitoes bite humans by sucking their blood increases when the environmental conditions are favourable for their survival.
   Reference: https://ij-healthgeographics.biomedcentral.com/articles/10.1186/s12942-024-00381-8

3. Parasite information

   • Type of malaria parasite(Plasmodium: Vivax, Falciparum, Malariae, Ovale)
   • Mosquito Density

4. Target Variable

   • Malaria test result (binary: positive/negative)

4.1 Supervised Learning

Definition: The model learns from labeled data — i.e., data with both input features and known outputs (Malaria test results).

Goal:Predict outcomes for new data based on learned patterns.

Common Tasks:

• Classification (predict categories): e.g., spam detection, disease diagnosis

• Regression (predict continuous values): e.g., house price prediction, temperature forecasting

Example: Training a model to predict malaria outcome (Positive/Negative) based on patient-level factors, environmental factors, Parasite information etc

4.2 Unsupervised Learning

Definition: The model learns patterns and structures from unlabeled data — no known outputs provided.

Goal: Find hidden structure or groupings in data.

Common Tasks:

• Clustering: group similar items (e.g., customer segmentation)

• Dimensionality Reduction: simplify datasets (e.g., for visualization or preprocessing)

Example:Group malaria patient records into distinct types of symptom profiles without predefined labels.

4.3 Reinforcement Learning (RL)

Definition:An agent learns by interacting with an environment and receiving rewards or penalties for actions.

Goal: Learn a sequence of actions that maximizes cumulative rewards over time.

Key Concepts:

• Agent: the learner/decision-maker

• Environment: where the agent operates

• Actions: choices the agent can make

• Reward: feedback signal for learning

• Policy: strategy mapping states to actions

Example:Training a robot to navigate teaching an AI to play chess or a video game.

NB: Machine Learning offers the potential to analyze vast amounts of healthcare data, identify complex patterns, and generate predictions that can be more accurate and efficient than traditional diagnostic methods. However, these models, particularly deep learning approaches, are often criticized for being “black-box” systems, where the rationale behind a prediction is not easily understood by clinicians or patients.

5.1 Breakdown of how AI models work:

1. Modeling: The first step is creating the model, which uses complex algorithms to analyze data and make decisions based on it

2. Training: The model is then “trained” by feeding it with large amounts of data. The more data the model processes, the more accurately it can perform its task

3. Inference: The final step is inference, where the trained model is used to make predictions or act on new, previously unseen data.

5.2 Explainable AI (XAI)

Refers to a set of tools and techniques that make the behavior of machine learning models more interpretable and understandable to humans.

Goals To clarify how and why AI models make certain predictions ensuring that both experts and non-experts can trust and effectively use these insights in real-world applications.

Why do we Need Explainable AI (XAI)

In healthcare, the lack of transparency in machine learning models raises significant concerns. Healthcare professionals require models that not only predict accurately but also explain why a particular decision was made. This is particularly important in scenarios like malaria prediction, where medical decisions directly impact patients’ health outcomes. Without explanations, it becomes challenging for clinicians to trust the model’s recommendations or make informed decisions based on them.

Explainable AI (XAI) has emerged as a solution to this challenge, aiming to provide interpretability for complex machine learning algorithms. XAI techniques make the decision making process of models more transparent by providing understandable, human-readable explanations of how predictions are made. This is crucial in healthcare, where model interpretability can enhance clinicians’ trust in AI-driven predictions, improve patient outcomes through better communication, and ensure ethical accountability.

5.3 Importance of XAI in Malaria Prediction

XAI helps bridge the gap between complex AI technologies and actionable insights for policymakers, healthcare workers, and researchers, allowing them to use AI insights more confidently in planning and intervention.

In healthcare, the stakes are high: AI-driven systems are used to make critical decisions regarding diagnoses, treatment plans, and patient care. For healthcare professionals to trust these systems, they must be able to understand the reasoning behind the recommendations and predictions made by the AI. Key reasons for the importance of explainability in healthcare include:

1. Trust and Transparency

   • Stakeholders (e.g., healthcare professionals and policymakers) are more likely to adopt models they understand.
   • Clinicians and patients are more likely to trust AI-driven decisions if they can understand and verify the reasoning behind them.

2. Clinical Decision Support

   • Explainable AI can enhance clinical decision-making by offering insights into why a certain diagnosis or treatment was suggested, allowing healthcare providers to incorporate AI recommendations into their decision-making process more effectively.

3. Actionable Insights

   • Helps identify key factors driving malaria spread and outcomes, such as climatic conditions or patient demographics.

4. Regulatory Compliance

   • Interpretability can meet requirements for explainability in healthcare systems.

   • Healthcare institutions are subject to regulations such as GDPR, and NIST standards,which emphasize transparency and patient consent. AI must be able to justify its decisions, particularly when it comes to sensitive topics such as predicting the likelihood of developing a disease like malaria, to avoid reinforcing existing biases or discrimination.

GDPR: A comprehensive data privacy law in the European Union that gives individuals control over their personal data.

NIST: A voluntary framework developed by the U.S. National Institute of Standards and Technology to help organizations manage risks associated with AI systems

5. Patient Safety

   • Transparent models enable healthcare professionals to identify potential errors, biases, or flaws in the AI’s predictions, reducing the likelihood of harmful recommendations.

5.4 Basic Workflow with DALEX

“DALEX” is powerful tool for interpreting machine learning models. It provides a unified framework for creating model explanations, making it easier to understand both global and local behavior of machine learning models.

1. Preparing the Malaria Dataset

• Before training models, ensure your data is clean and well-prepared.

2. Training Machine Learning Models

• Use the caret package to train models such as Logistic regression, SVM, Random Forest etc

3. Model Interpretability Techniques

-use the DALEX package for machine learning interpretations

Feature Importance using iml(Interpretable Machine Learning)

• Quantifies how much each feature contributes to the model predictions

• Identify most critical predictors of malaria risk

• Understand complex interactions between variables

SHAP (Shapley Additive Explanations)

SHAP has become one of the most widely used XAI methods due to its consistency and mathematical rigor. In malaria prediction models, SHAP can be used to calculate the exact contribution of each feature to the risk prediction, enabling a clear understanding of which factors are most influential

• Explain individual predictions with SHAP values.

• Explains individual predictions using game theory

• Example: Shows how rainfall or temperature affects malaria risk predictions.

Local Interpretable Model-Agnostic Explanations (LIME)

LIME can be applied to explain the reasons behind a specific prediction, such as why a particular patient is predicted to be at risk for malaria.

• Focuses on explaining individual predictions.

• Flexible enough to work with almost any supervised learning model

• Generates plots to illustrate feature contributions

• Example: Explains why a specific patient was classified as high risk

Partial Dependence Plots (PDP)

PDP show the relationship between a feature and the predicted outcome while keeping other features constant. This can be particularly useful in diabetes prediction models, as it allows clinicians to visualize how changes in a single factor (e.g., glucose levels or BMI) affect the likelihood of diabetes while accounting for other variables.

• Understand the relationship between features and predictions.

• Visualizes the relationship between features and the target prediction

• Understand non-linear relationships

5.5 Machine Learning Pipeline

9.1 Check for data structure

str(data)     

'data.frame':   1000 obs. of  18 variables:
 $ Age             : int  25 49 44 40 24 48 26 27 26 33 ...
 $ Gender          : chr  "Female" "Male" "Male" "Female" ...
 $ Rainfall        : num  218.4 62.3 230.6 145.8 250.8 ...
 $ Temperature     : num  26.1 22.9 31.5 20.5 31 26.9 29.3 29.3 25.8 27.2 ...
 $ Humidity        : num  78.4 54.4 88.3 41.6 49.7 82.9 60.5 84 85 63.3 ...
 $ Endemic_Zone    : chr  "Low risk" "Coastal" "Lake Basin" "Lake Basin" ...
 $ Mosquito_Density: num  66.5 167.7 156.7 72.3 46.1 ...
 $ Fever           : chr  "no" "no" "no" "yes" ...
 $ Chills          : chr  "no" "no" "no" "yes" ...
 $ Headache        : chr  "no" "no" "no" "yes" ...
 $ Nausea          : chr  "yes" "no" "no" "yes" ...
 $ Muscle_Aches    : chr  "yes" "yes" "yes" "yes" ...
 $ Fatigue         : chr  "no" "no" "no" "yes" ...
 $ falciparum      : chr  "no" "no" "yes" "yes" ...
 $ vivax           : chr  "no" "no" "no" "no" ...
 $ malariae        : chr  "no" "no" "yes" "yes" ...
 $ ovale           : chr  "no" "no" "no" "yes" ...
 $ Malaria_Result  : chr  "Negative" "Negative" "Negative" "Positive" ...

9.2 Convert all character variables to factors

data[] <- lapply(data, function(x) if(is.character(x)) as.factor(x) else x)

9.3 Diagnose the data set

library(gtsummary)    #create publication-ready summary tables for regression models and descriptive statistics
library(flextable)    #Creates highly customizable tables suitable for reporting and publication
library(dlookr)       #Analyzes the structure and quality of the dataset

diagnose(data) |> flextable()

#Explore individual columns/variables
#unique(data$Region)                          # unique values for single column
#table(data$Malaria_Result)                           # frequency for a single column
#table(data$Region, data$Intervention_Type) # frequencies for multiple columns

9.4 Check for zero variance predictors:

nzv <- nearZeroVar(data[,-18], saveMetrics = TRUE)
print(nzv)

                 freqRatio percentUnique zeroVar   nzv
Age               1.025641           3.5   FALSE FALSE
Gender            1.049180           0.2   FALSE FALSE
Rainfall          1.000000          80.9   FALSE FALSE
Temperature       1.000000          15.1   FALSE FALSE
Humidity          1.500000          48.0   FALSE FALSE
Endemic_Zone      1.196203           0.5   FALSE FALSE
Mosquito_Density  1.000000          77.3   FALSE FALSE
Fever             1.631579           0.2   FALSE FALSE
Chills            1.538071           0.2   FALSE FALSE
Headache          1.450980           0.2   FALSE FALSE
Nausea            1.481390           0.2   FALSE FALSE
Muscle_Aches      1.631579           0.2   FALSE FALSE
Fatigue           1.481390           0.2   FALSE FALSE
falciparum        2.533569           0.2   FALSE FALSE
vivax             2.597122           0.2   FALSE FALSE
malariae          2.257329           0.2   FALSE FALSE
ovale             2.424658           0.2   FALSE FALSE

##The results above show that there is no feature with zero variance
## Remove nzv

#data <- data[, !nzv$nzv]
#dim(data)

9.5 Visualizing the Target Variable (Malaria Test Results)

library(ggplot2)       ## for data visualization

# Plot Target variable using ggplot2 function
# Sample dataset
dt <- data.frame(
  Malaria_Result = c("Negative", "Positive"),
  Respondent = c(820, 180)) # Replace with actual numbers

# Calculate percentages
dt1 <- dt %>%
  mutate(Percentage = Respondent / sum(Respondent) * 100)

# Create the bar plot
dt1 |> ggplot(aes(x = Malaria_Result, y = Respondent, fill = Malaria_Result)) +
  geom_bar(stat = "identity", show.legend = TRUE) +
  geom_text(aes(label = paste0(Respondent, " (", round(Percentage, 1), "%)")),vjust = -0.5, size = 5) +
  labs(title = "Imbalance Malaria Data",
       x = "Malaria Test Result",
       y = "Respondent",
       fill = "Results") +
  theme_minimal() +
  theme(plot.title = element_text(hjust = 0.5)) # Align the title to the center

15.1 Plot of Confusion Matrix of Logistic Regression

#load required packages
library(reshape2)
library(scales)

# Create the confusion matrix
conf_matrix <- matrix(c(194, 11, 2, 43), nrow = 2, byrow = TRUE)

# Name the rows and columns
rownames(conf_matrix) <- c("Negative","Positive")
colnames(conf_matrix) <- c("Negative","Positive")

# Melt the matrix for ggplot
conf_df <- melt(conf_matrix)

# Plot
ggplot(conf_df, aes(x = Var2, y = Var1, fill = value)) +
  geom_tile(color = "white") +
  geom_text(aes(label = value), color = "red", size = 6) +
  scale_fill_gradient(low = "white", high = "purple") +
  labs(title = "Logistic Regreession Confusion Matrix", x = "Predicted label", y = "True label") +
  theme_minimal() +
  theme(
    plot.title = element_text(hjust = 0.5, face = "bold"),
    axis.title.x = element_text(face = "bold"),
    axis.title.y = element_text(face = "bold"),
    axis.text = element_text(face = "bold"))

## Model performance metric for classification Confusion matrix : is a performance evaluation tool used for classification models. It shows how well the model’s predicted classes match the actual classes.

Accuracy: The proportion of all predictions that are correct used to determine the performance of the Algorithms

\[ \text{Accuracy} = \frac{TP + TN}{TP + FP + FN + TN} =\frac{43 + 194}{43 + 11 + 2 + 194} = 0.948 \]

Interpretation:

• How often the classifier is correct overall.

-⚠️ Can be misleading if classes are imbalanced.

True Positive Rate (TPR): Also known as sensitivity or recall

• It is the ratio of correctly predicted positive observations to the actual positives.

\[ \text{Sensitivity} = \frac{TP}{TP + FN} = 0.956 \]

where TP is True Positives and FN is False Negatives.

Interpretation:

• Of all the true positive cases, how many did the model catch?

• 📌 Important in medical diagnosis, fraud detection.

False Positive Rate (FPR). Also known as specificity

• It is the ratio of incorrectly predicted positive observations to the actual negatives.

\[ \text{Specificity} = \frac{TN}{TN + FP} = 0.946 \]

where FP is False Positives and TN is True Negatives.

Interpretation:

• Of all the true negative cases, how many did the model correctly identify?

• 📌 Important to minimize false alarms.

Precision: The proportion of predicted positives that are actually positive

used to test the correctness of the model when it gives a positive outcome

\[ \text{Precision} = \frac{TP}{TP + FP} = 0.796 \]

Interpretation:

• How trustworthy are the positive predictions?

-📌 Important when false positives are costly (e.g., spam filters, cancer screening, malaria detection).

F1 score: The harmonic mean of precision and recall. It balances the two when both are important.

Used to assess the number of variables that are missing in the predictions which are positive target

\[ F_1 Score = \frac{2 \cdot \text{Precision} \cdot \text{Recall}}{\text{Precision} + \text{Recall}} = 0.869 \]

Interpretation:

• F1 reaches its best value at 1 (perfect precision and recall) and worst at 0.

• 📌 Useful in imbalanced classification tasks.

15.2 Importance of features in a Logistic regression Model

# Show relative importance of features
# vip::vip(LRModel)

# Alternatively using ggplot function
var_imp <-varImp(LRModel)
ggplot(var_imp, aes(x = reorder(Variable, Importance), y = importance)) +
  geom_bar(stat = "identity", fill = "tomato") +
  coord_flip() +
  xlab("Variable") +
  ylab("Importance") +
  ggtitle("Feature Importance Plot for LR Model") +
  theme(plot.title = element_text(hjust = 0.5)) # Align the title to the center

15.3 Prepare data for explain() function

library(DALEX)  ## powerful tool for interpreting machine learning models.

# Converts the target variable from categorical ("Positive"/"Negative") to numeric (1=for Positive, 0=otherwise) 
# Required for many ML models and for DALEX to interpret the output as binary classification.

train$Malaria_Result <- ifelse(train$Malaria_Result == "Positive", 1, 0) 

# Create the explainer Object
explainer_1 <- explain(model = LRModel,
                     data = train[, -which(names(train) =="Malaria_Result")],  # Exclude the target column
                     y = train$Malaria_Result,                                 # Target values as vector
                     label = "Local Explanation with DALEX for Logistics Regression")

Preparation of a new explainer is initiated
  -> model label       :  Local Explanation with DALEX for Logistics Regression 
  -> data              :  750  rows  17  cols 
  -> target variable   :  750  values 
  -> predict function  :  yhat.train  will be used (  default  )
  -> predicted values  :  No value for predict function target column. (  default  )
  -> model_info        :  package caret , ver. 7.0.1 , task classification (  default  ) 
  -> predicted values  :  numerical, min =  1.761836e-08 , mean =  0.2061251 , max =  0.9999998  
  -> residual function :  difference between y and yhat (  default  )
  -> residuals         :  numerical, min =  -0.9989722 , mean =  -0.02612506 , max =  0.8216142  
  A new explainer has been created!  

# This creates an explainer1 object used by the DALEX and ingredients packages for interpretability methods

# Select an instance to explain from test set or new unseen data 

set.seed(123)
new_observation_1 <- test[3, -which(names(test)=="Malaria_Result")]  # Select a test instance
#new_observation_1

# Break Down explanation for the instance
local_explanation_1 <- predict_parts(explainer_1, new_observation_1)

# Plot local explanation
plot(local_explanation_1)

15.4 Overview

The graph presents a local explanation of a RF model using the DALEX package. It visualizes how different predictor variables contribute to the prediction for a specific instance. The prediction is represented by the bar on the right, and the contributions of each variable are shown as horizontal bars.

15.5 Breakdown of Contributions

• Intercept: This baseline value represents the model’s prediction when all predictor variables are zero or absent. In this case, the intercept is xxxxxx

• Predictor Variables: Each predictor variable’s contribution is shown as a bar. The color indicates the direction of the contribution:

• Green: Positive contribution, meaning the variable increases the prediction.

• Red: Negative contribution, meaning the variable decreases the prediction.

• The length of the bar represents the magnitude of the contribution.

15.6 Overall Prediction

Summing up all the contributions (intercept + predictor variables), we arrive at the final prediction of xxxx This value represents the probability of a positive outcome, as RF model typically output probabilities.

This graph provides a valuable tool for understanding how a RF model arrives at a specific prediction. It highlights the relative importance of different predictor variables and their impact on the final outcome.

However, it is essential to consider the limitations and interpret the results in conjunction with other model evaluation metrics

15.7 SHAP(SHapley Additive exPlanations)

The Shapley value helps explain how much each feature contributes to the prediction made by a machine learning model. It provides a way to fairly distribute the “credit” for the model’s output across all input features. By visualizing the SHAP plot, you can understand not only which features are important, but also how specific feature values that are driving predictions for individual cases.

15.8 SHAP explanation

shap_values_1 <- predict_parts(explainer_1, new_observation_1, type = "shap")

plot(shap_values_1)

15.9 Overview

1. Leftward(negative): Indicates the feature is pushing the model prediction towards a negative class

2. Larger absolute SHAP values mean a feature has a stronger influence on the prediction.

3. Smaller SHAP values (close to zero) indicate that a feature has minimal influence on the model’s output for that instance

15.10 Partial Dependence Plots (PDP)

PDP show the relationship between a feature and the predicted outcome while keeping other features constant. This can be particularly useful in diabetes prediction models, as it allows clinicians to visualize how changes in a single factor (e.g., glucose levels or BMI) affect the likelihood of diabetes while accounting for other variables.

• Understand the relationship between features and predictions.

• Visualizes the relationship between features and the target prediction

• Understand non-linear relationships

## Partial Dependence Plot for Temperature

pdp <- model_profile(explainer_1, variables = "Temperature", type = "partial")
plot(pdp)

## Partial Dependence Plot for Humidity

pdp <- model_profile(explainer_1, variables = "Humidity", type = "partial")
plot(pdp)

16.1 Prediction & Evaluation of the RF model performance metrics

# Prediction using RF model
RFpred=predict(RFModel,newdata = test)

# Evaluation of the RF model performance metrics
RF_CM<- confusionMatrix(RFpred,as.factor(test$Malaria_Result), positive = "Positive", mode='everything')
RF_CM

Confusion Matrix and Statistics

          Reference
Prediction Negative Positive
  Negative      201        5
  Positive        4       40
                                          
               Accuracy : 0.964           
                 95% CI : (0.9328, 0.9834)
    No Information Rate : 0.82            
    P-Value [Acc > NIR] : 3.662e-12       
                                          
                  Kappa : 0.877           
                                          
 Mcnemar's Test P-Value : 1               
                                          
            Sensitivity : 0.8889          
            Specificity : 0.9805          
         Pos Pred Value : 0.9091          
         Neg Pred Value : 0.9757          
              Precision : 0.9091          
                 Recall : 0.8889          
                     F1 : 0.8989          
             Prevalence : 0.1800          
         Detection Rate : 0.1600          
   Detection Prevalence : 0.1760          
      Balanced Accuracy : 0.9347          
                                          
       'Positive' Class : Positive        
                                          

16.2 Plot of Confusion Matrix of Random Forest

#load required packages
library(reshape2)
library(scales)

# Create the confusion matrix
conf_matrix <- matrix(c(201, 5, 4, 40), nrow = 2, byrow = TRUE)

# Name the rows and columns
rownames(conf_matrix) <- c("Negative","Positive")
colnames(conf_matrix) <- c("Negative","Positive")

# Melt the matrix for ggplot
conf_df <- melt(conf_matrix)

# Plot
ggplot(conf_df, aes(x = Var2, y = Var1, fill = value)) +
  geom_tile(color = "white") +
  geom_text(aes(label = value), color = "black", size = 6) +
  scale_fill_gradient(low = "white", high = "purple") +
  labs(title = "Random Forest Confusion Matrix", x = "Predicted label", y = "True label") +
  theme_minimal() +
  theme(
    plot.title = element_text(hjust = 0.5, face = "bold"),
    axis.title.x = element_text(face = "bold"),
    axis.title.y = element_text(face = "bold"),
    axis.text = element_text(face = "bold"))

16.3 Importance of features in a Random Model

# Show relative importance of features
# vip::vip(RFModel)

# Alternatively using ggplot function
var_imp <-varImp(RFModel)
ggplot(var_imp, aes(x = reorder(Variable, Importance), y = importance)) +
  geom_bar(stat = "identity", fill = "tomato") +
  coord_flip() +
  xlab("Variable") +
  ylab("Importance") +
  ggtitle("Feature Importance Plot for RF Model") +
  theme(plot.title = element_text(hjust = 0.5)) # Align the title to the center

16.4 Prepare data for explain() function

library(DALEX)

# Converts the target variable from categorical ("Positive"/"Negative") to numeric (1 for Positive, 0 otherwise) 
# Required for many ML models and for DALEX to interpret the output as binary classification.

train$Malaria_Result <- ifelse(train$Malaria_Result == "Positive", 1, 0) 

# Create the explainer Object
explainer_2 <- explain(model = RFModel,
                     data = train[, -which(names(train) =="Malaria_Result")],  # Exclude the target column
                     y = train$Malaria_Result,                                 # Target values as vector
                     label = "Local Explanation with DALEX for Random Forest")

Preparation of a new explainer is initiated
  -> model label       :  Local Explanation with DALEX for Random Forest 
  -> data              :  750  rows  17  cols 
  -> target variable   :  750  values 
  -> predict function  :  yhat.train  will be used (  default  )
  -> predicted values  :  No value for predict function target column. (  default  )
  -> model_info        :  package caret , ver. 7.0.1 , task classification (  default  ) 
  -> predicted values  :  numerical, min =  0 , mean =  0.2224133 , max =  1  
  -> residual function :  difference between y and yhat (  default  )
  -> residuals         :  numerical, min =  -1 , mean =  -0.2224133 , max =  0  
  A new explainer has been created!  

# This creates an explainer object used by the DALEX and ingredients packages for interpretability methods

# Select an instance to explain
new_observation_2 <- test[3, -which(names(test)=="Malaria_Result")]  # Select a test instance
new_observation_2

  Age Gender Rainfall Temperature Humidity Endemic_Zone Mosquito_Density Fever
4  40 Female    145.8        20.5     41.6   Lake Basin             72.3   yes
  Chills Headache Nausea Muscle_Aches Fatigue falciparum vivax malariae ovale
4    yes      yes    yes          yes     yes        yes    no      yes   yes

# Break Down explanation for the instance
local_explanation_2 <- predict_parts(explainer_2, new_observation_2)

# Plot local explanation
plot(local_explanation_2)

## Overview

The graph presents a local explanation of a RF model using the DALEX package. It visualizes how different predictor variables contribute to the prediction for a specific instance. The prediction is represented by the bar on the right, and the contributions of each variable are shown as horizontal bars.

16.5 Breakdown of Contributions

• Intercept: This baseline value represents the model’s prediction when all predictor variables are zero or absent. In this case, the intercept is 0.221

• Predictor Variables: Each predictor variable’s contribution is shown as a bar. The color indicates the direction of the contribution:

• Green: Positive contribution, meaning the variable increases the prediction.

• Red: Negative contribution, meaning the variable decreases the prediction.

• The length of the bar represents the magnitude of the contribution.

16.6 Overall Prediction

Summing up all the contributions (intercept + predictor variables), we arrive at the final prediction of 0.95. This value represents the probability of a positive outcome, as RF model typically output probabilities.

This graph provides a valuable tool for understanding how a RF model arrives at a specific prediction. It highlights the relative importance of different predictor variables and their impact on the final outcome.

However, it is essential to consider the limitations and interpret the results in conjunction with other model evaluation metrics

16.7 SHAP(SHapley Additive exPlanations)

The Shapley value helps explain how much each feature contributes to the prediction made by a machine learning model. It provides a way to fairly distribute the “credit” for the model’s output across all input features. By visualizing the SHAP plot, you can understand not only which features are important, but also how specific feature values that are driving predictions for individual cases.

16.8 SHAP explanation

shap_values_2 <- predict_parts(explainer_2, new_observation_2, type = "shap")

plot(shap_values_2)

16.9 Overview

1. Leftward(negative): Indicates the feature is pushing the model prediction towards a negative class

2. Larger absolute SHAP values mean a feature has a stronger influence on the prediction.

3. Smaller SHAP values (close to zero) indicate that a feature has minimal influence on the model’s output for that instance

16.10 Partial Dependence Plots (PDP)

PDP show the relationship between a feature and the predicted outcome while keeping other features constant. This can be particularly useful in diabetes prediction models, as it allows clinicians to visualize how changes in a single factor (e.g., glucose levels or BMI) affect the likelihood of diabetes while accounting for other variables.

• Understand the relationship between features and predictions.

• Visualizes the relationship between features and the target prediction

• Understand non-linear relationships

## Partial Dependence Plot for Humidity

pdp <- model_profile(explainer_2, variables = "Humidity", type = "partial")
plot(pdp)

16.11 Support Vector Machine (SVM)

#load library
library(caret)

#Tune the grid
svm_tunegrid <- expand.grid(sigma = c(0.01, 0.1, 0.2), C = c(0.1, 1, 10))

tic()
set.seed(123)                ## Ensures that results are reproducible

# Training a model with standardization: This step is useful for models that are sensitive to the scale of the data such as:KNN, SVM, DT etc
SVMModel <- train(factor(Malaria_Result) ~ ., 
                  data = over, 
                  method = "svmRadial", 
                  trControl = control,
                  tuneGrid =svm_tunegrid,
                  preProcess= c("center", "scale"))      ##function to normalize the predictors

toc()

92.72 sec elapsed

16.12 View the Model

SVMModel

Support Vector Machines with Radial Basis Function Kernel 

1230 samples
  17 predictor
   2 classes: 'Negative', 'Positive' 

Pre-processing: centered (20), scaled (20) 
Resampling: Cross-Validated (10 fold, repeated 5 times) 
Summary of sample sizes: 1106, 1106, 1108, 1107, 1107, 1108, ... 
Resampling results across tuning parameters:

  sigma  C     ROC        Sens       Spec     
  0.01    0.1  0.9953629  0.9684876  0.9693601
  0.01    1.0  0.9958446  0.9687996  0.9687150
  0.01   10.0  0.9972916  0.9778689  0.9895717
  0.10    0.1  0.9982851  0.9792015  0.9846854
  0.10    1.0  0.9996592  0.9925225  0.9937969
  0.10   10.0  0.9995957  0.9915494  0.9931465
  0.20    0.1  0.9996107  0.9990270  0.9729773
  0.20    1.0  0.9995456  0.9990217  0.9921735
  0.20   10.0  0.9995456  0.9993443  0.9921735

ROC was used to select the optimal model using the largest value.
The final values used for the model were sigma = 0.1 and C = 1.

16.13 View the Best Tune

SVMModel$bestTune

  sigma C
5   0.1 1

16.14 View the Results

SVMModel$results

  sigma    C       ROC      Sens      Spec       ROCSD      SensSD     SpecSD
1  0.01  0.1 0.9953629 0.9684876 0.9693601 0.003822631 0.022073584 0.02232086
2  0.01  1.0 0.9958446 0.9687996 0.9687150 0.004190148 0.020157785 0.02335719
3  0.01 10.0 0.9972916 0.9778689 0.9895717 0.003600154 0.017037069 0.01390537
4  0.10  0.1 0.9982851 0.9792015 0.9846854 0.001962261 0.017669779 0.01936928
5  0.10  1.0 0.9996592 0.9925225 0.9937969 0.001098089 0.009959455 0.01138115
6  0.10 10.0 0.9995957 0.9915494 0.9931465 0.001174721 0.009973902 0.01194267
7  0.20  0.1 0.9996107 0.9990270 0.9729773 0.001392756 0.003890588 0.02753736
8  0.20  1.0 0.9995456 0.9990217 0.9921735 0.001960388 0.003911731 0.01287408
9  0.20 10.0 0.9995456 0.9993443 0.9921735 0.001960388 0.003245060 0.01287408

16.15 Plot the Best Model

plot(SVMModel)

16.16 Prediction & Evaluation of the SVM model performance metrics

SVMpred = predict(SVMModel, newdata = test)
SVM_CM <- confusionMatrix(SVMpred, as.factor(test$Malaria_Result), positive = "Positive", mode='everything')
SVM_CM

Confusion Matrix and Statistics

          Reference
Prediction Negative Positive
  Negative      202        8
  Positive        3       37
                                          
               Accuracy : 0.956           
                 95% CI : (0.9226, 0.9778)
    No Information Rate : 0.82            
    P-Value [Acc > NIR] : 9.723e-11       
                                          
                  Kappa : 0.8442          
                                          
 Mcnemar's Test P-Value : 0.2278          
                                          
            Sensitivity : 0.8222          
            Specificity : 0.9854          
         Pos Pred Value : 0.9250          
         Neg Pred Value : 0.9619          
              Precision : 0.9250          
                 Recall : 0.8222          
                     F1 : 0.8706          
             Prevalence : 0.1800          
         Detection Rate : 0.1480          
   Detection Prevalence : 0.1600          
      Balanced Accuracy : 0.9038          
                                          
       'Positive' Class : Positive        
                                          

16.17 Plot of Confusion Matrix of SVM

#load required packages
library(reshape2)
library(scales)
# Create the confusion matrix
conf_matrix <- matrix(c(200, 5, 4, 41), nrow = 2, byrow = TRUE)
# Name the rows and columns
rownames(conf_matrix) <- c("Negative","Positive")
colnames(conf_matrix) <- c("Negative","Positive")
# Melt the matrix for ggplot
conf_df <- melt(conf_matrix)
# Plot
ggplot(conf_df, aes(x = Var2, y = Var1, fill = value)) +
  geom_tile(color = "white") +
  geom_text(aes(label = value), color = "black", size = 6) +
  scale_fill_gradient(low = "white", high = "blue") +
  labs(title = "SVM Confusion Matrix", x = "Predicted label", y = "True label") +
  theme_minimal() +
  theme(
    plot.title = element_text(hjust = 0.5, face = "bold"),
    axis.title.x = element_text(face = "bold"),
    axis.title.y = element_text(face = "bold"),
    axis.text = element_text(face = "bold")
  )

16.18 Prepare data for explain() function

# Converts the target variable from categorical ("Positive"/"Negative") to numeric (1 for Positive, 0 otherwise)
train$Malaria_Result <- ifelse(train$Malaria_Result == "Positive", 1, 0)
# Create the explainer Object
explainer_3 <- explain(model = SVMModel,
                     data = train[, -which(names(train) =="Malaria_Result")],  # Exclude the target column
                     y = train$Malaria_Result,                                 # Target values as vector
                     label = "Local Explanation with DALEX for SVM")

Preparation of a new explainer is initiated
  -> model label       :  Local Explanation with DALEX for SVM 
  -> data              :  750  rows  17  cols 
  -> target variable   :  750  values 
  -> predict function  :  yhat.train  will be used (  default  )
  -> predicted values  :  No value for predict function target column. (  default  )
  -> model_info        :  package caret , ver. 7.0.1 , task classification (  default  ) 
  -> predicted values  :  numerical, min =  5.030309e-07 , mean =  0.1796074 , max =  0.9999985  
  -> residual function :  difference between y and yhat (  default  )
  -> residuals         :  numerical, min =  -0.9999985 , mean =  -0.1796074 , max =  -5.030309e-07  
  A new explainer has been created!  

This creates an explainer object used by the DALEX and ingredients packages for interpretability methods

new_observation_3 <- test[3, -which(names(test)=="Malaria_Result")]  # Select a test instance
new_observation_3

  Age Gender Rainfall Temperature Humidity Endemic_Zone Mosquito_Density Fever
4  40 Female    145.8        20.5     41.6   Lake Basin             72.3   yes
  Chills Headache Nausea Muscle_Aches Fatigue falciparum vivax malariae ovale
4    yes      yes    yes          yes     yes        yes    no      yes   yes

# Break Down explanation for the instance
local_explanation_3 <- predict_parts(explainer_3, new_observation_3)

# Plot local explanation
plot(local_explanation_2)

## Overview

The graph presents a local explanation of a RF model using the DALEX package. It visualizes how different predictor variables contribute to the prediction for a specific instance. The prediction is represented by the bar on the right, and the contributions of each variable are shown as horizontal bars.

16.19 Breakdown of Contributions

• Intercept: This baseline value represents the model’s prediction when all predictor variables are zero or absent. In this case, the intercept is 0.221

• Predictor Variables: Each predictor variable’s contribution is shown as a bar. The color indicates the direction of the contribution:

• Green: Positive contribution, meaning the variable increases the prediction.

• Red: Negative contribution, meaning the variable decreases the prediction.

• The length of the bar represents the magnitude of the contribution.

16.20 Overall Prediction

Summing up all the contributions (intercept + predictor variables), we arrive at the final prediction of 0.95. This value represents the probability of a positive outcome, as RF model typically output probabilities.

This graph provides a valuable tool for understanding how a RF model arrives at a specific prediction. It highlights the relative importance of different predictor variables and their impact on the final outcome.

However, it is essential to consider the limitations and interpret the results in conjunction with other model evaluation metrics

16.21 SHAP(SHapley Additive exPlanations)

The Shapley value helps explain how much each feature contributes to the prediction made by a machine learning model. It provides a way to fairly distribute the “credit” for the model’s output across all input features. By visualizing the SHAP plot, you can understand not only which features are important, but also how specific feature values that are driving predictions for individual cases.

16.22 SHAP explanation

shap_values_3 <- predict_parts(explainer_3, new_observation_3, type = "shap")

plot(shap_values_3)

16.23 Overview

1. Leftward(negative): Indicates the feature is pushing the model prediction towards a negative class

2. Larger absolute SHAP values mean a feature has a stronger influence on the prediction.

3. Smaller SHAP values (close to zero) indicate that a feature has minimal influence on the model’s output for that instance

17.1 View the Best Tune

KNNModel$bestTune

  k
1 5

17.2 Plot the Model

plot(KNNModel)

17.3 Prediction & Evaluation of the KNN model performance metrics

KNNpred = predict(KNNModel, newdata = test)
KNN_CM <- confusionMatrix(KNNpred, as.factor(test$Malaria_Result), positive = "Positive", mode='everything')
KNN_CM

Confusion Matrix and Statistics

          Reference
Prediction Negative Positive
  Negative      183        1
  Positive       22       44
                                          
               Accuracy : 0.908           
                 95% CI : (0.8652, 0.9408)
    No Information Rate : 0.82            
    P-Value [Acc > NIR] : 7.049e-05       
                                          
                  Kappa : 0.7364          
                                          
 Mcnemar's Test P-Value : 3.042e-05       
                                          
            Sensitivity : 0.9778          
            Specificity : 0.8927          
         Pos Pred Value : 0.6667          
         Neg Pred Value : 0.9946          
              Precision : 0.6667          
                 Recall : 0.9778          
                     F1 : 0.7928          
             Prevalence : 0.1800          
         Detection Rate : 0.1760          
   Detection Prevalence : 0.2640          
      Balanced Accuracy : 0.9352          
                                          
       'Positive' Class : Positive        
                                          

17.4 Plot of Confusion Matrix of KNN

#load required packages
library(reshape2)
library(scales)
# Create the confusion matrix
conf_matrix <- matrix(c(183, 1, 22, 44), nrow = 2, byrow = TRUE)
# Name the rows and columns
rownames(conf_matrix) <- c("Negative","Positive")
colnames(conf_matrix) <- c("Negative","Positive")
# Melt the matrix for ggplot
conf_df <- melt(conf_matrix)
# Plot
ggplot(conf_df, aes(x = Var2, y = Var1, fill = value)) +
  geom_tile(color = "white") +
  geom_text(aes(label = value), color = "black", size = 6) +
  scale_fill_gradient(low = "white", high = "blue") +
  labs(title = "KNN Confusion Matrix", x = "Predicted label", y = "True label") +
  theme_minimal() +
  theme(
    plot.title = element_text(hjust = 0.5, face = "bold"),
    axis.title.x = element_text(face = "bold"),
    axis.title.y = element_text(face = "bold"),
    axis.text = element_text(face = "bold")
  )

17.5 Prepare data for explain() function

# Converts the target variable from categorical ("Positive"/"Negative") to numeric (1 for Positive, 0 otherwise)
train$Malaria_Result <- ifelse(train$Malaria_Result == "Positive", 1, 0)
# Create the explainer Object
explainer_4 <- explain(model = KNNModel,
                     data = train[, -which(names(train) =="Malaria_Result")],  # Exclude the target column
                     y = train$Malaria_Result,                                 # Target values as vector
                     label = "Local Explanation with DALEX for KNN")

Preparation of a new explainer is initiated
  -> model label       :  Local Explanation with DALEX for KNN 
  -> data              :  750  rows  17  cols 
  -> target variable   :  750  values 
  -> predict function  :  yhat.train  will be used (  default  )
  -> predicted values  :  No value for predict function target column. (  default  )
  -> model_info        :  package caret , ver. 7.0.1 , task classification (  default  ) 
  -> predicted values  :  numerical, min =  0 , mean =  0.2194729 , max =  1  
  -> residual function :  difference between y and yhat (  default  )
  -> residuals         :  numerical, min =  -1 , mean =  -0.2194729 , max =  0  
  A new explainer has been created!  

This creates an explainer object used by the DALEX and ingredients packages for interpretability methods

new_observation_4 <- test[3, -which(names(test)=="Malaria_Result")]  # Select a test instance
new_observation_4

  Age Gender Rainfall Temperature Humidity Endemic_Zone Mosquito_Density Fever
4  40 Female    145.8        20.5     41.6   Lake Basin             72.3   yes
  Chills Headache Nausea Muscle_Aches Fatigue falciparum vivax malariae ovale
4    yes      yes    yes          yes     yes        yes    no      yes   yes

# Break Down explanation for the instance
local_explanation_4 <- predict_parts(explainer_4, new_observation_4)
# Plot local explanation
plot(local_explanation_4)

17.6 Overview

The graph presents a local explanation of a KNN model using the DALEX package. It visualizes how different predictor variables contribute to the prediction for a specific instance. The prediction is represented by the bar on the right, and the contributions of each variable are shown as horizontal bars.

17.7 SHAP explanation

shap_values_4 <- predict_parts(explainer_4, new_observation_4, type = "shap")

plot(shap_values_4)

17.8 Decision Tree (DT)

set.seed(123)                ## Ensures that results are reproducible
tic()
DTModel <- train(factor(Malaria_Result) ~ ., 
                  data = over, 
                  method = "rpart", 
                  trControl = control,
                  preProcess= c("center", "scale"))
toc()

1.64 sec elapsed

17.9 View the Model

DTModel

CART 

1230 samples
  17 predictor
   2 classes: 'Negative', 'Positive' 

Pre-processing: centered (20), scaled (20) 
Resampling: Cross-Validated (10 fold, repeated 5 times) 
Summary of sample sizes: 1106, 1106, 1108, 1107, 1107, 1108, ... 
Resampling results across tuning parameters:

  cp          ROC        Sens       Spec     
  0.02113821  0.8716873  0.8670333  0.8565944
  0.07073171  0.8278307  0.7677261  0.8757747
  0.60813008  0.6312374  0.7784981  0.4839767

ROC was used to select the optimal model using the largest value.
The final value used for the model was cp = 0.02113821.

17.10 View the Best Tune

DTModel$bestTune

          cp
1 0.02113821

17.11 Plot the Model

plot(DTModel)

17.12 Fancy Rpartplot

library(rpart.plot)
rpart.plot(DTModel$finalModel, 
           main = "Decision Tree for Malaria Prediction",
           extra = 104, 
           type = 3, 
           fallen.leaves = TRUE, 
           box.palette = "RdBu", 
           shadow.col = "gray", 
           nn = TRUE)

17.13 Prediction & Evaluation of the DT model performance metrics

DTpred = predict(DTModel, newdata = test)
DT_CM <- confusionMatrix(DTpred, as.factor(test$Malaria_Result), positive = "Positive", mode='everything')
DT_CM

Confusion Matrix and Statistics

          Reference
Prediction Negative Positive
  Negative      175        6
  Positive       30       39
                                          
               Accuracy : 0.856           
                 95% CI : (0.8063, 0.8971)
    No Information Rate : 0.82            
    P-Value [Acc > NIR] : 0.0780309       
                                          
                  Kappa : 0.5962          
                                          
 Mcnemar's Test P-Value : 0.0001264       
                                          
            Sensitivity : 0.8667          
            Specificity : 0.8537          
         Pos Pred Value : 0.5652          
         Neg Pred Value : 0.9669          
              Precision : 0.5652          
                 Recall : 0.8667          
                     F1 : 0.6842          
             Prevalence : 0.1800          
         Detection Rate : 0.1560          
   Detection Prevalence : 0.2760          
      Balanced Accuracy : 0.8602          
                                          
       'Positive' Class : Positive        
                                          

17.14 Plot of Confusion Matrix of DT

#load required packages
library(reshape2)
library(scales)
# Create the confusion matrix
conf_matrix <- matrix(c(175, 6, 30, 39), nrow = 2, byrow = TRUE)
# Name the rows and columns
rownames(conf_matrix) <- c("Negative","Positive")
colnames(conf_matrix) <- c("Negative","Positive")
# Melt the matrix for ggplot
conf_df <- melt(conf_matrix)
# Plot
ggplot(conf_df, aes(x = Var2, y = Var1, fill = value)) +
  geom_tile(color = "white") +
  geom_text(aes(label = value), color = "black", size = 6) +
  scale_fill_gradient(low = "white", high = "blue") +
  labs(title = "Decision Tree Confusion Matrix", x = "Predicted label", y = "True label") +
  theme_minimal() +
  theme(
    plot.title = element_text(hjust = 0.5, face = "bold"),
    axis.title.x = element_text(face = "bold"),
    axis.title.y = element_text(face = "bold"),
    axis.text = element_text(face = "bold")
  )

17.15 Prepare data for explain() function

# Converts the target variable from categorical ("Positive"/"Negative") to numeric (1 for Positive, 0 otherwise)
train$Malaria_Result <- ifelse(train$Malaria_Result == "Positive", 1, 0)
# Create the explainer Object
explainer_5 <- explain(model = DTModel,
                     data = train[, -which(names(train) =="Malaria_Result")],  # Exclude the target column
                     y = train$Malaria_Result,                                 # Target values as vector
                     label = "Local Explanation with DALEX for Decision Tree")

Preparation of a new explainer is initiated
  -> model label       :  Local Explanation with DALEX for Decision Tree 
  -> data              :  750  rows  17  cols 
  -> target variable   :  750  values 
  -> predict function  :  yhat.train  will be used (  default  )
  -> predicted values  :  No value for predict function target column. (  default  )
  -> model_info        :  package caret , ver. 7.0.1 , task classification (  default  ) 
  -> predicted values  :  numerical, min =  0.12749 , mean =  0.3173826 , max =  0.9013807  
  -> residual function :  difference between y and yhat (  default  )
  -> residuals         :  numerical, min =  -0.9013807 , mean =  -0.3173826 , max =  -0.12749  
  A new explainer has been created!  

# This creates an explainer object used by the DALEX and ingredients packages for interpretability methods
new_observation_5 <- test[3, -which(names(test)=="Malaria_Result")]  # Select a test instance
new_observation_5

  Age Gender Rainfall Temperature Humidity Endemic_Zone Mosquito_Density Fever
4  40 Female    145.8        20.5     41.6   Lake Basin             72.3   yes
  Chills Headache Nausea Muscle_Aches Fatigue falciparum vivax malariae ovale
4    yes      yes    yes          yes     yes        yes    no      yes   yes

# Break Down explanation for the instance
local_explanation_5 <- predict_parts(explainer_5, new_observation_5)
# Plot local explanation
plot(local_explanation_5)

17.16 SHAP explanation

shap_values_5 <- predict_parts(explainer_5, new_observation_5, type = "shap")

plot(shap_values_5)

17.17 Naive Bayes NB

set.seed(123)                ## Ensures that results are reproducible
tic()
NBModel <- train(factor(Malaria_Result) ~ ., 
                  data = over, 
                  method = "naive_bayes", 
                  trControl = control,
                  preProcess= c("center", "scale"))
toc()

2.72 sec elapsed

17.18 View the Model

NBModel

Naive Bayes 

1230 samples
  17 predictor
   2 classes: 'Negative', 'Positive' 

Pre-processing: centered (20), scaled (20) 
Resampling: Cross-Validated (10 fold, repeated 5 times) 
Summary of sample sizes: 1106, 1106, 1108, 1107, 1107, 1108, ... 
Resampling results across tuning parameters:

  usekernel  ROC        Sens       Spec     
  FALSE      0.9939957  0.9658805  0.9391274
   TRUE      0.9957419  0.9434479  0.9807827

Tuning parameter 'laplace' was held constant at a value of 0
Tuning
 parameter 'adjust' was held constant at a value of 1
ROC was used to select the optimal model using the largest value.
The final values used for the model were laplace = 0, usekernel = TRUE
 and adjust = 1.

17.19 View the Best Tune

NBModel$bestTune

  laplace usekernel adjust
2       0      TRUE      1

17.20 View the Results

NBModel$results

  usekernel laplace adjust       ROC      Sens      Spec       ROCSD     SensSD
1     FALSE       0      1 0.9939957 0.9658805 0.9391274 0.004204189 0.02127463
2      TRUE       0      1 0.9957419 0.9434479 0.9807827 0.003101067 0.02500433
      SpecSD
1 0.02930288
2 0.01672640

17.21 Plot the Model

plot(NBModel)

17.22 Prediction & Evaluation of the NB model performance metrics

NBpred = predict(NBModel, newdata = test)
NB_CM <- confusionMatrix(NBpred, as.factor(test$Malaria_Result), positive = "Positive", mode='everything')
NB_CM

Confusion Matrix and Statistics

          Reference
Prediction Negative Positive
  Negative      188        1
  Positive       17       44
                                          
               Accuracy : 0.928           
                 95% CI : (0.8886, 0.9568)
    No Information Rate : 0.82            
    P-Value [Acc > NIR] : 7.338e-07       
                                          
                  Kappa : 0.7858          
                                          
 Mcnemar's Test P-Value : 0.000407        
                                          
            Sensitivity : 0.9778          
            Specificity : 0.9171          
         Pos Pred Value : 0.7213          
         Neg Pred Value : 0.9947          
              Precision : 0.7213          
                 Recall : 0.9778          
                     F1 : 0.8302          
             Prevalence : 0.1800          
         Detection Rate : 0.1760          
   Detection Prevalence : 0.2440          
      Balanced Accuracy : 0.9474          
                                          
       'Positive' Class : Positive        
                                          

17.23 Plot of Confusion Matrix of NB

#load required packages
library(reshape2)
library(scales)
# Create the confusion matrix
conf_matrix <- matrix(c(188, 1, 17, 44), nrow = 2, byrow = TRUE)
# Name the rows and columns
rownames(conf_matrix) <- c("Negative","Positive")
colnames(conf_matrix) <- c("Negative","Positive")
# Melt the matrix for ggplot
conf_df <- melt(conf_matrix)
# Plot
ggplot(conf_df, aes(x = Var2, y = Var1, fill = value)) +
  geom_tile(color = "white") +
  geom_text(aes(label = value), color = "black", size = 6) +
  scale_fill_gradient(low = "white", high = "blue") +
  labs(title = "Naive Bayes Confusion Matrix", x = "Predicted label", y = "True label") +
  theme_minimal() +
  theme(
    plot.title = element_text(hjust = 0.5, face = "bold"),
    axis.title.x = element_text(face = "bold"),
    axis.title.y = element_text(face = "bold"),
    axis.text = element_text(face = "bold")
  )

17.24 Prepare data for explain() function

# Converts the target variable from categorical ("Positive"/"Negative") to numeric (1 for Positive, 0 otherwise)
train$Malaria_Result <- ifelse(train$Malaria_Result == "Positive", 1, 0)
# Create the explainer Object
explainer_6 <- explain(model = NBModel,
                     data = train[, -which(names(train) =="Malaria_Result")],  # Exclude the target column
                     y = train$Malaria_Result,                                 # Target values as vector
                     label = "Local Explanation with DALEX for Naive Bayes")

Preparation of a new explainer is initiated
  -> model label       :  Local Explanation with DALEX for Naive Bayes 
  -> data              :  750  rows  17  cols 
  -> target variable   :  750  values 
  -> predict function  :  yhat.train  will be used (  default  )
  -> predicted values  :  No value for predict function target column. (  default  )
  -> model_info        :  package caret , ver. 7.0.1 , task classification (  default  ) 
  -> predicted values  :  numerical, min =  2.431903e-06 , mean =  0.2270044 , max =  0.9999973  
  -> residual function :  difference between y and yhat (  default  )
  -> residuals         :  numerical, min =  -0.9999973 , mean =  -0.2270044 , max =  -2.431903e-06  
  A new explainer has been created!  

This creates an explainer object used by the DALEX and ingredients packages for interpretability methods

new_observation_6 <- test[2, -which(names(test)=="Malaria_Result")]  # Select a test instance
new_observation_6

  Age Gender Rainfall Temperature Humidity Endemic_Zone Mosquito_Density Fever
3  44   Male    230.6        31.5     88.3   Lake Basin            156.7    no
  Chills Headache Nausea Muscle_Aches Fatigue falciparum vivax malariae ovale
3     no       no     no          yes      no        yes    no      yes    no

# Break Down explanation for the instance
local_explanation_6 <- predict_parts(explainer_6, new_observation_6)
# Plot local explanation
plot(local_explanation_6)

shap_values_6 <- predict_parts(explainer_6, new_observation_6, type = "shap")
plot(shap_values_6)

17.25 ASSIGNMENT(Try for other Models)