Decision tree-based learning and laboratory data mining: an efficient approach to amebiasis testing

Abstract

Background: Amebiasis represents a significant global health concern. This is especially evident in developing countries, where infections are more common. The primary diagnostic method in laboratories involves the microscopy of stool samples. However, this approach can sometimes result in the misinterpretation of amebiasis as other gastroenteritis (GE) conditions. The goal of the work is to produce a machine learning (ML) model that uses laboratory findings and demographic information to automatically predict amebiasis.

Method: Data extracted from Jordanian electronic medical records (EMR) between 2020 and 2022 comprised 763 amebic cases and 314 nonamebic cases. Patient demographics, clinical signs, microscopic diagnoses, and leukocyte counts were used to train eight decision tree algorithms and compare their accuracy of predictions. Feature ranking and correlation methods were implemented to enhance the accuracy of classifying amebiasis from other conditions.

Results: The primary dependent variables distinguishing amebiasis include the percentage of neutrophils, mucus presence, and the counts of red blood cells (RBCs) and white blood cells (WBCs) in stool samples. Prediction accuracy and precision ranged from 92% to 94.6% when employing decision tree classifiers including decision tree (DT), random forest (RF), XGBoost, AdaBoost, and gradient boosting (GB). However, the optimized RF model demonstrated an area under the curve (AUC) of 98% for detecting amebiasis from laboratory data, utilizing only 300 estimators with a max depth of 20. This study highlights that amebiasis is a significant health concern in Jordan, responsible for 17.22% of all gastroenteritis episodes in this study. Male sex and age were associated with higher incidence of amebiasis (P = 0.014), with over 25% of cases occurring in infants and toddlers.

Conclusions: The application of ML to EMR can accurately predict amebiasis. This finding significantly contributes to the emerging use of ML as a decision support system in parasitic disease diagnosis.

Keywords: E. histolytica; Amebiasis; Decision tree; Electronic medical records (EMR); Feature selection; Jordan; Leukocytosis; Machine learning; Microscopic diagnosis; Stool RBCs.

Fig. 1

a The top features account for 95% of the data’s importance in predicting amebiasis. b A correlation heat map where strong positive correlations are denoted by red, strong negative correlations by blue, and a lack of color indicates weak correlation

Fig. 2

A point-biserial correlation heatmap, where strong positive correlations are denoted by red, strong negative correlations by blue, and a lack of color indicates weak correlation

Fig. 3

Comparison of clinical manifestations, age, and sex of the Jordanian patients with confirmed cases of amebiasis. a Age distribution of the cases and b age distribution according to symptoms being diagnosed. Significant difference when P-value < 0.05. n = 763

Fig. 4

Distribution of laboratory findings, clinical manifestations, age, and sex of Jordanian patients with confirmed cases of amebiasis gastroenteritis. Severity features indicated by WBC count (n = 763)

Fig. 5

Distribution of laboratory findings, clinical manifestations, age, and sex of Jordanian patients with confirmed cases of amebiasis. Severity features indicated by neutrophil percentage (n = 763)

Fig. 6

Grid search process to find the best parameters for RF (top). AUC = 0.98 results when using RF with its best parameters (bottom left), and the confusion matrix results using RF with its best parameters (bottom right)

Fig. 7

Three-dimensional visualization of three features of our data