Development and application of machine learning models for hematological disease diagnosis using routine laboratory...

Area of Science:

• Hematology
• Machine Learning
• Medical Diagnostics

Background:

• Hematological diseases are increasing, necessitating early and accurate diagnosis for improved patient prognosis.
• Traditional diagnostic methods can be time-consuming and resource-intensive.

Purpose of the Study:

• To develop and evaluate machine learning models for the early diagnosis of hematological diseases.
• To compare the performance of these models against human hematologists.
• To create a user-friendly diagnostic platform.

Main Methods:

• Utilized 54 clinical and laboratory parameters to train 7 machine learning models.
• Employed feature selection and ensemble methods to optimize model performance.
• Interpreted model decisions using SHapley Additive exPlanations (SHAP).

Main Results:

• Two ensemble models, EnMod1-46 (46 features) and EnMod2-12 (12 features), showed high accuracy (up to 0.804) and AUC (up to 0.964) in diagnosing 16 hematological diseases.
• SHAP analysis identified key parameters like Platelets (PLT), White Blood Cell (WBC) count, Mean Corpuscular Volume (MCV), Hemoglobin (HGB), Red Blood Cell (RBC) count, and age.
• Models outperformed junior hematologists and matched senior hematologists, with a user-friendly platform developed based on the simplified model.

Conclusions:

• Developed accurate and efficient machine learning models for hematological disease screening.
• The simplified model (EnMod2-12) offers a practical solution, especially for resource-limited settings.
• The diagnostic platform enhances accessibility and risk assessment for hematological conditions.