Development of a screening model for APL using cell population data and deep learning-extracted WBC scattergram...

Area of Science:

• Hematology
• Machine Learning in Medicine
• Diagnostic Technologies

Background:

• Acute promyelocytic leukemia (APL) is a high-risk subtype of acute myeloid leukemia.
• Rapid diagnosis is crucial to reduce early mortality, but current methods are time-consuming and challenging in resource-limited settings.

Purpose of the Study:

• To develop a novel machine learning approach for immediate APL suspicion using routine laboratory data.
• To provide a new diagnostic possibility for under-resourced hospitals.

Main Methods:

• A two-stage machine learning model (RFC-S) was developed using multi-center retrospective data.
• VGG-16 networks extracted 3D scatterplot features from routine blood tests (DIFF and WNB channels).
• The model was optimized using recursive feature elimination and threshold tuning, with external validation.

Main Results:

• The RFC-S model achieved high diagnostic accuracy with an AUC of 0.9893 (test set) and 0.9979 (external validation).
• It demonstrated 98.15% sensitivity and 95.52% specificity, outperforming conventional methods.
• SHAP analysis identified key scattergram features driving predictions; no additional tests are required.

Conclusions:

• The RFC-S model offers an innovative, accurate, and computationally efficient approach to APL screening.
• Its ability to use existing lab data makes it invaluable for resource-constrained settings.
• This practical tool can aid early APL identification and reduce diagnostic delays in diverse clinical environments.