International testing and refinement of AI algorithms predicting acute leukemia subtypes from routine laboratory...

Area of Science:

• Computational Biology and Bioinformatics
• Hematology and Oncology
• Medical Informatics

Background:

• Health disparities significantly limit access to timely diagnosis and treatment for patients with acute leukemia.
• Artificial Intelligence (AI) presents a potential avenue to mitigate these disparities in leukemia care.
• Existing AI diagnostic tools may face challenges with accuracy and patient inclusion thresholds.

Purpose of the Study:

• To evaluate an AI tool designed for acute leukemia diagnosis using standard laboratory results.
• To improve the accuracy and robustness of AI-driven leukemia diagnosis, particularly for patients initially excluded by confidence thresholds.
• To adapt the AI algorithm for improved performance in pediatric leukemia cases.

Main Methods:

• Retrospective analysis of a diverse, international cohort of 6206 leukemia patients from 20 centers.
• Initial testing of a pretrained AI algorithm for leukemia diagnosis.
• Implementation of an ensemble method (Isolation Forest and Local Outlier Factor) to enhance diagnostic accuracy and reduce patient exclusion.

Main Results:

• The pretrained AI achieved high AUROC metrics (0.94 for AML, 0.98 for APL, 0.84 for ALL) but excluded a large percentage of patients (70.8-92.5%).
• The ensemble method significantly improved AUROC for AML from 0.72 to 0.84 on a hold-out test set, while only excluding 12.1% of patients.
• The AI algorithm was successfully retrained for pediatric leukemia patients.

Conclusions:

• AI holds promise for improving acute leukemia diagnosis and potentially reducing health disparities.
• Ensemble AI methods enhance diagnostic accuracy and patient inclusivity compared to single-model approaches.
• Further development and validation are needed to fully integrate AI tools into clinical practice for diverse leukemia patient populations.