Acute myeloid leukemia (AML) and acute lymphoblastic leukemia (ALL) are hematologic malignancies characterized by the rapid and uncontrolled proliferation in the bone marrow precursor cells. AML involves the myeloid lineage, with blocked differentiation and accumulation of blasts, leading to failure of normal hematopoiesis. ALL affects the lymphoid lineage, predominated by lymphoblastic blasts, and frequently involves the central nervous system. Both have aggressive progression, requiring early diagnosis and immediate treatment. Diagnostic and prognostic accuracy is essential to guide therapeutic strategies, and artificial intelligence (AI) and machine learning (ML) techniques have emerged as valuable tools to optimize these processes.

To conduct a literature review on the application of machine learning (ML) and artificial intelligence (AI) techniques in the diagnosis and prognosis of acute myeloid leukemia (AML) and acute lymphoblastic leukemia (ALL).

This study is an integrative literature review conducted in PubMed and LILACS databases, using the terms “leukemias,” “machine learning,” “artificial intelligence,” “diagnosis,” and “prognosis,” covering the last 5 years, in both Portuguese and English. Original studies and reviews with quantitative data on technical performance and clinical applicability were included.

In AML, convolutional neural networks (CNNs) demonstrated high diagnostic performance, with a mean accuracy of 95.57% and sensitivity of 85.81%. Multilayer perceptron (MLP) models, using 52 clinical variables, were applied to survival prognosis, with accuracy ranging from 62.1% to 68.5%, depending on the treatment. In ALL, Naïve Bayes classifiers achieved 96.15% accuracy in the identification of blasts in microscopic images. Advanced deep learning techniques, such as YOLOv8 and YOLOv11 combined with ResNet50 and Inception-ResNet-v2 architectures, reached up to 99.7% accuracy in the automatic detection of leukemic cells.

The findings indicate that AI techniques, particularly CNNs, provide high diagnostic accuracy in AML and ALL, surpassing traditional methods in speed and consistency. The lower sensitivity compared to accuracy in AML suggests the need for refinement to reduce false negatives, which are critical in early diagnosis. The modest accuracy of MLP models in AML prognosis reflects the biological complexity and clinical heterogeneity of the disease, indicating that integrative approaches may improve outcome prediction. In ALL, the high accuracy of deep learning models reinforces their potential for immediate clinical application, particularly in high-demand settings and in regions with limited availability of specialists.

Machine learning and artificial intelligence represent significant advances in improving the diagnosis and prognosis of AML and ALL. The high accuracy achieved highlights their complementary and, in some cases, superior role compared to conventional methods. However, challenges such as diagnostic sensitivity and prognostic complexity require ongoing refinement and validation. The integration of these technologies may contribute to therapeutic personalization and improved clinical outcomes.