A Reliable Machine Learning Approach applied to Single-Cell Classification in Acute Myeloid Leukemia

Area of Science:

• Medical Informatics
• Machine Learning
• Computational Biology

Background:

• Machine learning applications in medicine are growing, but clinical deployment is limited.
• A key challenge is the inability of current models to generalize to new patient data, leading to unreliable predictions.
• Assessing the reliability of machine learning classifications is crucial for clinical trust and adoption.

Purpose of the Study:

• To propose and evaluate a novel measure for assessing the reliability of machine learning classifications in clinical settings.
• To determine if a new data instance is similar enough to the training set to ensure a trustworthy prediction.
• To differentiate between reliable and unreliable predictions for new, unseen medical data.

Main Methods:

• Developed a new reliability measure based on instance similarity to the training dataset.
• Evaluated whether a new instance would be considered informative by an instance selection algorithm.
• Tested the method on simulated data and a real-world case of Acute Myeloid Leukemia cell classification.

Main Results:

• The proposed reliability measure successfully distinguished between reliable and unreliable classifications.
• Demonstrated the method's effectiveness in both simulated environments and a practical clinical scenario.
• Showcased the ability to identify trustworthy predictions for distinguishing tumor from normal cells in leukemia patients.

Conclusions:

• The novel reliability measure enhances trust in machine learning predictions for clinical applications.
• This approach addresses the generalization gap, improving the reliability of AI in healthcare.
• The method provides a practical tool for validating machine learning model outputs on new patient data.