Learning classification models with soft-label information.

Area of Science:

• Medical Informatics
• Machine Learning
• Computational Biology

Background:

• Clinical data classification models require expert labeling, which is time-consuming and costly.
• Efficiently learning from limited labeled clinical data is crucial for developing automated medical models.

Purpose of the Study:

• To propose and evaluate a new machine learning approach for improved binary classification models using soft labels.
• To reduce the time and cost associated with expert data labeling in medical applications.

Main Methods:

• Developed two methods for learning from soft labels: probabilistic/numeric and ordinal categorical.
• Applied and compared these methods to an alerting model for heparin-induced thrombocytopenia.
• Evaluated model performance using the area under the receiver operating characteristic curve (AUC) on 377 patient instances labeled by three experts.

Main Results:

• The proposed approach demonstrated more efficient learning of classification models compared to traditional methods.
• Significant improvements in AUC scores were observed, particularly when learning from a small number of examples.
• The use of soft labels enhanced classification model performance.

Conclusions:

• A novel framework utilizing auxiliary soft-label information from human experts offers a promising direction for medical classification.
• This approach effectively reduces the time and cost burdens of expert data labeling.
• The findings highlight the potential of soft labels for improving machine learning efficiency in medicine.