Efficient automated error detection in medical data using deep-learning and label-clustering.

Area of Science:

• Medical Data Science
• Artificial Intelligence in Healthcare
• Machine Learning for Data Quality

Background:

• Medical datasets frequently contain errors due to subjective testing, biological complexity, and privacy constraints.
• Manual error detection is challenging for experts due to data scale and lack of context.
• Existing methods for training AI on noisy data include model robustness, regularization, loss functions, or data subset selection.

Purpose of the Study:

• To develop an automated algorithm for detecting errors in medical datasets.
• To improve the efficiency and accuracy of training artificial intelligence (AI) models on potentially mislabeled medical data.
• To reduce the computational resources required for error detection in large-scale medical datasets.

Main Methods:

• A deep-learning algorithm was combined with a label-clustering approach for automated error detection.
• The method was evaluated on datasets with synthetically introduced label flips.
• Performance was compared against a previous model consensus approach and noise-robust loss functions.

Main Results:

• The automated method achieved up to 85% accuracy in identifying synthetic label errors.
• It required up to 93% fewer computing resources than the consensus approach.
• Trained AI models showed improved stability and accuracy, with up to 45% improvement (from 69% to over 99%) in one case, and significant gains in both binary and multi-class classification tasks.

Conclusions:

• Automated, a priori detection of errors in medical data is feasible without human oversight.
• The proposed deep-learning and label-clustering method offers a computationally efficient and effective solution for improving medical AI model training.
• This approach enhances AI model performance and training stability by addressing data quality issues.