A machine learning model based on readers' characteristics to predict their performances in reading screening...

Area of Science:

• Radiology
• Machine Learning
• Medical Imaging Analysis

Background:

• Assessing radiologist performance is crucial for mammography screening quality.
• Objective performance metrics like AUC and lesion sensitivity are key indicators.
• Reader characteristics may influence performance.

Purpose of the Study:

• To develop and validate a machine learning model predicting radiologist performance.
• To utilize reader characteristics and case difficulty for performance prediction.
• To evaluate the model's ability to categorize readers into performance groups.

Main Methods:

• Collected data from 905 radiologists and breast physicians interpreting mammographic images.
• Gathered reader demographics (reading volume, experience) and case difficulty.
• Employed ensemble of regression trees to predict AUC and lesion sensitivity.
• Validated model performance using leave-one-out cross-validation.

Main Results:

• Achieved a Pearson correlation of 0.60 between predicted and actual AUC (p < 0.001).
• Model demonstrated 0.86 performance in differentiating readers in the first and fourth AUC quartiles.
• Model achieved 0.91 AUC for distinguishing readers in the first vs. fourth quartile based on lesion sensitivity.

Conclusions:

• A machine learning model can effectively categorize radiologists as high- or low-performing.
• This predictive model can enhance quality assurance in screening programs.
• Optimizing double-reading practices can be facilitated by this approach.