The Subgroup Imperative: Chest Radiograph Classifier Generalization Gaps in Patient, Setting, and Pathology Subgroups

Doctor's Orders-Why Radiologists Should Consider Adjusting Commercial Machine Learning Applications in Chest Radiography to Fit Their Specific Needs

This retrospective study evaluated a commercial deep learning (DL) software for chest radiographs and explored its performance in different scenarios. A total of 477 patients (284 male, 193 female, mean age 61.4 (44.7-78.1) years) were included. For the reference standard, two radiologists performed independent readings on seven diseases, thus reporting 226 findings in 167 patients. An autonomous DL reading was performed separately and evaluated against the gold standard regarding accuracy, sensitivity and specificity using ROC analysis. The overall average AUC was 0.84 (95%-CI 0.76-0.92) with an optimized DL sensitivity of 85% and specificity of 75.4%. The best results were seen in pleural effusion with an AUC of 0.92 (0.885-0.955) and sensitivity and specificity of each 86.4%. The data also showed a significant influence of sex, age, and comorbidity on the level of agreement between gold standard and DL reading. About 40% of cases could be ruled out correctly when screening for only one specific disease with a sensitivity above 95% in the exploratory analysis. For the combined reading of all abnormalities at once, only marginal workload reduction could be achieved due to insufficient specificity. DL applications like this one bear the prospect of autonomous comprehensive reporting on chest radiographs but for now require human supervision. Radiologists need to consider possible bias in certain patient groups, e.g., elderly and women. By adjusting their threshold values, commercial DL applications could already be deployed for a variety of tasks, e.g., ruling out certain conditions in screening scenarios and offering high potential for workload reduction.

Empirical data drift detection experiments on real-world medical imaging data

While it is common to monitor deployed clinical artificial intelligence (AI) models for performance degradation, it is less common for the input data to be monitored for data drift - systemic changes to input distributions. However, when real-time evaluation may not be practical (eg., labeling costs) or when gold-labels are automatically generated, we argue that tracking data drift becomes a vital addition for AI deployments. In this work, we perform empirical experiments on real-world medical imaging to evaluate three data drift detection methods' ability to detect data drift caused (a) naturally (emergence of COVID-19 in X-rays) and (b) synthetically. We find that monitoring performance alone is not a good proxy for detecting data drift and that drift-detection heavily depends on sample size and patient features. Our work discusses the need and utility of data drift detection in various scenarios and highlights gaps in knowledge for the practical application of existing methods.