Confounding variables can degrade generalization performance of radiological deep learning models

Background: There is interest in using convolutional neural networks (CNNs) to analyze medical imaging to provide computer aided diagnosis (CAD). Recent work has suggested that image classification CNNs may not generalize to new data as well as previously believed. We assessed how well CNNs generalized across three hospital systems for a simulated pneumonia screening task. Methods and Findings: A cross-sectional design with multiple model training cohorts was used to evaluate model generalizability to external sites. 158,323 chest radiographs were drawn from three institutions: NIH (112,120 from 30,805 patients), Mount Sinai Hospital (MSH; 42,396 from 12,904 patients), July 16, 2018 1/15 ar X iv :1 80 7. 00 43 1v 2 [ cs .C V ] 1 3 Ju l 2 01 8 and Indiana (IU; 3,807 radiographs from 3,683 patients). These patient populations had age mean (S.D.) 46.9 (16.6), 63.2 (16.5), and 49.6 (17), and percent female 43.5%, 44.8%, and 57.1%, respectively. We assessed individual models using area under the receiver operating characteristic curve (AUC) for radiographic findings consistent with pneumonia and compared performance on different test sets with DeLong’s test. The prevalence of pneumonia was high enough at MSH (34.2%) relative to NIH and IU (1.2% and 1.0%) that merely sorting by hospital system achieved an AUC of 0.861 on the joint MSH-NIH dataset. Models trained on data from either NIH or MSH had equivalent performance on IU (p-values 0.580 and 0.273, respectively) and inferior performance on data from each other relative to an internal test set (i.e., new data from within the hospital system used for training data; p-values both < 0.001). The highest internal performance was achieved by combining training and test data from MSH and NIH (AUC 0.931, 95% C.I. 0.927-0.936), but this model demonstrated significantly lower external performance at IU (AUC 0.815, 95% C.I. 0.745-0.885, P = 0.001). To test the effect of pooling data from sites with disparate pneumonia prevalence, we used stratified subsampling to generate MSH-NIH cohorts that only differed in disease prevalence between training data sites. When both training data sites had the same pneumonia prevalence, the model performed consistently on external IU data (P = 0.88). When a ten-fold difference in pneumonia rate was introduced between sites, internal test performance improved compared to the balanced model (10x MSH risk P < 0.001; 10x NIH P = 0.002), but this outperformance failed to generalize to IU (MSH 10x P < 0.001; NIH 10x P = 0.027). CNNs were able to directly detect hospital system of a radiograph for 99.95% NIH (22,050/22,062) and 99.98% MSH (8,386/8,388) radiographs. The primary limitation of our approach and the available public data is that we cannot fully assess what other factors might be contributing to hospital system-specific biases. Conclusions: Pneumonia screening CNNs achieved better internal than external performance in 3 / 5 natural comparisons. When models were trained on pooled data from sites with different pneumonia prevalence, they performed better on new pooled data from these sites but not on external data. CNNs robustly identified hospital system and department within a hospital which can have large differences in disease burden and may confound disease predictions.