Validating the accuracy of deep learning for the diagnosis of pneumonia on chest x-ray against a robust multimodal reference diagnosis: a post hoc analysis of two prospective studies

Background Artificial intelligence (AI) seems promising in diagnosing pneumonia on chest x-rays (CXR), but deep learning (DL) algorithms have primarily been compared with radiologists, whose diagnosis can be not completely accurate. Therefore, we evaluated the accuracy of DL in diagnosing pneumonia on CXR using a more robust reference diagnosis. Methods We trained a DL convolutional neural network model to diagnose pneumonia and evaluated its accuracy in two prospective pneumonia cohorts including 430 patients, for whom the reference diagnosis was determined a posteriori by a multidisciplinary expert panel using multimodal data. The performance of the DL model was compared with that of senior radiologists and emergency physicians reviewing CXRs and that of radiologists reviewing computed tomography (CT) performed concomitantly. Results Radiologists and DL showed a similar accuracy on CXR for both cohorts (p >= 0.269): cohort 1, radiologist 1 75.5% (95% confidence interval 69.1-80.9), radiologist 2 71.0% (64.4-76.8), DL 71.0% (64.4-76.8); cohort 2, radiologist 70.9% (64.7-76.4), DL 72.6% (66.5-78.0). The accuracy of radiologists and DL was significantly higher (p <= 0.022) than that of emergency physicians (cohort 1 64.0% [57.1-70.3], cohort 2 63.0% [55.6-69.0]). Accuracy was significantly higher for CT (cohort 1 79.0% [72.8-84.1], cohort 2 89.6% [84.9-92.9]) than for CXR readers including radiologists, clinicians, and DL (all p-values < 0.001). Conclusions When compared with a robust reference diagnosis, the performance of AI models to identify pneumonia on CXRs was inferior than previously reported but similar to that of radiologists and better than that of emergency physicians. Relevance statement: The clinical relevance of AI models for pneumonia diagnosis may have been overestimated. AI models should be benchmarked against robust reference multimodal diagnosis to avoid overestimating its performance. Key point center dot We evaluated an openly-access convolutional neural network (CNN) model to diagnose pneumonia on CXRs. center dot CNN was validated against a strong multimodal reference diagnosis. center dot In our study, the CNN performance (area under the receiver operating characteristics curve 0.74) was lower than that previously reported when validated against radiologists' diagnosis (0.99 in a recent meta-analysis). center dot The CNN performance was significantly higher than emergency physicians' (p <= 0.022) and comparable to that of board-certified radiologists (p >= 0.269).