Underspecification Presents Challenges for Credibility in Modern Machine Learning

ML models often exhibit unexpectedly poor behavior when they are deployed in real-world domains. We identify underspecification as a key reason for these failures. An ML pipeline is underspecified when it can return many predictors with equivalently strong held-out performance in the training domain. Underspecification is common in modern ML pipelines, such as those based on deep learning. Predictors returned by underspecified pipelines are often treated as equivalent based on their training domain performance, but we show here that such predictors can behave very differently in deployment domains. This ambiguity can lead to instability and poor model behavior in practice, and is a distinct failure mode from previously identified issues arising from structural mismatch between training and deployment domains. We show that this problem appears in a wide variety of practical ML pipelines, using examples from computer vision, medical imaging, natural language processing, clinical risk prediction based on electronic health records, and medical genomics. Our results show the need to explicitly account for underspecification in modeling pipelines that are intended for real-world deployment in any domain.

Quantitative and Qualitative Evaluation of Explainable Deep Learning Methods for Ophthalmic Diagnosis

Background: The lack of explanations for the decisions made by algorithms such as deep learning has hampered their acceptance by the clinical community despite highly accurate results on multiple problems. Recently, attribution methods have emerged for explaining deep learning models, and they have been tested on medical imaging problems. The performance of attribution methods is compared on standard machine learning datasets and not on medical images. In this study, we perform a comparative analysis to determine the most suitable explainability method for retinal OCT diagnosis. Methods: A commonly used deep learning model known as Inception v3 was trained to diagnose 3 retinal diseases - choroidal neovascularization (CNV), diabetic macular edema (DME), and drusen. The explanations from 13 different attribution methods were rated by a panel of 14 clinicians for clinical significance. Feedback was obtained from the clinicians regarding the current and future scope of such methods. Results: An attribution method based on a Taylor series expansion, called Deep Taylor was rated the highest by clinicians with a median rating of 3.85/5. It was followed by two other attribution methods, Guided backpropagation and SHAP (SHapley Additive exPlanations). Conclusion: Explanations of deep learning models can make them more transparent for clinical diagnosis. This study compared different explanations methods in the context of retinal OCT diagnosis and found that the best performing method may not be the one considered best for other deep learning tasks. Overall, there was a high degree of acceptance from the clinicians surveyed in the study. Keywords: explainable AI, deep learning, machine learning, image processing, Optical coherence tomography, retina, Diabetic macular edema, Choroidal Neovascularization, Drusen

Deep Learning vs. Human Graders for Classifying Severity Levels of Diabetic Retinopathy in a Real-World Nationwide Screening Program

Deep learning algorithms have been used to detect diabetic retinopathy (DR) with specialist-level accuracy. This study aims to validate one such algorithm on a large-scale clinical population, and compare the algorithm performance with that of human graders. 25,326 gradable retinal images of patients with diabetes from the community-based, nation-wide screening program of DR in Thailand were analyzed for DR severity and referable diabetic macular edema (DME). Grades adjudicated by a panel of international retinal specialists served as the reference standard. Across different severity levels of DR for determining referable disease, deep learning significantly reduced the false negative rate (by 23%) at the cost of slightly higher false positive rates (2%). Deep learning algorithms may serve as a valuable tool for DR screening.