Denoising Autoencoders for Learning from Noisy Patient-Reported Data

Healthcare datasets often include patientreported values, such as mood, symptoms, and meals, which can be subject to varying levels of human error. Improving the accuracy of patient-reported data could help in several downstream tasks, such as remote patient monitoring. In this study, we propose a novel denoising autoencoder (DAE) approach to denoise patient-reported data, drawing inspiration from recent work in computer vision. Our approach is based on the observation that noisy patient-reported data are often collected alongside higher fidelity data collected from wearable sensors. We leverage these auxiliary data to improve the accuracy of the patient-reported data. Our approach combines key ideas from DAEs with co-teaching to iteratively filter and learn from clean patient-reported samples. Applied to the task of recovering carbohydrate values for blood glucose management in diabetes, our approach reduces noise (MSE) in patient-reported carbohydrates from 72g(2) (95% CI: 54-93) to 18g(2) (13-25), outperforming the best baseline (33g(2) (27-43)). Notably, our approach achieves strong performance with only access to patientreported target values, making it applicable to many settings where ground truth data may be unavailable.