Medical Data Analysis Using AutoML Frameworks

Recently, there has been a growing interest in applying machine learning (ML) and deep learning to medical big data and smart healthcare. However, it can be challenging to possess both domain knowledge of medical data and expertise in ML necessary for conducting research in this area. In this regard, an automated machine learning (AutoML) framework can be a solution, which automates an ML pipeline-building process from data cleansing and preprocessing to model building, optimization, and performance evaluation. Since AutoML can reduce domain dependency in ML and quickly build high-performance models, it has a high potential for use in the medical and healthcare fields. Therefore, in this paper, we aim to validate the effectiveness of various AutoML frameworks using structured and unstructured medical data such as electronic medical records, medical images, and signal data. More specifically, we compare the performance of representative AutoML frameworks and simple handcrafted models in terms of various metrics such as the area under the curve, accuracy, and F1 score. Our experimental results show that AutoML effectively handles structured and unstructured data in general, but it still needs to be further improved to deal with data that is imbalanced or requires special preprocessing, such as signal data.