Timenet time series classification series#
In this work, we consider application of TimeNet to phenotyping and in-hospital mortality tasks for multivariate clinical time series classification. TimeNet-based features have been shown to be useful for various tasks including ECG classification. We provide experimental details and observations in Sections 8 and 9, respectively, and comparison of two approaches in Section 10, and finally conclude in Section 11. In Section 7, we provide details of cohort selection and datasets. We provide an overview of the proposed domain and task adaptation approaches in Section 4, followed by their details in Sections 5 and 6, respectively. The rest of the paper is organized as follows: In Section 2, we present related work, and provide details of an existing off-the-shelf pre-trained deep RNN, namely TimeNet, in Section 3. Through empirical evaluation of patient phenotyping and mortality prediction tasks on MIMIC-III benchmark dataset, we demonstrate that our transfer learning approaches yield data- and compute-efficient classification models that require little training or fine-tuning effort while yielding classification performance that is comparable with models with hand-crafted features or carefully trained domain-specific deep networks benchmarked in. We also study the trade-off between domain adaptation and task adaptation with respect to the amount of labeled data from the healthcare domain, and show that leveraging pre-trained models from other (seemingly unrelated) domains can be useful in scenarios where task adaptation may be ineffective, e.g., when labeled data for the target task as well as its related tasks within the healthcare domain is scarce. More specifically, we consider two scenarios: (i) extract features from a pre-trained network and use them to build models for target task and (ii) initialize deep network for target task using parameters of a pre-trained network and then fine-tune using labeled training data for target task. In this work, we consider two simple yet effective approaches to transfer the knowledge captured in pre-trained deep RNNs for new target tasks in healthcare domain.
However, fine-tuning a large number of parameters with a small-labeled dataset may still result in overfitting, and requires careful regularization (as we show in Section 9 through empirical evaluation). Transfer learning via fine-tuning parameters of pre-trained models for end tasks has been recently considered for medical applications as well.
Also, it has been argued that transferring weights even from distant tasks can be better than using random initial weights in neural networks. It has been shown that pre-trained networks can learn to extract a rich set of generic features that can then be applied to a wide range of other similar tasks. Moreover, fine-tuning a pre-trained network for the target task is often faster and easier than constructing and training a new network from scratch. For example, training a deep network on a diverse set of images can provide useful features for images from unseen domains. Transfer learning is known to mitigate this: It enables knowledge transfer from neural networks trained on a source task (domain) with sufficient training instances to a related target task (domain) with few training instances. However, like most deep learning approaches, RNNs are prone to overfitting when labeled training data is scarce, and often require careful and computationally expensive hyper-parameter tuning effort. With various medical parameters being recorded over a period of time in EHR databases, recurrent neural networks (RNNs) can be an effective way to model the sequential aspects of EHR data and, in turn, enable applications in diagnoses, mortality prediction, and estimating length of stay. As a result, there has been a rapid growth in the applications of deep learning to various clinical prediction tasks from electronic health records, e.g., Doctor AI for medical diagnosis, Deep Patient to predict future diseases in patients, and DeepR to predict unplanned readmission after discharge. On the other hand, deep learning approaches enable end-to-end learning without the need of hand-crafted and domain-specific features, and have recently produced promising results for various clinical prediction tasks. Traditional machine learning approaches often require careful domain-specific feature engineering to achieve good prediction performance.
Electronic health records (EHR) consisting of the medical history of patients are useful in various clinical applications such as diagnosis and recommending medicine.
0 kommentar(er)
