ConCare: Personalized Clinical Feature Embedding via Capturing the Healthcare Context

Abstract

Predicting the patient's clinical outcome from the historical electronic medical records (EMR) is a fundamental research problem in medical informatics. Most deep learning-based solutions for EMR analysis concentrate on learning the clinical visit embedding and exploring the relations between visits. Although those works have shown superior performances in healthcare prediction, they fail to explore the personal characteristics during the clinical visits thoroughly. Moreover, existing works usually assume that the more recent record weights more in the prediction, but this assumption is not suitable for all conditions. In this paper, we propose ConCare to handle the irregular EMR data and extract feature interrelationship to perform individualized healthcare prediction. Our solution can embed the feature sequences separately by modeling the time-aware distribution. ConCare further improves the multi-head self-attention via the cross-head decorrelation, so that the inter-dependencies among dynamic features and static baseline information can be effectively captured to form the personal health context. Experimental results on two real-world EMR datasets demonstrate the effectiveness of ConCare. The medical findings extracted by ConCare are also empirically confirmed by human experts and medical literature.

Set-Sequence-Graph: A Multi-View Approach Towards Exploiting Reviews for Recommendation

“…As redundant representations easily lead to over-fitting and bad generalizability [9], we propose to mitigate this issue via decorrelating these views. Similar to [8,29], we measure the correlation of two features based on their covariance. Let ( , ) denote a pair of scalar features.…”

AdaCare: Explainable Clinical Health Status Representation Learning via Scale-Adaptive Feature Extraction and Recalibration

“…Recently, due to the remarkable representation learning ability of deep neural networks, many deep learning-based models have been developed to tackle such prediction tasks by using EMR data, including mortality prediction (Ma et al 2020), disease diagnosis prediction (Lee et al 2018), and patient phenotype identification (Baytas et al 2017). Usually, those models first embed the EMR data into lowdimensional feature space to learn the dense representation of the patients' health status and then perform specific clinical analysis tasks based on such representation.…”

EffiCare: Better Prognostic Models via Resource-Efficient Health Embeddings

“…For many machine learning tasks we can observe that more complex models with more data tend to outperform the previous state-of-art model. Particularly deep learning approaches can implicitly learn to extract and apply task relevant features [1][2][3][4][5][6][7][8] . However, complex models require more compute resources and large amounts of data to learn well -usually the more data the better.…”

“…Moreover, a highly limited feature set combined with relatively small dataset does not provide enough input information to fully utilize deep learning methods. This problem is compounded by an increasing trend to adopt the latest large, complex models with many learnable parameters from domains like computer vision or natural language processing [1][2][3][4][5][6][7][8] , where data types are much more homogeneous and large scale data can be readily exploited. On the other hand, Tomašev et al 10 provide a recent example of the performance benefits gained from using all instead of an expert-selected feature set.…”

“…Using smaller resource-efficient models with feature-embeddings, we can handle an amount of data that would be resourceprohibitive for large models -i.e. we train over hundreds of features, while previous (complex) models used only a few hand-picked features [5][6][7][8][9] , given comparable compute setups. By combining two unsupervised pre-training methods 11, 12 , 1 Project code: https://github.com/oguzserbetci/EffiCare.…”