[论文解读] MiME: Multilevel Medical Embedding of Electronic Health Records for Predictive Healthcare

Edward Choi, Cao Xiao+2 more|arXiv (Cornell University)|Oct 22, 2018

Machine Learning in Healthcare参考文献 38被引用 99

一句话总结

MiME learns multilevel embeddings from EHR data by modeling diagnosis–treatment relationships with auxiliary tasks, achieving strong predictive performance especially on small datasets.

ABSTRACT

Deep learning models exhibit state-of-the-art performance for many predictive healthcare tasks using electronic health records (EHR) data, but these models typically require training data volume that exceeds the capacity of most healthcare systems. External resources such as medical ontologies are used to bridge the data volume constraint, but this approach is often not directly applicable or useful because of inconsistencies with terminology. To solve the data insufficiency challenge, we leverage the inherent multilevel structure of EHR data and, in particular, the encoded relationships among medical codes. We propose Multilevel Medical Embedding (MiME) which learns the multilevel embedding of EHR data while jointly performing auxiliary prediction tasks that rely on this inherent EHR structure without the need for external labels. We conducted two prediction tasks, heart failure prediction and sequential disease prediction, where MiME outperformed baseline methods in diverse evaluation settings. In particular, MiME consistently outperformed all baselines when predicting heart failure on datasets of different volumes, especially demonstrating the greatest performance improvement (15% relative gain in PR-AUC over the best baseline) on the smallest dataset, demonstrating its ability to effectively model the multilevel structure of EHR data.

研究动机与目标

• Address data volume insufficiency in deep learning for EHR data by leveraging inherent multilevel structure of codes.
• Learn multilevel embeddings that capture diagnosis–treatment interactions within visits.
• Improve predictive performance on tasks like heart failure prediction using auxiliary prediction tasks without external labels.

提出的方法

• Represent a visit as a set of Dx objects each with a Dx code and associated treatments.
• Compute diagnosis object embeddings o_i from Dx code embeddings and interactions with treatments via g(d_i, m_i).
• Aggregate to visit embedding v using a skip-connected top-down formulation and derive patient representation h for final prediction.
• Use auxiliary prediction tasks to predict Dx and treatment codes from o_i, guiding embeddings without extra labeling.
• Incorporate bilinear-like interaction modeling through g(d_i, m_i) and W_m to capture Dx–Rx interactions.
• Train end-to-end with a target task (e.g., heart failure prediction) plus auxiliary losses (L_aux) for Dx/m predicted codes.

实验结果

研究问题

• RQ1 Can the inherent multilevel structure of EHR data be exploited to improve predictive performance when data is limited?
• RQ2 Do auxiliary tasks based on Dx–treatment relationships improve the quality and generalizability of visit and patient embeddings?
• RQ3 How does MiME compare to baselines that flatten codes or inject ontology-based knowledge in predicting heart failure and sequential diseases?

主要发现

• MiME outperforms baselines across heart failure prediction tasks on datasets of varying sizes, with a 15% relative gain in PR-AUC on the smallest dataset when using MiME with auxiliary tasks.
• MiME consistently surpasses baselines in heart failure prediction and sequential disease prediction across different data volumes and visit complexities.
• Auxiliary tasks improve generalization, especially on smaller or more complex datasets, with MiME aux achieving higher PR-AUC than MiME alone in multiple settings.
• MiME demonstrates robustness to reduced data and captures Dx–treatment interactions more effectively than models using only code hierarchies or flattened representations.