[论文解读] Deep EHR: Chronic Disease Prediction Using Medical Notes

Jingshu Liu, Zachariah Zhang+1 more|arXiv (Cornell University)|Aug 15, 2018

Machine Learning in Healthcare参考文献 9被引用 48

一句话总结

本文提出一个多任务深度学习框架，将非结构化的医疗笔记与结构化数据结合起来，以预测慢性疾病的发病，并且笔记有助于提升性能，否定处理（negation handling）可提高准确性。

ABSTRACT

Early detection of preventable diseases is important for better disease management, improved inter-ventions, and more efficient health-care resource allocation. Various machine learning approacheshave been developed to utilize information in Electronic Health Record (EHR) for this task. Majorityof previous attempts, however, focus on structured fields and lose the vast amount of information inthe unstructured notes. In this work we propose a general multi-task framework for disease onsetprediction that combines both free-text medical notes and structured information. We compareperformance of different deep learning architectures including CNN, LSTM and hierarchical models.In contrast to traditional text-based prediction models, our approach does not require disease specificfeature engineering, and can handle negations and numerical values that exist in the text. Ourresults on a cohort of about 1 million patients show that models using text outperform modelsusing just structured data, and that models capable of using numerical values and negations in thetext, in addition to the raw text, further improve performance. Additionally, we compare differentvisualization methods for medical professionals to interpret model predictions.

研究动机与目标

• Motivate early detection of preventable diseases using the rich information in EHRs.
• Propose a general multi-task architecture that fuses free-text notes with structured numerical data.
• Evaluate various deep learning architectures (CNN, LSTM, hierarchical models) on large real-world EHR data.
• Address negations and provide visualization methods to aid clinical interpretability.

提出的方法

• Represent notes with word embeddings (PubMed and in-house StarSpace embeddings) and feed through CNN, LSTM, or hierarchical models.
• Incorporate numerical lab/vital sign values extracted from notes and demographics as additional inputs.
• Use a multi-task framework with disease-specific outputs and masked loss to handle patient-disease eligibility overlaps.
• Apply a negation tagging step (NegEx) to improve interpretation of negated findings.
• Compare baseline models (logistic regression with labs/demographics, TF-IDF text) to deep learning models.
• Provide visualization methods (log-odds based) to identify high-impact n-grams and phrases.

实验结果

研究问题

• RQ1 Can unstructured medical notes improve prediction of disease onset beyond structured data alone?
• RQ2 Which deep learning architectures best leverage notes for multi-disease onset prediction?
• RQ3 How do negations and numerical values in notes affect predictive performance?
• RQ4 What visualization methods best help clinicians interpret model predictions?

主要发现

模型	Heart Failure AUC	Kidney Failure AUC	Stroke AUC
Log Reg Lab/Demo	0.781	0.724	0.70
LSTM Lab/Demo	0.813	0.743	0.699
Logistic Reg Notes	0.810	0.752	0.708
CNN PubMed Embeddings	0.844	0.799	0.711
CNN Single Task	0.847	0.796	0.706
CNN	0.854	0.802	0.714
CNN + Neg Tag	0.867	0.811	0.727
CNN + Neg Tag + Dense	0.880	0.812	0.733
CNN + Neg Tag + Dense + Lab/Demo	0.893	0.822	0.749
BiLSTM	0.869	0.807	0.738
BiLSTM + Neg Tag	0.875	0.811	0.745
BiLSTM + Neg Tag + Dense	0.892	0.823	0.739
BiLSTM + Neg Tag + Dense + Lab/Demo	0.900	0.833	0.753
Enc CNN-LSTM	0.859	0.797	0.727
Enc CNN-LSTM + Lab/Demo	0.885	0.812	0.740

• Models using medical notes outperform those using only lab and demographic data.
• Deep learning models with notes achieve higher AUC than logistic regression baselines with TF-IDF features.
• Incorporating negation tagging and combining labs/demographics yields the best performance across CHF, KF, and stroke.
• Best model (BiLSTM with negation tagging, dense layer, plus lab/demographic inputs) achieved the highest AUC across all three diseases.
• Log-odds based visualizations provide more intuitive explanations than gradient-based methods.