[Paper Review] Improving Hospital Mortality Prediction with Medical Named Entities and Multimodal Learning

Mengqi Jin, Mohammad Taha Bahadori+12 more|arXiv (Cornell University)|Nov 29, 2018

Topic Modeling19 references62 citations

TL;DR

The paper proposes a multimodal model that combines structured ICU data with text representations derived from clinical notes and medical named entities, improving in-hospital mortality prediction over a structured-data baseline.

ABSTRACT

Clinical text provides essential information to estimate the acuity of a patient during hospital stays in addition to structured clinical data. In this study, we explore how clinical text can complement a clinical predictive learning task. We leverage an internal medical natural language processing service to perform named entity extraction and negation detection on clinical notes and compose selected entities into a new text corpus to train document representations. We then propose a multimodal neural network to jointly train time series signals and unstructured clinical text representations to predict the in-hospital mortality risk for ICU patients. Our model outperforms the benchmark by 2% AUC.

Motivation & Objective

• Motivate using unstructured clinical notes to enhance mortality risk prediction.
• Leverage an internal medical NER service to extract entities and negations from notes.
• Create two text representations (NoteEmb and EntityEmb) and integrate them with time-series data.
• Evaluate a multimodal architecture versus a baseline LSTM on the MIMIC-III benchmark.
• Quantify performance gains attributable to clinical notes and entities.

Proposed method

• Use Doc2VecC to generate embeddings from daily aggregated notes (NoteEmb).
• Use a neural NER service to extract medical entities and negations, producing EntityEmb.
• Compare two architectures: a baseline LSTM on structured data (Vital) and a multimodal model that fuses text embeddings with time-series signals.
• Train and evaluate models on 42,276 ICU stays from MIMIC-III with 70/15/15 train/validation/test split.
• Replicate the Harutyunyan et al. benchmark setup for reproducibility and comparability.

Experimental results

Research questions

• RQ1 Does incorporating daily aggregated clinical notes improve mortality prediction over structured data alone?
• RQ2 Do embeddings from medical named entities provide additional predictive value over raw note embeddings?
• RQ3 Does a multimodal architecture outperform a pure LSTM using the same data modalities?
• RQ4 What is the quantitative impact of NoteEmb and EntityEmb on AU-ROC and AU-PRC metrics?

Key findings

• Multimodal model with NoteEmb improves AU-ROC to 0.8669 (Vital+NoteEmb, Multi-modal) vs 0.8531 (Vital baseline, LSTM).
• Multimodal model with EntityEmb achieves AU-ROC of 0.8734 (Vital+EntityEmb, Multi-modal) vs 0.8703 (Vital+EntityEmb, LSTM).
• NoteEmb with LSTM yields AU-ROC 0.8496, AU-PRC 0.5040; EntityEmb with LSTM yields AU-ROC 0.8703, AU-PRC 0.5470.
• Overall, embeddings from medical entities consistently outperform tokenized note embeddings in AU-ROC.
• The multimodal architectures demonstrate improved performance over the pure LSTM baseline across feature sets.
• The improvements are directly comparable to the published benchmark under the same data processing flow.