[Paper Review] Improving Clinical Document Understanding on COVID-19 Research with Spark NLP

Veysel Kocaman, David Talby|arXiv (Cornell University)|Dec 7, 2020

Machine Learning in Healthcare23 references4 citations

TL;DR

This paper presents a scalable, high-accuracy clinical NLP pipeline built on Spark NLP that improves clinical document understanding in COVID-19 research by recognizing over 100 entity types and detecting assertion statuses (present, absent, conditional, someone-else) with state-of-the-art performance. The system leverages pre-trained Bi-LSTM-CNN-Char models and scales efficiently across distributed clusters, enabling fast, accurate extraction of structured insights from the CORD-19 dataset.

ABSTRACT

Following the global COVID-19 pandemic, the number of scientific papers studying the virus has grown massively, leading to increased interest in automated literate review. We present a clinical text mining system that improves on previous efforts in three ways. First, it can recognize over 100 different entity types including social determinants of health, anatomy, risk factors, and adverse events in addition to other commonly used clinical and biomedical entities. Second, the text processing pipeline includes assertion status detection, to distinguish between clinical facts that are present, absent, conditional, or about someone other than the patient. Third, the deep learning models used are more accurate than previously available, leveraging an integrated pipeline of state-of-the-art pretrained named entity recognition models, and improving on the previous best performing benchmarks for assertion status detection. We illustrate extracting trends and insights, e.g. most frequent disorders and symptoms, and most common vital signs and EKG findings, from the COVID-19 Open Research Dataset (CORD-19). The system is built using the Spark NLP library which natively supports scaling to use distributed clusters, leveraging GPUs, configurable and reusable NLP pipelines, healthcare specific embeddings, and the ability to train models to support new entity types or human languages with no code changes.

Motivation & Objective

• To address the challenge of extracting actionable clinical insights from the rapidly growing volume of unstructured COVID-19 scientific literature.
• To improve clinical text mining by recognizing over 100 diverse entity types, including social determinants of health and adverse events, beyond standard biomedical terms.
• To implement a robust assertion status detection system that distinguishes between present, absent, conditional, and other-patient clinical facts.
• To build a scalable, production-ready NLP pipeline using Spark NLP that supports distributed inference and multi-language deployment with minimal code changes.
• To demonstrate the pipeline’s utility in extracting meaningful trends—such as common symptoms, vital signs, and EKG findings—from the CORD-19 dataset.

Proposed method

• The system uses a Bi-LSTM-CNN-Char deep learning architecture for named entity recognition (NER), pre-trained on clinical and biomedical datasets.
• It integrates a Bi-LSTM-based assertion status detection module to classify clinical entities based on contextual relevance to the patient.
• The pipeline is implemented using the Spark NLP library, enabling distributed processing across clusters and GPU acceleration.
• The system supports zero-code extension to new languages (e.g., German, Spanish) and new entity types through fine-tuning with annotated data.
• A unified, composable NLP pipeline is constructed by chaining NER and assertion detection steps on Spark DataFrames, enabling efficient, parallel execution.
• The pipeline is evaluated on the CORD-19 dataset using benchmarking on both local and cluster deployments to measure speed and scalability.

Experimental results

Research questions

• RQ1 Can a scalable NLP pipeline recognize over 100 distinct clinical and biomedical entity types in unstructured COVID-19 literature?
• RQ2 How accurately can assertion status detection distinguish between clinically relevant states such as present, absent, conditional, or someone-else?
• RQ3 To what extent does the Spark NLP pipeline scale in performance when deployed on distributed clusters compared to single-machine execution?
• RQ4 What insights can be extracted from the CORD-19 dataset using structured, normalized clinical entities and their assertion statuses?
• RQ5 Can the pipeline be extended to new languages and entity types with minimal code changes and without retraining from scratch?

Key findings

• The NER model achieves state-of-the-art performance on benchmark datasets, recognizing over 100 entity types including social determinants of health, anatomy, risk factors, and adverse events.
• The assertion status detection model sets a new SOTA benchmark, accurately classifying clinical facts as present, absent, conditional, or related to another person.
• Tokenization speed improved by 20x and NER inference by 3.5x when running on a 10-node cluster compared to a single machine.
• The pipeline successfully extracted meaningful clinical trends from the CORD-19 dataset, including frequent symptoms, disorders, vital signs, and EKG findings.
• The system supports zero-code extension to new languages (e.g., German, Spanish) and new entity types with minimal reconfiguration, thanks to Spark NLP’s modular architecture.
• The integration of pre-trained models into a scalable, distributed pipeline enables efficient, production-grade clinical text mining on large-scale biomedical corpora.