Semi-Supervised Natural Language Approach for Fine-Grained Classification of Medical Reports

Abstract

Although machine learning has become a powerful tool to augment doctors in clinical analysis, the immense amount of labeled data that is necessary to train supervised learning approaches burdens each development task as time and resource intensive. The vast majority of dense clinical information is stored in written reports, detailing pertinent patient information. The challenge with utilizing natural language data for standard model development is due to the complex and unstructured nature of the modality. In this research, a model pipeline was developed to utilize an unsupervised approach to train an encoder-language model, a bidirectional recurrent neural network, to generate document encodings; which then can be used as features passed into a decoder-classifier model that requires magnitudes less labeled data than previous approaches to differentiate between fine-grained disease classes accurately. The language model was trained on unlabeled radiology reports from the Massachusetts General Hospital Radiology Department (n=218,159) and terminated with a loss of 1.62 and a word prediction accuracy of 62%. The classification models were trained on three labeled datasets of head CT studies of reported patients, presenting large vessel occlusion (n=1403), acute ischemic strokes (n=331), and intracranial hemorrhage (n=4350), to identify a variety of different findings directly from the radiology report data; resulting in AUCs of 0.98, 0.95, and 0.99, respectively, for the large vessel occlusion, acute ischemic stroke, and intracranial hemorrhage datasets. The output encodings are able to be used in conjunction with imaging data, to create models that can process a multitude of different modalities. The ability to automatically extract relevant features from textual data allows for faster model development and integration of Preprint. Under review.

A Systematic Review of Natural Language Processing Applied to Radiology Reports

“…Banerjee et al [Banerjee et al, 2017] found that there was not much difference between a uni-gram approach and a Word2vec embedding, hypothesising this was due to their narrow domain, intracranial haemorrhage. However, the NLP research field has seen a move towards bi-directional encoder representations from transformers (BERT) based embedding models not reflected in our analysis, with only one study using BERT generated embeddings [Deshmukh et al, 2019]. Embeddings from BERT are thought to be superior as they can deliver better contextual representations and result in improved task performance.…”

“…This category was not found in Pons' work. Methods considered a range of conditions including intracranial haemorrhage [Jnawali et al, 2019, Banerjee et al, 2017, aneurysms [K los et al, 2018], brain metastases [Deshmukh et al, 2019], ischaemic stroke , Garg et al, 2019, and several classified on types and severity of conditions e.g., [Deshmukh et al, 2019, Shin et al, 2017, Wheater et al, 2019, Gorinski et al, 2019. Studies focused on breast imaging considered aspects such as predicting lesion malignancy from BI-RADS descriptors [Bozkurt et al, 2016], breast cancer subtypes [Patel et al, 2017], and extracting or inferring BI-RADS categories, such as [Banerjee et al, 2019a, Miao et al, 2018.…”