Evaluating shallow and deep learning strategies for the 2018 n2c2 shared task on clinical text classification

Abstract

Objective

Automated clinical phenotyping is challenging because word-based features quickly turn it into a high-dimensional problem, in which the small, privacy-restricted, training datasets might lead to overfitting. Pretrained embeddings might solve this issue by reusing input representation schemes trained on a larger dataset. We sought to evaluate shallow and deep learning text classifiers and the impact of pretrained embeddings in a small clinical dataset.

Materials and Methods

We participated in the 2018 National NLP Clinical Challenges (n2c2) Shared Task on cohort selection and received an annotated dataset with medical narratives of 202 patients for multilabel binary text classification. We set our baseline to a majority classifier, to which we compared a rule-based classifier and orthogonal machine learning strategies: support vector machines, logistic regression, and long short-term memory neural networks. We evaluated logistic regression and long short-term memory using both self-trained and pretrained BioWordVec word embeddings as input representation schemes.

Results

Rule-based classifier showed the highest overall micro F₁ score (0.9100), with which we finished first in the challenge. Shallow machine learning strategies showed lower overall micro F₁ scores, but still higher than deep learning strategies and the baseline. We could not show a difference in classification efficiency between self-trained and pretrained embeddings.

Discussion

Clinical context, negation, and value-based criteria hindered shallow machine learning approaches, while deep learning strategies could not capture the term diversity due to the small training dataset.

Conclusion

Shallow methods for clinical phenotyping can still outperform deep learning methods in small imbalanced data, even when supported by pretrained embeddings.

INTRODUCTION

Background and significance

Clinical narratives stored in electronic health records show considerable variability in format and quality, with their natural language sometimes being described as idiosyncratic.¹ On the one hand, structured data in electronic health records are often created for administrative purposes only and are thus biased toward those diagnoses and procedure codes relevant for billing. On the other hand, semantic tagging of unstructured clinical texts, generally considered the most detailed source of information, is not commonly used and requires prospective planning.² Nevertheless, there is an increasing demand to unlock unstructured data to foster primary and secondary uses.³

One of such secondary uses would be to support observational research such as cohort, cross-sectional, and case-control studies.⁴ A system that analyzes the content of clinical narratives to recruit patients according to selection criteria could help mitigate sampling bias (eg, by creating a matched control group in a case-control study or by drawing both case and control groups from large cohorts).⁵

To process such data for secondary uses, natural language processing (NLP) techniques must be employed to structure the meaning behind human language into a computer-readable representation. However, researchers cannot easily reuse NLP models from the general domain on clinical text due to significant linguistic differences, especially its tendency toward brevity, often characterized as telegraphic style.

A common challenge in building such systems for the clinical domain is the lack of public corpora of annotated clinical narratives due to privacy concerns. While the availability of huge data silos in the general domain sparked the big data revolution by using complex neural networks to model the diversity of the human language with human-like accuracy, the same has not yet happened in small data scenarios, in which models have commonly been trained from scratch. To address that, there has been a rise of interest in transfer learning methods to reuse models trained on large collections in restricted settings with minimal annotation effort.

Transfer learning

One can transfer knowledge from a larger dataset using various approaches, depending on whether the source and target labels are available and what is reused.⁶ One common approach is the so-called feature representation transfer, in which an input representation scheme learned in an unsupervised way in a large corpus is reused in a small annotated dataset.⁶ Nonetheless, Goodfellow et al⁷ argue that the popularity of this approach has declined, because deep learning achieves human-level performance when large labeled datasets are available and Bayesian methods outperform pretraining on small data.

In the NLP area, Mikolov et al⁸ eased feature representation transfer with the release of the word2vec embeddings, trained on around 100 billion words from a Google News corpus. However, clinical text typically shows a low coverage rate in this model due to rare words and misspellings, which has driven the search for alternative input representation schemes.⁹

Bojanowski et al¹⁰ proposed enriching word vectors with subword information to take morphology into account. In parallel, Joulin et al¹¹ released fastText, an efficient implementation of multinomial logistic regression to allow large-scale linear text classification, in clear contrast to the trend of deep learning approaches. More recently, the National Center for Biotechnology Information used fastText to train word embeddings on around 30 million documents from PubMed and the MIMIC-III (Medical Information Mart for Intensive Care)¹² clinical dataset and released BioWordVec.^13,14 Taken together, these resources may help address clinical idiosyncrasies that hinder transfer learning from big data to the clinical domain.

Clinical text classification

Clinical text classification, also referred to as text-based patient phenotyping,^15–17 aims at automatically assigning a finite set of labels to raw clinical text.^18,19 Historically, several strategies have been employed to address this problem, from rule-based systems, known to provide near-optimal results,²⁰ to systems based on machine learning (ML), including support vector machines (SVMs),¹⁸ naive Bayes,²¹ and decision trees.^21–24 More recently, approaches based on deep learning have been studied, including long short-term memory (LSTM) using hand-engineered features,²⁵ as well as convolutional neural networks (CNNs) with rule-extracted trigger phrases²⁶ and word2vec embeddings.²⁷

When introducing BioWordVec, Chen et al¹⁴ showed that CNNs trained with fastText embeddings obtained from PubMed and MIMIC-III improved results of a clinical text classification task when compared with models trained with embeddings from each corpus separately or without embeddings at all, even though no comparison was made to embeddings trained on the target dataset. Roberts²⁸ evaluated the impact of word2vec embeddings trained on multiple corpora (including the target dataset) by applying LSTM and CNN models to 2 downstream tasks: a concept recognition task and a multiclass text classification task, respectively. He showed that models with embeddings trained on several corpora outperformed models with embeddings trained on a single collection but did not consider embeddings with subword information.

2014 i2b2/UTHealth shared task track 2

To promote NLP research in the health domain, the National Center for Biomedical Computing has been organizing since 2006 the Informatics for Integrating Biology and Bedside (i2b2) challenges (https://www.i2b2.org/). The 2014 i2b2/UTHealth shared task track 2 explored the problem of clinical text classification in small data and asked participants to classify patients according to 8 heart disease risk factors (diabetes mellitus, cardiovascular disease [CAD], hypertension, hyperlipidemia, obesity, smoking, family history, and medication).

Participating teams explored several strategies, from rule-based systems to hybrid systems, with different combinations of features and machine learning algorithms.²⁹ The abundance of hybrid systems led the organizers to conclude there was no consensus about which approaches were better suited for the task. They also found that pseudo-tables encoded in text and CAD indicators were especially hard for most of the teams, resulting in comparatively low F₁ scores.

The best team in 2014 obtained an overall micro F₁ score of 0.9276 by reannotating two-thirds of the training corpus and then training SVM models associated with custom-built lexica to classify triggers for each risk factor.³⁰ Documents were preprocessed to identify section headers, negation markers, modality words, and other output from the ConText tool,³¹ but did not use other syntactic and semantic cues. They also showed that such fine-grained annotations could have helped other automated systems.

Kotfila and Uzuner³² performed a systematic comparison of feature spaces, weighting schemes, kernels, and training data sizes regarding the efficiency of SVM classifiers trained on the same data. They reported that minimal feature spaces (only lowercased alphabetic tokens) performed as well as combinations with lexically normalized tokens and semantic concepts extracted via MetaMap;³³tf-idf was not a significant factor to determine efficiency (compared with a count-based weighting scheme); and linear kernels were not statistically significantly worse than radial kernels. Finally, they concluded that larger corpora might not be necessary to achieve high efficiency with SVM models.

In 2018, the Department of Medical Informatics of the Harvard Medical School assumed the organization of what is now called National NLP Clinical Challenges (n2c2) (https://n2c2.dbmi.hms.harvard.edu). The first track focused on the problem of cohort building for clinical trials and framed it as a multilabel binary text classification task.

Research problem

It is challenging to automatically classify clinical text because word-based features quickly turn this task into a high-dimensional problem. On top of that, large corpora are seldom shared due to ethical concerns, while training complex models on small datasets may lead to overfitting. Pretrained embeddings might solve this issue by reusing unsupervised input representation schemes trained on a larger dataset and fine-tuning them using a small annotated dataset.

Therefore, considering (1) previous satisfactory results with shallow methods^11,32 that challenge approaches based on deep learning and (2) the recent availability of pretrained embeddings with subword information in the clinical domain¹⁴ that supports training deeper models in small data scenarios, we decided to participate in the 2018 n2c2 shared task track 1 and use its data to contribute results that may help elucidating these topics. To the best of our knowledge, this is the first study to assess the BioWordVec pretrained embeddings in a text classification task apart from the original work.

Objective

We sought to evaluate shallow and deep learning text classifiers and the impact of pretrained embeddings with subword information in a small clinical dataset.

Hypothesis

We hypothesize that shallow strategies for text classification outperform deep learning strategies in small clinical datasets and that pretrained embeddings increase classification efficiency in the clinical domain.

MATERIALS AND METHODS

Data

We participated in the 2018 n2c2 shared task track 1 and received a small training dataset (70%) with 202 annotated files (patients) containing 887 medical narratives written in English. Two months afterward, we received an extra test set (30%) with 86 new patients (377 narratives) so that we could run our systems and send our results.

Each file had a sequence of 2-5 narratives and was annotated at the patient level with “met” or “not met” for 13 criteria (see Table 1). Of the 13 criteria, 6 were highly imbalanced in the dataset (1 class with <10% of the 202 training samples), 1 semibalanced (with “met” only in approximately 20% of the samples), and the remaining 6 balanced (minority class with at least one-third of the examples). Moreover, 2 criteria were value-dependent: Hba1c and Creatinine.

Table 1.

Overview of the target classification criteria in the 2018 n2c2 shared task track 1

Criterion	Balance	Description
Abdominal	Balanced	History of intra-abdominal surgery.
Advanced-cad	Balanced	Presence of advanced cardiovascular disease.
Alcohol-abuse	Imbalanced	Current weekly alcohol use over recommended limits.
Asp-for-mi	Semibalanced	Use of aspirin to prevent myocardial infarction.
Creatinine	Balanced	Serum creatinine above the normal limit.
Dietsupp-2mos	Balanced	Use of dietary supplements in the last two months.
Drug-abuse	Imbalanced	Drug abuse.
English	Imbalanced	The patient can speak English.
Hba1c	Balanced	Glycated hemoglobin levels between 6.5% and 9.5%.
Keto-1yr	Imbalanced	Ketoacidosis in the last year.
Major-diabetes	Balanced	Major complication due to diabetes.
Makes-decisions	Imbalanced	The patient can make decisions by himself.
Mi-6mos	Imbalanced	Myocardial infarction in the last six months.

Balanced criteria had the minority class with at least one-third of samples; the semibalanced criterion Asp-for-mi had “met” in around 20% of samples and imbalanced criteria had 1 class with <10% of the training samples.

Evaluation metrics

Participating teams were evaluated using precision, recall, and F₁ score across the thirteen criteria and the 2 possible classification outputs: “met” and “not met.” Overall F₁ was the simple mean of the individual F₁ scores for the classes “met” and “not met.” The final metric used for ranking was overall micro F₁ score in the test set. We additionally considered “met” and “not met” as positive and negative outcomes, respectively, and thus also report overall accuracy per criterion.

Preprocessing

We preprocessed input text as follows: (1) removal of spurious whitespaces (as defined by the Java programming language), (2) sentence detection using a customized rule-based algorithm that deals with common abbreviations and artificial new lines, (3) tokenization by the Unicode Text Segmentation algorithm (http://unicode.org/reports/tr29/) as implemented by Lucene 7.5.0 (https://lucene.apache.org/), (4) lowercasing, (5) stop words removal using the SMART³⁴ system’s list of 524 common words, and (6) punctuation removal.

Word embeddings

We used the BioWordVec (https://github.com/ncbi-nlp/BioSentVec) embeddings with subword information pretrained on PubMed and MIMIC-III as available online. To evaluate its impact, we also trained word embeddings from scratch in the target dataset (n2c2) using fastText with the same hyperparameters: window size of 20, learning rate of 0.05, negative sample size of 10, and maximum length of word n-grams set to 6.

Methods

Table 2 shows an overview of the assessed strategies. We evaluated only orthogonal (nonhybrid) strategies to ease comparison among methods. Our baseline was a majority classifier, which always assigns the dominant class seen in training data. We also built a rule-based classifier (RBC) to better understand the data and the noise present therein. We then trained ML-based classifiers using SVMs, logistic regression (LR), and long short-term memory (LSTM) recurrent neural networks. Additionally, we explored both self-trained (SELF) and pretrained embeddings (PRE) with subword information for approaches using word embeddings as the input representation scheme (LR and LSTM). We contributed the results of RBC, SVMs, and a variant of LSTM (described in Supplementary Appendix A) as official runs for participation at the n2c2 shared task.

Table 2.

Overview of the evaluated methods and their characteristics

Acronym	Classification method	Word embeddings
Baseline	Majority	N/A
RBC	Rule-based classifier	N/A
SVM	Support vector machine	N/A
SELF-LR	Logistic regression	Self-trained
PRE-LR	Logistic regression	Pretrained
SELF-LSTM	Long short-term memory	Self-trained
PRE-LSTM	Long short-term memory	Pretrained

Pretrained word embeddings were obtained from BioWordVec.

N/A: not applicable.

Our extensible Java framework is available on GitHub at https://github.com/bst-mug/n2c2 under the open-source Apache License version 2.

Rule-based classifier

We developed a rule-based approach using both regular expressions and textual markers, extended in 4 criteria (Advanced-cad, Asp-for-mi, Major-diabetes, and Mi-6mos) with negation and context detection. For the value-dependent criteria Creatinine and Hba1c, we extracted the corresponding value using a regular expression and compared it to manually defined thresholds, namely $1.4<\text{creatinine}<\hspace{0.17em}10$ and $6.5\le \mathit{hba}1c\le 9.5$. For the remaining criteria, we manually identified typical text snippets from the training set (such as “elevated creatinine”) that, when found, would classify a patient for a given criterion. We enriched these text markers with negative “lookaround” regular expressions to invalidate the marker when it referred to (1) a negated context (eg, “denies ischemia”), (2) drug allergies (“allergy to aspirin”), (3) distant history (“STEMI in 2008”); or (4) family history (“FH with NSTEMI”).

Support vector machines

We explored SVMs trained on a bag-of-words representation of the input documents using tf-idf (term frequency – inverse document frequency). As text is typically linearly separable,¹⁸ we then applied SVM with a linear kernel. We used the Weka³⁵ 3.8.2 framework with a LibSVM³⁶ wrapper to train the SVM classifier. We could not significantly improve the overall micro F₁ score in the training set by using any of the following: (1) cost hyperparameter optimization³⁷ (default: 1), (2) a lower number of features to avoid overfitting (default: 1000), or (3) L2 normalization on the SVM objective function. Thus, we kept the default values wherever possible.

Logistic regression

LR is a linear machine learning method that is equivalent to a single-layer perceptron (a single-layer feedforward neural network) with a logistic function as output instead of a step function. We trained LR using fastText with 100 epochs, learning rate of 0.50, window size of 5, cross entropy loss function, and a single thread to make results reproducible. We represented the input text as the average of either self-trained word embeddings (SELF-LR) or pretrained BioWordVec embeddings (PRE-LR).

Long short-term memory

We explored a deep learning approach based on a recurrent neural network composed of LSTM cells,³⁸ a type of architecture that is generally used to model time series events,²⁵ but can also be used to model natural language for domain-specific tasks.^39,40 Our network had a single LSTM layer with 64 cells and a RNN output layer with sigmoid activation and cross entropy loss function. We employed the Deeplearning4j (https://deeplearning4j.org) 0.9.1 framework to model the LSTM neural network. We used Adam⁴¹ for gradient-based optimization with a learning rate of 0.02 and trained the network for 25 epochs with a dropout rate of 50% to prevent overfitting. Similar to LR, we represented the input text as a sequence of either self-trained word embeddings (SELF-LSTM) or pretrained BioWordVec embeddings (PRE-LSTM).

RESULTS

Tables 3 and 4 depict overall F₁ score and accuracy per criterion, on the test set of the proposed methods when compared with the baseline, a majority classifier. We present detailed results by target class in Supplementary Appendix B.

Table 3.

Overall F₁ score per criterion on the test set of the evaluated strategies when compared with the baseline, a majority classifier

Criterion	Baseline	RBC	SVM	SELF-LR	PRE-LR	SELF-LSTM	PRE-LSTM
Abdominal	0.3944	0.8720	0.6028	0.5681	0.5959	0.4930	0.5146
Advanced-cad	0.3435	0.7902	0.7281	0.7109	0.6838	0.5865	0.4788
Alcohol-abuse	0.4911	0.4881	0.4911	0.4911	0.4911	0.4911	0.4881
Asp-for-mi	0.4416	0.7095	0.6063	0.5962	0.6060	0.4948	0.4416
Creatinine	0.4189	0.8071	0.6532	0.7180	0.7399	0.4788	0.5322
Dietsupp-2mos	0.3385	0.9185	0.5814	0.6150	0.6261	0.5903	0.4640
Drug-abuse	0.4911	0.6910	0.4911	0.4911	0.4881	0.4850	0.4881
English	0.4591	0.8644	0.4591	0.4591	0.4591	0.5253	0.5176
Hba1c	0.3723	0.9382	0.6267	0.5393	0.5770	0.4682	0.5137
Keto-1yr	0.5000	0.5000	0.5000	0.5000	0.5000	0.5000	0.5000
Major-diabetes	0.3333	0.8369	0.7555	0.7518	0.7420	0.4883	0.5435
Makes-decisions	0.4911	0.4911	0.4911	0.4911	0.4911	0.4911	0.4881
Mi-6mos	0.4756	0.8752	0.6815	0.4756	0.4756	0.4658	0.4691
Overall (macro)	0.4270	0.7525	0.5899	0.5698	0.5751	0.5045	0.4953
Overall (micro)	0.7608	0.9100	0.8035	0.8017	0.8063	0.7362	0.7377

Overall F₁ score is the simple mean of the F₁ scores for the classes “met” and “not met.”

PRE-LR: pretrained logistic regression; PRE-LSTM: pretrained long short-term memory; RBC: rule-based classifier; SELF-LR: self-trained logistic regression; SELF-LSTM: self-trained long short-term memory; SVM: support vector machine.

Table 4.

Overall accuracy per criterion on the test set of the evaluated strategies when compared with the baseline, a majority classifier

Criterion	Baseline	RBC	SVM	SELF-LR	PRE-LR	SELF-LSTM	PRE-LSTM
Abdominal	0.6512	0.8837	0.6512	0.6279	0.6628	0.5233	0.6047
Advanced-cad	0.5233	0.7907	0.7326	0.7209	0.6977	0.5465	0.5465
Alcohol-abuse	0.9651	0.9535	0.9651	0.9651	0.9651	0.9535	0.9651
Asp-for-mi	0.7907	0.8605	0.7558	0.7674	0.7791	0.7442	0.7791
Creatinine	0.7209	0.8372	0.7209	0.7674	0.7907	0.5698	0.6395
Dietsupp-2mos	0.5116	0.9186	0.5814	0.6163	0.6279	0.6047	0.4651
Drug-abuse	0.9651	0.9651	0.9651	0.9651	0.9535	0.9651	0.9651
English	0.8488	0.9419	0.8488	0.8488	0.8488	0.8372	0.8488
Hba1c	0.5930	0.9419	0.6512	0.5814	0.6047	0.6047	0.5465
Keto-1yr	1.0000	1.0000	1.0000	1.0000	1.0000	1.0000	1.0000
Major-diabetes	0.5000	0.8372	0.7558	0.7558	0.7442	0.5349	0.5465
Makes-decisions	0.9651	0.9651	0.9651	0.9651	0.9651	0.9651	0.9651
Mi-6mos	0.9070	0.9651	0.9302	0.9070	0.9070	0.9070	0.9070
Overall	0.7648	0.9123	0.8095	0.8068	0.8113	0.7504	0.7522

Overall accuracy is calculated with “met” and “not met” being considered as positive and negative outcomes, respectively.

PRE-LR: pretrained logistic regression; PRE-LSTM: pretrained long short-term memory; RBC: rule-based classifier; SELF-LR: self-trained logistic regression; SELF-LSTM: self-trained long short-term memory; SVM: support vector machine.

Baseline

As expected, individual F₁ scores for the majority classifier did not exceed 0.5000, but due to an imbalance among target classes (the most extreme example being Keto-1yr), accuracy and overall micro F₁ scores reached high values, thereby setting the baseline at F₁ = 0.7608 and A = 0.7648.

Rule-based classifier

With respect to individual F₁ scores and accuracies, the RBC showed better efficiency than the baseline for every criterion except Alcohol-abuse, on which the RBC had a single false positive due to a missed negation. Rules improved Dietsupp-2mos the most, with absolute accuracy (F₁) increase of 0.4070 (0.5800). Imbalanced criteria such as Alcohol-abuse and Makes-decisions presented the worst F₁ scores; however, owing to micro-averaging, such criteria did not significantly affect the overall efficiency. Conversely, we observed the lowest accuracies in the criteria Advanced-cad, Creatinine, and Major-diabetes.

Support vector machines

Considering individual F₁ scores, SVMs performed equal or better than the baseline on each criterion; however, considering individual accuracies, SVMs performed worse than the baseline for Asp-for-mi due to an increase in false negatives (shown as a decrease in recall for “met” and precision for “not met” in Supplementary Table B3). Imbalanced criteria such as English, Alcohol-abuse, and Makes-decisions presented the lowest F₁ scores. Considering accuracy, Dietsupp-2mos had the lowest results, followed by Abdominal and Hba1c.

Logistic regression

Both SELF-LR and PRE-LR showed equal or better F₁ scores than the baseline for all criteria (except Drug-Abuse for PRE-LR); considering accuracy, SELF-LR (PRE-LR) had better results on 4 (6) criteria: Advanced-cad, Creatinine, Dietsupp-2mos, and Major-diabetes (Abdominal, Advanced-cad, Creatinine, Dietsupp-2mos, Hba1c, and Major-diabetes); slightly worse results on 3 (2) criteria: Abdominal, Asp-for-mi, and Hba1c (Asp-for-mi and Drug-abuse); and was in a tie in the remaining 6 (5) criteria. Similar to SVM, imbalanced criteria had the lowest F₁ scores and Hba1c showed the lowest accuracy results for both SELF-LR and PRE-LR.

Long short-term memory

With respect to individual F₁ scores, SELF-LSTM (PRE-LSTM) showed worse results than the baseline in the criteria Drug-Abuse and Mi-6mos (Alcohol-Abuse, Drug-Abuse, Makes-decisions, and Mi-6mos). Conversely, with respect to the accuracy, SELF-LSTM (PRE-LSTM) showed better accuracy results than the baseline in the criteria Advanced-cad, Dietsupp-2mos, Hba1c, and Major-diabetes (Advanced-cad and Major-diabetes). As expected, SELF-LSTM (PRE-LSTM) mostly improved F₁ score for balanced criteria, with Dietsupp-2mos (Major-diabetes) showing the most substantial absolute increase in F₁ score: 0.2518 (0.2102).

DISCUSSION

The results in the previous section showed that RBC had the highest classification efficiency, followed by shallow methods (SVM and LR), which had similar scores among themselves and above the baseline. Apart from overall macro F₁ score, the deep learning method (LSTM) showed results worse than the baseline. We could not show however a significant difference in classification efficiency of using pretrained word embeddings when compared with embeddings trained on the n2c2 dataset.

We further analyzed false positives and false negatives to obtain deeper insights about the data, discuss our limitations, and propose future work.

False positives and negatives analysis

A high textual diversity compared with the amount of available balanced training data contributed to setting the limits of our approaches. For example, although in our context “chest pain” could be considered a synonym for angina (and thus a predictor for Advanced-cad), it was commonly used in negated sentences such as “denies […] chest pain,” “negative for […] chest pain,” and “chest pain free.” To keep our methods orthogonal, we did not explore the impact of rule-based or automated negation detection.

Similarly, context also played a role. For instance, “renal transplant” was a common indicator for an intra-abdominal surgery (and thus a predictor for Abdominal); we found, however, that its meaning was sometimes changed by nearby words, as in the excerpt “postponing her renal transplant.” Even though a bag-of-bigrams approach might have helped, training such a model would have needed more data due to the larger dimensionality.

A specific case of context is family history. It showed more prominently in the selection criteria Makes-decisions (in which “dementia” would have been an important feature if it had not been used in sentences such as “Father had dementia”) and Mi-6mos (in which “MI” was an indicator for myocardial infarction not only for the patient but also for relatives, such as in the sentence “father died of MI”).

Another common source of errors for our strategies were value-based criteria such as Hba1c and Creatinine. To handle numbers in ML strategies without rule-based normalization, we would have needed (1) a large amount of data to capture all values and (2) a nonlinear approach to model both a low and a high threshold. Conversely, we also needed to define a clear classification threshold for rule-based approaches, for which we found inconsistent examples in the training set (eg, creatinine in serum for values close to 1.4 mg/dL).

Limitations

It is known that rule-based approaches do not generalize well and present a maintenance burden. To avoid that, we kept our markers to the bare minimum and used regular expressions only when needed. Together with automated negation and context detection, we believe our method could be reused in other English-speaking institutions with minimal effort. Meanwhile, our shallow approaches (SVM and LR) are language and domain independent; therefore, their reuse in other institutions could be even simpler.

Likewise, we selected features for the SVM classifier based on word frequency only. A better approach would have been to employ output-based or statistical-based methods such as chi-square and information gain. The lack of feature selection may have been the reason the criterion Dietsupp-2mos had the lowest SVM results: there is a vast quantity of dietary supplements, often unique in the collection. Nevertheless, our LR approach used a 200-dimensional word vector as input representation, a vector space which should keep semantically-close words nearby, and similarly showed low results for the mentioned criteria.

Furthermore, we did not completely investigate the impact of text normalization. Even though our ML-based methods employed basic tokenization, lowercasing, and stemming, we did not resolve short forms nor misspellings. Our manual data inspection and the high RBC results showed that these were not crucial issues in this dataset, but other domains and languages might require further preprocessing.

Finally, some documents had pseudo-tables for laboratory data, from which we could not extract the proper pieces of information. Our manual analysis of training data showed, however, that physicians would usually emphasize abnormal laboratory results in the text and thus we could capture it using straightforward strategies. In a real clinical setting, structured data may be directly accessible from laboratory systems and thus constitute a better exploratory avenue.

Future work

Future work might explore the dependence between criteria (eg, an episode of myocardial infarction in the last 6 months would trigger not only Mi-6mos, but potentially also Advanced-cad and Asp-for-mi). Experienced NLP researchers might also experiment with sentence parsing to unlock the meaning behind sentences such as “BUN and creatinine were 32 and 1.2.”

We opted for independent strategies to ease method comparison and promote interpretability in the shared-task scenario. A real system could benefit from a hybrid approach, using (1) an ensemble of methods (eg, weighted linear combination of signals provided by each approach), (2) stacking strategies (eg, SVM trained on top of count features extracted by the rule-based approach), or (3) a mixed approach (eg, RBCs for imbalanced and value-based criteria and ML-based classifiers for more balanced and complex criteria).

Finally, recent developments in transfer learning that allow reuse of full NLP models trained on large data may help the clinical domain more than input pretraining alone. Special attention should be devoted to Google’s BERT⁴² and Universal Sentence Encoder,⁴³ built on top of the ULMFiT⁴⁴ and the ELMo⁴⁵ models, with pretrained models released not only for the general domain but quite recently also for the clinical domain, the so-called ClinicalBERT.⁴⁶

CONCLUSION

We participated in the 2018 n2c2 shared task track 1 and used its dataset to evaluate shallow and deep learning strategies and the impact of recently released pretrained embeddings for multilabel text classification in small clinical data. We also built a rule-based classifier to provide us with a deeper understanding of the underlying data. We submitted to the shared task the results of 3 orthogonal strategies (RBC, SVM, and a variant of LSTM) to support method comparison. We also contributed our extensible Java framework to the community under an open-source license.

Our rule-based classifier showed the highest overall micro F₁ score of 0.9100, with which we finished first in the shared task. Shallow strategies showed lower overall micro F₁ scores (SVM: 0.8035, SELF-LR: 0.8017, PRE-LR: 0.8063), but still higher than the deep learning strategy (SELF-LSTM: 0.7362, PRE-LSTM: 0.7377) and the baseline (0.7608) set to a majority classifier. We could not show however a significant difference in classification efficiency of using pretrained word embeddings when compared with embeddings trained on the n2c2 target dataset.

Together with the inter-rater agreement scores released by the task organizers, our top-ranking RBC contributed to practical upper bounds for each selection criteria, which might guide other researchers with directions for further improvement. It also provides the community with a reliable method for clinical phenotyping, which can also be reused in a fine-grained way and thus allow further experiments with deep learning approaches.

We also discussed that clinical context, negation, and value-based criteria hindered shallow machine learning approaches. Even though pretrained word embeddings could not alleviate these issues, we see potential in novel transfer learning techniques that allow reuse not only of feature representation schemes but also full classification models, thus bridging the gap between big and small data.

Taken together, our study suggests that rule-based and shallow methods for clinical phenotyping can still outperform deep learning methods in small imbalanced data, even when augmented with pretrained embeddings with subword information.

FUNDING

MO is funded by the Brazilian National Research Council - CNPq (project number 206892/2014-4).

AUTHOR CONTRIBUTIONS

MO analyzed false positives/negatives and implemented the code framework and the rule-based, support vector machine, and logistic regression approaches. AK analyzed the data, implemented zoning, and presented the work at the conference. MK implemented the long short-term memory and the bidirectional long short-term memory approach. ZK analyzed false positives/negatives, revised the manuscript, and gave overall clinical feedback.

SUPPLEMENTARY MATERIAL

Supplementary material is available at Journal of the American Medical Informatics Association online.

Conflict of interest statement

None declared.