Anthony Rios

An empirical evaluation of supervised learning approaches in assigning diagnosis codes to electronic medical records

BACKGROUND Diagnosis codes are assigned to medical records in healthcare facilities by trained coders by reviewing all physician authored documents associated with a patients visit. This is a necessary and complex task involving coders adhering to coding guidelines and coding all assignable codes. With the popularity of electronic medical records (EMRs), computational approaches to code assignment have been proposed in the recent years. However, most efforts have focused on single and often short clinical narratives, while realistic scenarios warrant full EMR level analysis for code assignment. OBJECTIVE We evaluate supervised learning approaches to automatically assign international classification of diseases (ninth revision) - clinical modification (ICD-9-CM) codes to EMRs by experimenting with a large realistic EMR dataset. The overall goal is to identify methods that offer superior performance in this task when considering such datasets. METHODS We use a dataset of 71,463 EMRs corresponding to in-patient visits with discharge date falling in a two year period (2011-2012) from the University of Kentucky (UKY) Medical Center. We curate a smaller subset of this dataset and also use a third gold standard dataset of radiology reports. We conduct experiments using different problem transformation approaches with feature and data selection components and employing suitable label calibration and ranking methods with novel features involving code co-occurrence frequencies and latent code associations. RESULTS Over all codes with at least 50 training examples we obtain a micro F-score of 0.48. On the set of codes that occur at least in 1% of the two year dataset, we achieve a micro F-score of 0.54. For the smaller radiology report dataset, the classifier chaining approach yields best results. For the smaller subset of the UKY dataset, feature selection, data selection, and label calibration offer best performance. CONCLUSIONS We show that datasets at different scale (size of the EMRs, number of distinct codes) and with different characteristics warrant different learning approaches. For shorter narratives pertaining to a particular medical subdomain (e.g., radiology, pathology), classifier chaining is ideal given the codes are highly related with each other. For realistic in-patient full EMRs, feature and data selection methods offer high performance for smaller datasets. However, for large EMR datasets, we observe that the binary relevance approach with learning-to-rank based code reranking offers the best performance. Regardless of the training dataset size, for general EMRs, label calibration to select the optimal number of labels is an indispensable final step.

Convolutional neural networks for biomedical text classification: application in indexing biomedical articles

Building high accuracy text classifiers is an important task in biomedicine given the wealth of information hidden in unstructured narratives such as research articles and clinical documents. Due to large feature spaces, traditionally, discriminative approaches such as logistic regression and support vector machines with n-gram and semantic features (e.g., named entities) have been used for text classification where additional performance gains are typically made through feature selection and ensemble approaches. In this paper, we demonstrate that a more direct approach using convolutional neural networks (CNNs) outperforms several traditional approaches in biomedical text classification with the specific use-case of assigning medical subject headings (or MeSH terms) to biomedical articles. Trained annotators at the national library of medicine (NLM) assign on an average 13 codes to each biomedical article, thus semantically indexing scientific literature to support NLMs PubMed search system. Recent evidence suggests that effective automated efforts for MeSH term assignment start with binary classifiers for each term. In this paper, we use CNNs to build binary text classifiers and achieve an absolute improvement of over 3% in macro F-score over a set of selected hard-to-classify MeSH terms when compared with the best prior results on a public dataset. Additional experiments on 50 high frequency terms in the dataset also show improvements with CNNs. Our results indicate the strong potential of CNNs in biomedical text classification tasks.

Supervised Extraction of Diagnosis Codes from EMRs: Role of Feature Selection, Data Selection, and Probabilistic Thresholding

Extracting diagnosis codes from medical records is a complex task carried out by trained coders by reading all the documents associated with a patients visit. With the popularity of electronic medical records (EMRs), computational approaches to code extraction have been proposed in the recent years. Machine learning approaches to multi-label text classification provide an important methodology in this task given each EMR can be associated with multiple codes. In this paper, we study the the role of feature selection, training data selection, and probabilistic threshold optimization in improving different multi-label classification approaches. We conduct experiments based on two different datasets: a recent gold standard dataset used for this task and a second larger and more complex EMR dataset we curated from the University of Kentucky Medical Center. While conventional approaches achieve results comparable to the state-of-the-art on the gold standard dataset, on our complex in-house dataset, we show that feature selection, training data selection, and probabilistic thresholding provide significant gains in performance.

Extracting Drug-Drug Interactions with Word and Character-Level Recurrent Neural Networks

Drug-drug interactions (DDIs) are known to be responsible for nearly a third of all adverse drug reactions. Hence several current efforts focus on extracting signal from EMRs to prioritize DDIs that need further exploration. To this end, being able to extract explicit mentions of DDIs in free text narratives is an important task. In this paper, we explore recurrent neural network (RNN) architectures to detect and classify DDIs from unstructured text using the DDIExtraction dataset from the SemEval 2013 (task 9) shared task. Our methods are in line with those used in other recent deep learning efforts for relation extraction including DDI extraction. However, to our knowledge, we are the first to investigate the potential of character-level RNNs (Char-RNNs) for DDI extraction (and relation extraction in general). Furthermore, we explore a simple but effective model bootstrapping method to (a). build model averaging ensembles, (b). derive confidence intervals around mean micro-F scores (MMF), and (c). assess the average behavior of our methods. Without any rule based filtering of negative examples, a popular heuristic used by most earlier efforts, we achieve an MMF of 69.13. By adding simple replicable heuristics to filter negative instances we are able to achieve an MMF of 70.38. Furthermore, our best ensembles produce micro F-scores of 70.81 (without filtering) and 72.13 (with filtering), which are superior to metrics reported in published results. Although Char-RNNs turnout to be inferior to regular word based RNN models in overall comparisons, we find that ensembling models from both architectures results in nontrivial gains over simply using either alone, indicating that they complement each other.

Ordinal Convolutional Neural Networks for Predicting RDoC Positive Valence Psychiatric Symptom Severity Scores

BACKGROUND The CEGS N-GRID 2016 Shared Task in Clinical Natural Language Processing (NLP) provided a set of 1000 neuropsychiatric notes to participants as part of a competition to predict psychiatric symptom severity scores. This paper summarizes our methods, results, and experiences based on our participation in the second track of the shared task. OBJECTIVE Classical methods of text classification usually fall into one of three problem types: binary, multi-class, and multi-label classification. In this effort, we study ordinal regression problems with text data where misclassifications are penalized differently based on how far apart the ground truth and model predictions are on the ordinal scale. Specifically, we present our entries (methods and results) in the N-GRID shared task in predicting research domain criteria (RDoC) positive valence ordinal symptom severity scores (absent, mild, moderate, and severe) from psychiatric notes. METHODS We propose a novel convolutional neural network (CNN) model designed to handle ordinal regression tasks on psychiatric notes. Broadly speaking, our model combines an ordinal loss function, a CNN, and conventional feature engineering (wide features) into a single model which is learned end-to-end. Given interpretability is an important concern with nonlinear models, we apply a recent approach called locally interpretable model-agnostic explanation (LIME) to identify important words that lead to instance specific predictions. RESULTS Our best model entered into the shared task placed third among 24 teams and scored a macro mean absolute error (MMAE) based normalized score (100·(1-MMAE)) of 83.86. Since the competition, we improved our score (using basic ensembling) to 85.55, comparable with the winning shared task entry. Applying LIME to model predictions, we demonstrate the feasibility of instance specific prediction interpretation by identifying words that led to a particular decision. CONCLUSION In this paper, we present a method that successfully uses wide features and an ordinal loss function applied to convolutional neural networks for ordinal text classification specifically in predicting psychiatric symptom severity scores. Our approach leads to excellent performance on the N-GRID shared task and is also amenable to interpretability using existing model-agnostic approaches.

Analyzing the Moving Parts of a Large-Scale Multi-label Text Classification Pipeline: Experiences in Indexing Biomedical Articles

Medical subject headings (MeSH) is a controlled hierarchical vocabulary used by the National Library of Medicine (NLM) to index biomedical articles. In the 2014 version of MeSH terminology there are a total of 27,149 terms. Librarians at the NLM tag each biomedical article to be indexed for the PubMed literature search system with terms from MeSH. This means the human indexers look at each articles full text and index it with a small set of descriptors, 13 on average, from over 27,000 descriptors available in MeSH. There have been many recent attempts to automate this process focused on using the article title and abstract text to predict MeSH terms for the corresponding article. There has also been an open automated biomedical indexing challenge, BioASQ [1], that started in 2013. The best general supervised learning framework in these challenges has been a pipeline with four different components: 1. pre-processing and feature extraction; 2. employing the binary relevance and/or nearest neighbor approaches to select a set of candidate terms; 3. ranking these candidate terms using corresponding informative features; and 4. applying label calibration to dynamically predict the number of top terms to be included in the final selection for the current instance. The specific details in how each of these components is implemented determines the performance variations of various entries in the challenge. In this paper, we analyze these moving parts of the MeSH indexing multilabel classification pipeline with experiments involving different combinations. Our best combination achieves ≈ 1% increase in micro F-score compared with the top performing team across the five weeks of the final batch of the BioASQ 2014 challenge. The main take away from our efforts is that small improvements/modifications to different components of the pipeline can offer moderate improvements to the overall performance of the method. Our experiences show that, at least thus far, top performances have resulted mostly due to these improvements rather than drastic changes of the core methodology.

Data and systems for medication-related text classification and concept normalization from Twitter: Insights from the Social Media Mining for Health (SMM4H) 2017 shared task

Abstract Objective We executed the Social Media Mining for Health (SMM4H) 2017 shared tasks to enable the community-driven development and large-scale evaluation of automatic text processing methods for the classification and normalization of health-related text from social media. An additional objective was to publicly release manually annotated data. Materials and Methods We organized 3 independent subtasks: automatic classification of self-reports of 1) adverse drug reactions (ADRs) and 2) medication consumption, from medication-mentioning tweets, and 3) normalization of ADR expressions. Training data consisted of 15 717 annotated tweets for (1), 10 260 for (2), and 6650 ADR phrases and identifiers for (3); and exhibited typical properties of social-media-based health-related texts. Systems were evaluated using 9961, 7513, and 2500 instances for the 3 subtasks, respectively. We evaluated performances of classes of methods and ensembles of system combinations following the shared tasks. Results Among 55 system runs, the best system scores for the 3 subtasks were 0.435 (ADR class F1-score) for subtask-1, 0.693 (micro-averaged F1-score over two classes) for subtask-2, and 88.5% (accuracy) for subtask-3. Ensembles of system combinations obtained best scores of 0.476, 0.702, and 88.7%, outperforming individual systems. Discussion Among individual systems, support vector machines and convolutional neural networks showed high performance. Performance gains achieved by ensembles of system combinations suggest that such strategies may be suitable for operational systems relying on difficult text classification tasks (eg, subtask-1). Conclusions Data imbalance and lack of context remain challenges for natural language processing of social media text. Annotated data from the shared task have been made available as reference standards for future studies (http://dx.doi.org/10.17632/rxwfb3tysd.1).

Extracting chemical–protein relations with ensembles of SVM and deep learning models

Abstract Mining relations between chemicals and proteins from the biomedical literature is an increasingly important task. The CHEMPROT track at BioCreative VI aims to promote the development and evaluation of systems that can automatically detect the chemical–protein relations in running text (PubMed abstracts). This work describes our CHEMPROT track entry, which is an ensemble of three systems, including a support vector machine, a convolutional neural network, and a recurrent neural network. Their output is combined using majority voting or stacking for final predictions. Our CHEMPROT system obtained 0.7266 in precision and 0.5735 in recall for an F-score of 0.6410 during the challenge, demonstrating the effectiveness of machine learning-based approaches for automatic relation extraction from biomedical literature and achieving the highest performance in the task during the 2017 challenge. Database URL: http://www.biocreative.org/tasks/biocreative-vi/track-5/

Generalizing biomedical relation classification with neural adversarial domain adaptation

Motivation Creating large datasets for biomedical relation classification can be prohibitively expensive. While some datasets have been curated to extract protein-protein and drug-drug interactions (PPIs and DDIs) from text, we are also interested in other interactions including gene-disease and chemical-protein connections. Also, many biomedical researchers have begun to explore ternary relationships. Even when annotated data are available, many datasets used for relation classification are inherently biased. For example, issues such as sample selection bias typically prevent models from generalizing in the wild. To address the problem of cross-corpora generalization, we present a novel adversarial learning algorithm for unsupervised domain adaptation tasks where no labeled data are available in the target domain. Instead, our method takes advantage of unlabeled data to improve biased classifiers through learning domain-invariant features via an adversarial process. Finally, our method is built upon recent advances in neural network (NN) methods. Results We experiment by extracting PPIs and DDIs from text. In our experiments, we show domain invariant features can be learned in NNs such that classifiers trained for one interaction type (protein-protein) can be re-purposed to others (drug-drug). We also show that our method can adapt to different source and target pairs of PPI datasets. Compared to prior convolutional and recurrent NN-based relation classification methods without domain adaptation, we achieve improvements as high as 30% in F1-score. Likewise, we show improvements over state-of-the-art adversarial methods. Availability and implementation Experimental code is available at https://github.com/bionlproc/adversarial-relation-classification. Supplementary information Supplementary data are available at Bioinformatics online.