Limitations of Transformers on Clinical Text Classification
Not Available
Engineering topics
Publications and source records attributed to Alawad, Mohammed.
Not Available
Abstract Background Automated text classification has many important applications in the clinical setting; however, obtaining labelled data for training machine learning and deep learning models is often difficult and expensive. Active learning techniques may mitigate this challenge by reducing the amount of labelled data required to effectively train a model. In this study, we analyze the effectiveness of 11 active learning algorithms on classifying subsite and histology from cancer pathology reports using a Convolutional Neural Network as the text classification model. Results We compare the performance of each active learning strategy using two differently sized datasets and two different classification tasks. Our results show that on all tasks and dataset sizes, all active learning strategies except diversity-sampling strategies outperformed random sampling, i.e., no active learning. On our large dataset (15K initial labelled samples, adding 15K additional labelled samples each iteration of active learning), there was no clear winner between the different active learning strategies. On our small dataset (1K initial labelled samples, adding 1K additional labelled samples each iteration of active learning), marginal and ratio uncertainty sampling performed better than all other active learning techniques. We found that compared to random sampling, active learning strongly helps performance on rare classes by focusing on underrepresented classes. Conclusions Active learning can save annotation cost by helping human annotators efficiently and intelligently select which samples to label. Our results show that a dataset constructed using effective active learning techniques requires less than half the amount of labelled data to achieve the same performance as a dataset constructed using random sampling.
This paper presents a multimodal data representation to improve the performance of deep learning models for extracting cancer key characteristics from unstructured text in pathology reports. Specifically, in addition to using the text as the input to deep learning models, we use concept unique identifiers (CUIs) as another source of information to the models. We analyze the performance of different text and CUI data representations, including word embeddings and bag of embeddings (BOE), with a convolutional neural network (CNN) and a fully connected multilayer perceptron neural network (MLP-NN). The high level document embeddings from text and CUI inputs are combined by concatenating them and then applying a classifier. The model is used for extracting cancer subsite and histology from pathology reports. These two classification tasks have a large number of labels, i.e. 317 for subsite and 556 for histology, with extreme class imbalance. We compare the performance of the developed DL models across the two tasks based on micro- and macro-F1 scores. The evaluation shows that a multi-channel DL model that utilizes text represented by word embeddings and CUIs represented by BOE outperforms other DL models. Also, this approach significantly improves the model performance on low prevalence classes.
Not Available
Individual electronic health records (EHRs) and clinical reports are often part of a larger sequence—for example, a single patient may generate multiple reports over the trajectory of a disease. In applications such as cancer pathology reports, it is necessary not only to extract information from individual reports, but also to capture aggregate information regarding the entire cancer case based off case-level context from all reports in the sequence. In this paper, we introduce a simple modular add-on for capturing case-level context that is designed to be compatible with most existing deep learning architectures for text classification on individual reports. We test our approach on a corpus of 431,433 cancer pathology reports, and we show that incorporating case-level context significantly boosts classification accuracy across six classification tasks—site, subsite, laterality, histology, behavior, and grade. We expect that with minimal modifications, our add-on can be applied towards a wide range of other clinical text-based tasks.
Rapidly expanding, new telecommunications infrastructure based on 5G technologies will disrupt and transform how we design, build, operate, and optimize scientific infrastructure and the experiments and services enabled by that infrastructure, from continental-scale sensor networks to centralized scientific user facilities, from intelligent Internet of Things devices to supercomputers. Concurrently, 5G will introduce, or exacerbate, challenges related to protecting infrastructure and associated scientific data as well as to fully leveraging opportunities related to expanded infrastructure scale and complexity. The U.S. Department of Energy (DOE) Office of Science operates scientific infrastructure, supporting some of the nation’s most advanced intellectual discoveries, spanning the country and including 30 world-class user facilities from supercomputers to accelerators. Along with field experiments and remote observatories, every aspect of DOE’s scientific enterprise will be affected by 5G, which amounts to a complete renovation of the underpinnings of the nation’s information infrastructure. In this report we explore the scientific opportunities and new research challenges associated with 5G, ranging from scalability to heterogeneity to cybersecurity. The rapid commercial deployment of 5G opens the opportunity to rethink and reinvent DOE’s scientific infrastructure and experimentation, from intelligent sensor networks at unprecedented scales to a digital continuum of cyberinfrastructure spanning low-power sensors, high-performance computing embedded within and at the edge of the network, and DOE’s large-scale user instrument and computing facilities. New programming paradigms, workflow and data frameworks, and AI-based system design, operation, and autonomous adaptation and optimization will be necessary in order to exploit these new opportunities. Field deployments and centralized scientific instruments can also be revolutionized, moving (without traditional performance penalties) from wired to wireless connectivity for data and control systems, improving flexibility, and opening new sensing modalities, including the use of the 5G electromagnetic spectrum itself as an environmental probe. For DOE science, in contrast to commercial 5G applications and settings, devices will be deployed in extreme environments such as cryogenically cooled instrument control systems and in remote settings with harsh conditions, requiring the design of new materials for RF communication and edge processing to operate in these regimes. Concurrently, 5G infrastructure comprises both hardware and sophisticated software systems - currently closed and proprietary. The cybersecurity challenges to 5G-empowered reinvention mirror the complexity and variety of new 5G features, from virtualization to private network slices to ubiquitous access. Research is also needed in order to accelerate the development of secure and open 5G software infrastructure, reducing reliance on hardware and software produced outside the United States and providing the transparency and rigorous evaluation and testing afforded through open software. Twelve broad research thrusts are laid out in four chapters, with a companion fifth chapter (and three additional research thrusts) underscoring the needs and opportunities for an aggressive testbed program co-designed by networking experts and scientists involved in the 15 research thrusts. The urgency of undertaking this research is fueled by a global, accelerating deployment of new telecommunications infrastructure that is designed for entertainment and commercial applications - barely scratching the surface of what 5G can do to extend U.S. leadership in scientific discovery.
Collaboration among cancer registries is essential to develop accurate, robust, and generalizable deep learning models for automated information extraction from cancer pathology reports. Sharing data presents a serious privacy issue, especially in biomedical research and healthcare delivery domains. Distributing pretrained deep learning (DL) models has been proposed to avoid critical data sharing. However, there is growing recognition that collaboration among clinical institutes through DL model distribution exposes new security and privacy vulnerabilities. These vulnerabilities increase in natural language processing (NLP) applications, in which the dataset vocabulary with word vector representations needs to be associated with the other model parameters. In this paper, we propose a novel privacy-preserving DL model distribution across cancer registries for information extraction from cancer pathology reports with privacy and confidentiality considerations. The proposed approach exploits the adversarial training framework to distinguish private features from shared features among different datasets. It only shares registry-invariant model parameters, without sharing raw data nor registry-specific model parameters among cancer registries. Thus, it protects both the data and the trained model simultaneously. We compare our proposed approach to single-registry models, and a model trained on centrally hosted data from different cancer registries. The results show that the proposed approach significantly outperforms the single-registry models and achieves statistically indistinguishable micro and macro F1-score as compared to the centralized model.
Objective: In this work we implement 2 different multitask learning (MTL) techniques, hard parameter sharing and cross-stitch, to train a word-level convolutional neural network (CNN) specifically designed for automatic extraction of cancer data from unstructured text in pathology reports. We show the importance of learning related information extraction (IE) tasks leveraging shared representations across the tasks to achieve state-of-the-art performance in classification accuracy and computational efficiency. Materials and Methods: Multitask CNN (MTCNN) attempts to tackle document information extraction by learning to extract multiple key cancer characteristics simultaneously. We trained our MTCNN to perform 5 information extraction tasks: (1) primary cancer site (65 classes), (2) laterality (4 classes), (3) behavior (3 classes), (4) histological type (63 classes), and (5) histological grade (5 classes). We evaluated the performance on a corpus of 95 231 pathology documents (71 223 unique tumors) obtained from the Louisiana Tumor Registry. We compared the performance of the MTCNN models against single-task CNN models and 2 traditional machine learning approaches, namely support vector machine (SVM) and random forest classifier (RFC). Results: MTCNNs offered superior performance across all 5 tasks in terms of classification accuracy as compared with the other machine learning models. Based on retrospective evaluation, the hard parameter sharing and cross-stitch MTCNN models correctly classified 59.04% and 57.93% of the pathology reports respectively across all 5 tasks. The baseline models achieved 53.68% (CNN), 46.37% (RFC), and 36.75% (SVM). Based on prospective evaluation, the percentages of correctly classified cases across the 5 tasks were 60.11% (hard parameter sharing), 58.13% (cross-stitch), 51.30% (single-task CNN), 42.07% (RFC), and 35.16% (SVM). Moreover, hard parameter sharing MTCNNs outperformed the other models in computational efficiency by using about the same number of trainable parameters as a single-task CNN. Conclusions: The hard parameter sharing MTCNN offers superior classification accuracy for automated coding support of pathology documents across a wide range of cancers and multiple information extraction tasks while maintaining similar training and inference time as those of a single task–specific model.
We introduce a deep learning architecture, hierarchical self-attention networks (HiSANs), designed for classifying pathology reports and show how its unique architecture leads to a new state-of-the-art in accuracy, faster training, and clear interpretability. We evaluate performance on a corpus of 374,899 pathology reports obtained from the National Cancer Institute's (NCI) Surveillance, Epidemiology, and End Results (SEER) program. Each pathology report is associated with five clinical classification tasks – site, laterality, behavior, histology, and grade. We compare the performance of the HiSAN against other machine learning and deep learning approaches commonly used on medical text data – Naive Bayes, logistic regression, convolutional neural networks, and hierarchical attention networks (the previous state-of-the-art). We show that HiSANs are superior to other machine learning and deep learning text classifiers in both accuracy and macro F-score across all five classification tasks. Compared to the previous state-of-the-art, hierarchical attention networks, HiSANs not only are an order of magnitude faster to train, but also achieve about 1% better relative accuracy and 5% better relative macro F-score.
Deep learning (DL) has been used for many natural language processing (NLP) tasks due to its superior performance as compared to traditional machine learning approaches. In DL models for NLP, words are represented using word embeddings, which capture both semantic and syntactic information in text. However, 90-95% of the DL trainable parameters are associated with the word embeddings, resulting in a large storage or memory footprint. Therefore, reducing the number of word embedding parameters is critical, especially with the increase of vocabulary size. In this work, we propose a novel approximate word embeddings approach for convolutional neural networks (CNNs) used for text classification tasks. The proposed approach significantly reduces the number of model trainable parameters without noticeably sacrificing in computing performance accuracy. Compared to other techniques, our proposed word embeddings technique does not require modifications to the DL model architecture. We evaluate the performance of the the proposed word embeddings on three classification tasks using two datasets, composed of Yelp and Amazon reviews. The results show that the proposed method can reduce the number of word embeddings parameters by 98% and 99% for the Yelp and Amazon datasets respectively, with no drop in computing accuracy.
Pathology reports are a main source of data for cancer surveillance programs. Manual coding of pathology reports is labor-intensive but necessary for obtaining labeled data to train automated information extraction systems. In this study, we investigated semi-supervised deep learning, improving the performance of a multitask information extraction system for automated annotation of pathology reports. We used a set of over 374,000 pathology reports from the Louisiana Tumor Registry and a novel convolutional attention-based auto-encoder. We performed a set of experiments comparing supervised training augmented with unlabeled data at 1%, 5%, 10%, and 50% of the original data size. We also compared the impact of extending text processing to include unlabeled tokens. We find that semi-supervised training consistently improved individual performance with increased micro-averaged F-scores between 0.012 and 0.064 and increased macro-averaged F-scores of up to 0.158. This demonstrates that semantic information learned via unsupervised learning can be used to improve supervised clinical task performance.
Automated text information extraction from cancer pathology reports is an active area of research to support national cancer surveillance. A well-known challenge is how to develop information extraction tools with robust performance across cancer registries. In this study we investigated whether transfer learning (TL) with a convolutional neural network (CNN) can facilitate cross-registry knowledge sharing. Specifically, we performed a series of experiments to determine whether a CNN trained with single-registry data is capable of transferring knowledge to another registry or whether developing a cross-registry knowledge database produces a more effective and generalizable model. Using data from two cancer registries and primary tumor site and topography as the information extraction task of interest, our study showed that TL results in 6.90% and 17.22% improvement of classification m acro F -score o ver the baseline single-registry models. Detailed analysis illustrated that the observed improvement is evident in the low prevalence classes.
Finding optimal hyperparameters is necessary to identify the best performing deep learning models but the process is costly. In this paper, we applied model-based optimization, also known as Bayesian optimization, using the CANDLE framework implemented on a High-Performance Computing environment. As a use case we selected information extraction from cancer pathology reports using a multi-task convolutional neural network, and hierarchical convolutional attention network to be optimized. We utilized a synthesized text corpus of 8,000 training cases and 2,000 validation cases with four types of clinical task labels including primary cancer site, laterality, behavior, and histological grade. We demonstrated that hyperparameter optimization using the CANDLE framework is a feasible approach with respect to both scalability and clinical task performance.
Here, FPGA-based heterogeneous computing platform, due to its extreme logic reconfigurability, emerges to be a strong contender as computing fabric in modern AI. As a result, various FPGA-based accelerators for deep CNN—the key driver of modern AI—have been proposed due to their advantages of high performance, reconfigurability, and fast development round, etc. In general, the consensus among researchers is that, although FPGA-based accelerator can achieve much higher energy efficiency, its raw computing performance lags behind when compared with GPUs with similar logic density. In this paper, we develop an alternative methodology to efficiently implement CNNs with FPGAs that outperform GPUs in terms of both power consumption and performance. Our key idea is to design a scalable hardware architecture and circuit design for large-scale CNNs that leverages a stochastic-based computing principle. Specifically, there are three major performance advantages. First, all key components of our deep learning CNN are designed and implemented to compute stochastically, thus achieving excellent computing performance and energy efficiency. Second, because our proposed CNN architecture enables a stream-mode computing, all of its stages can process even the partial results from preceding stages, therefore not incurring unnecessary latency due to data dependency. Finally, our FPGA-based deep CNN also provides a superior hardware scalability when compared with conventional FPGA implementations by reducing the bandwidth requirement between layers. The results show that our proposed CNN architecture significantly outperforms all previous FPGA-based deep CNN implementation approaches. It achieves 1.58x more GOPS, 6.42x more GOPS/Slice, and 10.92x more GOPS/W when compared with state-of-the-art CNN architecture. The top-5 accuracy of stochastic VGG-16 CNN is 86.77 percent with 18.91 fps frame rate.
Deep learning has surged in popularity and proven to be effective for various artificial intelligence applications including information extraction from cancer pathology reports. Since word representation is a core unit that enables deep learning algorithms to understand words and be able to perform NLP, this representation must include as much information as possible to help these algorithms achieve high classification performance. Therefore, in this work in addition to the distributional information of words in large sized corpora, we use UMLS vocabulary resources to enrich the vector space representation of words with the semantic relations between words. These resources provide many terminologies pertaining to cancer. The refined word embeddings are used with a convolutional neural (CNN) model to extract four data elements from cancer pathology reports; ICD-O-3 tumor topography codes, tumor laterality, behavior, and histological grade. We observed that using UMLS vocabulary resources to enrich word embeddings of CNN models consistently outperformed CNN models without pre-training word embeddings and even with pre-trained word embeddings on a domain specific corpus across all four tasks. The results show marginal improvement on the laterality task, but a significant improvement on the other tasks, especially for the macro-f score. Specifically, the improvements are 3%, 13%, and 15% for tumor site, histological grade, and behavior tasks, respectively. This approach is encouraging to enrich word embeddings with more clinical data resources to be used for information abstraction tasks from clinical pathology reports.
Information extraction and coding of free-text pathology reports is an important activity for cancer registries to support national cancer surveillance. Cancer registrars must process high volumes of pathology reports on an annual basis. In this study, we investigated an automated approach using a coarse-to-fine training of convolutional neural networks (CNNs) for extracting the primary site, histological grade and laterality from unstructured cancer pathology text reports. Our proposed training scheme consists of two stages. In the first stage, the multi-task learning (MTL) with hard parameter sharing approach is used to train a multi-task MT-CNN model for all the tasks. Then, the TM-CNN parameters are used to initialize a CNN model for each task to be fine trained individually using its corresponding dataset. The performance of our proposed model was compared against a state-of-the-art CNN and the commonly used SVM classifier. We observed that the proposed model consistently outperformed the base line models, especially for the less prevalent classes. Specifically, the proposed training approach achieved a micro-F score of 0.7749 over 12 ICD- O-3 topography codes which is a significant improvement as compared with state-of-the-art CNN (0.7101) and the SVM (0.6019) classifiers. Also, the results demonstrate the potential of the proposed method for handling class imbalance within each task. It significantly improves macro-F score by 24% and 12% of the primary site and histology grade tasks, respectively.
With large deep neural networks (DNNs) neces- sary to solve complex and data-intensive problems, energy efficiency is a key bottleneck for effectively deploying DL in the real world. Deep spiking NNs have gained much research at- tention recently due to the interest in building biological neural networks and the availability of neuromorphic platforms, which can be orders of magnitude more energy efficient compared to CPUs and GPUs. Although spiking NNs have proven to be an efficient technique for solving many machine learning and computer vision problems, to the best of our knowledge, this is the first attempt to adapt spiking NNs to sparse datasets. In this paper, we study the behaviour of spiking NNs in handling NLP datasets and the sparsity in their data representation. Then, we propose a novel framework for spiking NN using the concept of stochastic computing. Specifically, instead of generating spike trains with firing rates proportional to the intensity of each value in the feature set separately, the whole feature set is treated as a distribution function and a stochastic spiking train that follow this distribution is generated. This framework reduces the connectivity between NN layers from O(N) to O(logN). Also, it encodes input data differently and make suitable to handle sparse datasets. Finally, the framework achieves high energy efficiency since it uses Integrate and Fire neurons same as conventional spiking NNs. The results show that our proposed stochastic-based SNN achieves nearly the same accuracy as the original DNN on MNIST dataset, and it has better performance than state- of-the-art SNN. Besides that stochastic-based SNN is energy efficient, where the fully connected DNN, the conventional SNN, and the data normalized SNN consume 38.24, 1.83, and 1.85 -times more energy than the stochastic-based SNN, respectively. For sparse datasets, including IMDb and In-House clinical datasets, stochastic-based SNN achieves performance comparable to that of the conventional DNN. However, the conventional spiking NN has a significant decline in classification accuracy.