Chief complaint classification with recurrent neural networks

Integrating domain knowledge for biomedical text analysis into deep learning: A survey

The past decade has witnessed an explosion of textual information in the biomedical field. Biomedical texts provide a basis for healthcare delivery, knowledge discovery, and decision-making. Over the same period, deep learning has achieved remarkable performance in biomedical natural language processing, however, its development has been limited by well-annotated datasets and interpretability. To solve this, researchers have considered combining domain knowledge (such as biomedical knowledge graph) with biomedical data, which has become a promising means of introducing more information into biomedical datasets and following evidence-based medicine. This paper comprehensively reviews more than 150 recent literature studies on incorporating domain knowledge into deep learning models to facilitate typical biomedical text analysis tasks, including information extraction, text classification, and text generation. We eventually discuss various challenges and future directions.

Performance of emergency triage prediction of an open access natural language processing based chatbot application (ChatGPT): A preliminary, scenario-based cross-sectional study

OBJECTIVES

Artificial intelligence companies have been increasing their initiatives recently to improve the results of chatbots, which are software programs that can converse with a human in natural language. The role of chatbots in health care is deemed worthy of research. OpenAI's ChatGPT is a supervised and empowered machine learning-based chatbot. The aim of this study was to determine the performance of ChatGPT in emergency medicine (EM) triage prediction.

METHODS

This was a preliminary, cross-sectional study conducted with case scenarios generated by the researchers based on the emergency severity index (ESI) handbook v4 cases. Two independent EM specialists who were experts in the ESI triage scale determined the triage categories for each case. A third independent EM specialist was consulted as arbiter, if necessary. Consensus results for each case scenario were assumed as the reference triage category. Subsequently, each case scenario was queried with ChatGPT and the answer was recorded as the index triage category. Inconsistent classifications between the ChatGPT and reference category were defined as over-triage (false positive) or under-triage (false negative).

RESULTS

Fifty case scenarios were assessed in the study. Reliability analysis showed a fair agreement between EM specialists and ChatGPT (Cohen's Kappa: 0.341). Eleven cases (22%) were over triaged and 9 (18%) cases were under triaged by ChatGPT. In 9 cases (18%), ChatGPT reported two consecutive triage categories, one of which matched the expert consensus. It had an overall sensitivity of 57.1% (95% confidence interval [CI]: 34-78.2), specificity of 34.5% (95% CI: 17.9-54.3), positive predictive value (PPV) of 38.7% (95% CI: 21.8-57.8), negative predictive value (NPV) of 52.6 (95% CI: 28.9-75.6), and an F1 score of 0.461. In high acuity cases (ESI-1 and ESI-2), ChatGPT showed a sensitivity of 76.2% (95% CI: 52.8-91.8), specificity of 93.1% (95% CI: 77.2-99.2), PPV of 88.9% (95% CI: 65.3-98.6), NPV of 84.4 (95% CI: 67.2-94.7), and an F1 score of 0.821. The receiver operating characteristic curve showed an area under the curve of 0.846 (95% CI: 0.724-0.969, P < 0.001) for high acuity cases.

CONCLUSION

The performance of ChatGPT was best when predicting high acuity cases (ESI-1 and ESI-2). It may be useful when determining the cases requiring critical care. When trained with more medical knowledge, ChatGPT may be more accurate for other triage category predictions.

The Use and Structure of Emergency Nurses' Triage Narrative Data: Scoping Review

BACKGROUND

Emergency departments use triage to ensure that patients with the highest level of acuity receive care quickly and safely. Triage is typically a nursing process that is documented as structured and unstructured (free text) data. Free-text triage narratives have been studied for specific conditions but never reviewed in a comprehensive manner.

OBJECTIVE

The objective of this paper was to identify and map the academic literature that examines triage narratives. The paper described the types of research conducted, identified gaps in the research, and determined where additional review may be warranted.

METHODS

We conducted a scoping review of unstructured triage narratives. We mapped the literature, described the use of triage narrative data, examined the information available on the form and structure of narratives, highlighted similarities among publications, and identified opportunities for future research.

RESULTS

We screened 18,074 studies published between 1990 and 2022 in CINAHL, MEDLINE, Embase, Cochrane, and ProQuest Central. We identified 0.53% (96/18,074) of studies that directly examined the use of triage nurses' narratives. More than 12 million visits were made to 2438 emergency departments included in the review. In total, 82% (79/96) of these studies were conducted in the United States (43/96, 45%), Australia (31/96, 32%), or Canada (5/96, 5%). Triage narratives were used for research and case identification, as input variables for predictive modeling, and for quality improvement. Overall, 31% (30/96) of the studies offered a description of the triage narrative, including a list of the keywords used (27/96, 28%) or more fulsome descriptions (such as word counts, character counts, abbreviation, etc; 7/96, 7%). We found limited use of reporting guidelines (8/96, 8%).

CONCLUSIONS

The breadth of the identified studies suggests that there is widespread routine collection and research use of triage narrative data. Despite the use of triage narratives as a source of data in studies, the narratives and nurses who generate them are poorly described in the literature, and data reporting is inconsistent. Additional research is needed to describe the structure of triage narratives, determine the best use of triage narratives, and improve the consistent use of triage-specific data reporting guidelines.

INTERNATIONAL REGISTERED REPORT IDENTIFIER (IRRID)

RR2-10.1136/bmjopen-2021-055132.

Diagnostic trajectories in primary care at 12 months: an observational cohort study

BACKGROUND

Little is known about the epidemiology of diagnosis in primary care.

METHODS

A prospective observational cohort study was conducted of adults presenting between August and December 2018 to primary care clinics across two health systems with an undiagnosed medical problem. Primary outcomes were (1) likelihood of a definitive diagnosis by 12 months and (2) time to diagnosis. Multivariate logistic regression was used to assess for factors associated with the likelihood of reaching a diagnosis, and multivariable Cox regression was used to assess for factors associated with time to diagnosis. Bivariate models were used to explore unadjusted relationships between the cases' organ systems and likelihood of and time to diagnosis.

RESULTS

Among 410 cases in a diverse patient population, 206 (50.2%) reached a final diagnosis within 12 months, with a median time to diagnosis of 5 days (interquartile range = 0-46). Among these cases, 32.4% reached a diagnosis within the first month. A majority of cases not diagnosed within a month of the first presentation remained undiagnosed at 12 months. The likelihood of diagnosis and time to diagnosis did not differ by clinician or patient characteristics, clinicians' level of diagnostic uncertainty, chronicity of the medical issue, or visit type. There were no significant associations between organ system and likelihood of time to diagnosis.

CONCLUSION

Patients presenting with new or unresolved problems in ambulatory primary care often remain undiagnosed after a year. There were no provider or patient-level variables associated with such lack of diagnosis. The causes, contributors, and consequences of lack of timely diagnosis and potential solutions require further research.

Influenza, dengue and common cold detection using LSTM with fully connected neural network and keywords selection

Symptom-based machine learning models for disease detection are a way to reduce the workload of doctors when they have too many patients. Currently, there are many research studies on machine learning or deep learning for disease detection or clinical departments classification, using text of patient’s symptoms and vital signs. In this study, we used the Long Short-term Memory (LSTM) with a fully connected neural network model for classification, where the LSTM model was used to receive the patient’s symptoms text as input data. The fully connected neural network was used to receive other input data from the patients, including body temperature, age, gender, and the month the patients received care in. In this research, a data preprocessing algorithm was improved by using keyword selection to reduce the complexity of input data for overfitting problem prevention. The results showed that the LSTM with fully connected neural network model performed better than the LSTM model. The keyword selection method also increases model performance.

Discovering the Arrow of Time in Machine Learning

Machine learning (ML) is increasingly useful as data grow in volume and accessibility. ML can perform tasks (e.g., categorisation, decision making, anomaly detection, etc.) through experience and without explicit instruction, even when the data are too vast, complex, highly variable, full of errors to be analysed in other ways. Thus, ML is great for natural language, images, or other complex and messy data available in large and growing volumes. Selecting ML models for tasks depends on many factors as they vary in supervision needed, tolerable error levels, and ability to account for order or temporal context, among many other things. Importantly, ML methods for tasks that use explicitly ordered or time-dependent data struggle with errors or data asymmetry. Most data are (implicitly) ordered or time-dependent, potentially allowing a hidden ‘arrow of time’ to affect ML performance on non-temporal tasks. This research explores the interaction of ML and implicit order using two ML models to automatically classify (a non-temporal task) tweets (temporal data) under conditions that balance volume and complexity of data. Results show that performance was affected, suggesting that researchers should carefully consider time when matching appropriate ML models to tasks, even when time is only implicitly included.

Natural Language Processing to Extract Meaningful Information from Patient Experience Feedback

BACKGROUND

Due to reimbursement tied in part to patients' perception of their care, hospitals continue to stress obtaining patient feedback and understanding it to plan interventions to improve patients' experience. We demonstrate the use of natural language processing (NLP) to extract meaningful information from patient feedback obtained through Press Ganey surveys.

METHODS

The first step was to standardize textual data programmatically using NLP libraries. This included correcting spelling mistakes, converting text to lowercase, and removing words that most likely did not carry useful information. Next, we converted numeric data pertaining to each category based on sentiment and care aspect into charts. We selected care aspect categories where there were more negative comments for more in-depth study. Using NLP, we made tables of most frequently appearing words, adjectives, and bigrams. Comments with frequent words/combinations underwent further study manually to understand factors contributing to negative patient feedback. We then used the positive and negative comments as the training dataset for a neural network to perform sentiment analysis on sentences obtained by splitting mixed reviews.

RESULTS

We found that most of the comments were about doctors and nurses, confirming the important role patients ascribed to these two in patient care. "Room," "discharge" and "tests and treatments" were the three categories that had more negative than positive comments. We then tabulated commonly appearing words, adjectives, and two-word combinations. We found that climate control, housekeeping and noise levels in the room, time delays in discharge paperwork, conflicting information about discharge plan, frequent blood draws, and needle sticks were major contributors to negative patient feedback. None of this information was available from numeric data alone.

CONCLUSION

NLP is an effective tool to gain insight from raw textual patient feedback to extract meaningful information, making it a powerful tool in processing large amounts of patient feedback efficiently.

Machine learning for syndromic surveillance using veterinary necropsy reports

The use of natural language data for animal population surveillance represents a valuable opportunity to gather information about potential disease outbreaks, emerging zoonotic diseases, or bioterrorism threats. In this study, we evaluate machine learning methods for conducting syndromic surveillance using free-text veterinary necropsy reports. We train a system to detect if a necropsy report from the Wisconsin Veterinary Diagnostic Laboratory contains evidence of gastrointestinal, respiratory, or urinary pathology. We evaluate the performance of several machine learning algorithms including deep learning with a long short-term memory network. Although no single algorithm was superior, random forest using feature vectors of TF-IDF statistics ranked among the top-performing models with F1 scores of 0.923 (gastrointestinal), 0.960 (respiratory), and 0.888 (urinary). This model was applied to over 33,000 necropsy reports and was used to describe temporal and spatial features of diseases within a 14-year period, exposing epidemiological trends and detecting a potential focus of gastrointestinal disease from a single submitting producer in the fall of 2016.

Twitter mining using semi-supervised classification for relevance filtering in syndromic surveillance

We investigate the use of Twitter data to deliver signals for syndromic surveillance in order to assess its ability to augment existing syndromic surveillance efforts and give a better understanding of symptomatic people who do not seek healthcare advice directly. We focus on a specific syndrome-asthma/difficulty breathing. We outline data collection using the Twitter streaming API as well as analysis and pre-processing of the collected data. Even with keyword-based data collection, many of the tweets collected are not be relevant because they represent chatter, or talk of awareness instead of an individual suffering a particular condition. In light of this, we set out to identify relevant tweets to collect a strong and reliable signal. For this, we investigate text classification techniques, and in particular we focus on semi-supervised classification techniques since they enable us to use more of the Twitter data collected while only doing very minimal labelling. In this paper, we propose a semi-supervised approach to symptomatic tweet classification and relevance filtering. We also propose alternative techniques to popular deep learning approaches. Additionally, we highlight the use of emojis and other special features capturing the tweet's tone to improve the classification performance. Our results show that negative emojis and those that denote laughter provide the best classification performance in conjunction with a simple word-level n-gram approach. We obtain good performance in classifying symptomatic tweets with both supervised and semi-supervised algorithms and found that the proposed semi-supervised algorithms preserve more of the relevant tweets and may be advantageous in the context of a weak signal. Finally, we found some correlation (r = 0.414, p = 0.0004) between the Twitter signal generated with the semi-supervised system and data from consultations for related health conditions.

Natural language generation for electronic health records

One broad goal of biomedical informatics is to generate fully-synthetic, faithfully representative electronic health records (EHRs) to facilitate data sharing between healthcare providers and researchers and promote methodological research. A variety of methods existing for generating synthetic EHRs, but they are not capable of generating unstructured text, like emergency department (ED) chief complaints, history of present illness, or progress notes. Here, we use the encoder–decoder model, a deep learning algorithm that features in many contemporary machine translation systems, to generate synthetic chief complaints from discrete variables in EHRs, like age group, gender, and discharge diagnosis. After being trained end-to-end on authentic records, the model can generate realistic chief complaint text that appears to preserve the epidemiological information encoded in the original record-sentence pairs. As a side effect of the model’s optimization goal, these synthetic chief complaints are also free of relatively uncommon abbreviation and misspellings, and they include none of the personally identifiable information (PII) that was in the training data, suggesting that this model may be used to support the de-identification of text in EHRs. When combined with algorithms like generative adversarial networks (GANs), our model could be used to generate fully-synthetic EHRs, allowing healthcare providers to share faithful representations of multimodal medical data without compromising patient privacy. This is an important advance that we hope will facilitate the development of machine-learning methods for clinical decision support, disease surveillance, and other data-hungry applications in biomedical informatics.

Natural language generation for electronic health records

One broad goal of biomedical informatics is to generate fully-synthetic, faithfully representative electronic health records (EHRs) to facilitate data sharing between healthcare providers and researchers and promote methodological research. A variety of methods existing for generating synthetic EHRs, but they are not capable of generating unstructured text, like emergency department (ED) chief complaints, history of present illness, or progress notes. Here, we use the encoder-decoder model, a deep learning algorithm that features in many contemporary machine translation systems, to generate synthetic chief complaints from discrete variables in EHRs, like age group, gender, and discharge diagnosis. After being trained end-to-end on authentic records, the model can generate realistic chief complaint text that appears to preserve the epidemiological information encoded in the original record-sentence pairs. As a side effect of the model's optimization goal, these synthetic chief complaints are also free of relatively uncommon abbreviation and misspellings, and they include none of the personally identifiable information (PII) that was in the training data, suggesting that this model may be used to support the de-identification of text in EHRs. When combined with algorithms like generative adversarial networks (GANs), our model could be used to generate fully-synthetic EHRs, allowing healthcare providers to share faithful representations of multimodal medical data without compromising patient privacy. This is an important advance that we hope will facilitate the development of machine-learning methods for clinical decision support, disease surveillance, and other data-hungry applications in biomedical informatics.