Evaluating the ChatGPT family of models for biomedical reasoning and classification.

OBJECTIVE: Large language models (LLMs) have shown impressive ability in biomedical question-answering, but have not been adequately investigated for more specific biomedical applications. This study investigates ChatGPT family of models (GPT-3.5, GPT-4) in biomedical tasks beyond question-answering. MATERIALS AND METHODS: We evaluated model performance with 11 122 samples for two fundamental tasks in the biomedical domain-classification (n = 8676) and reasoning (n = 2446). The first task involves classifying health advice in scientific literature, while the second task is detecting causal relations in biomedical literature. We used 20% of the dataset for prompt development, including zero- and few-shot settings with and without chain-of-thought (CoT). We then evaluated the best prompts from each setting on the remaining dataset, comparing them to models using simple features (BoW with logistic regression) and fine-tuned BioBERT models. RESULTS: Fine-tuning BioBERT produced the best classification (F1: 0.800-0.902) and reasoning (F1: 0.851) results. Among LLM approaches, few-shot CoT achieved the best classification (F1: 0.671-0.770) and reasoning (F1: 0.682) results, comparable to the BoW model (F1: 0.602-0.753 and 0.675 for classification and reasoning, respectively). It took 78 h to obtain the best LLM results, compared to 0.078 and 0.008 h for the top-performing BioBERT and BoW models, respectively. DISCUSSION: The simple BoW model performed similarly to the most complex LLM prompting. Prompt engineering required significant investment. CONCLUSION: Despite the excitement around viral ChatGPT, fine-tuning for two fundamental biomedical natural language processing tasks remained the best strategy.

Large language models to identify social determinants of health in electronic health records.

Social determinants of health (SDoH) play a critical role in patient outcomes, yet their documentation is often missing or incomplete in the structured data of electronic health records (EHRs). Large language models (LLMs) could enable high-throughput extraction of SDoH from the EHR to support research and clinical care. However, class imbalance and data limitations present challenges for this sparsely documented yet critical information. Here, we investigated the optimal methods for using LLMs to extract six SDoH categories from narrative text in the EHR: employment, housing, transportation, parental status, relationship, and social support. The best-performing models were fine-tuned Flan-T5 XL for any SDoH mentions (macro-F1 0.71), and Flan-T5 XXL for adverse SDoH mentions (macro-F1 0.70). Adding LLM-generated synthetic data to training varied across models and architecture, but improved the performance of smaller Flan-T5 models (delta F1 + 0.12 to +0.23). Our best-fine-tuned models outperformed zero- and few-shot performance of ChatGPT-family models in the zero- and few-shot setting, except GPT4 with 10-shot prompting for adverse SDoH. Fine-tuned models were less likely than ChatGPT to change their prediction when race/ethnicity and gender descriptors were added to the text, suggesting less algorithmic bias (p < 0.05). Our models identified 93.8% of patients with adverse SDoH, while ICD-10 codes captured 2.0%. These results demonstrate the potential of LLMs in improving real-world evidence on SDoH and assisting in identifying patients who could benefit from resource support.

Use of Artificial Intelligence Chatbots for Cancer Treatment Information.

This survey study examines the performance of a large language model chatbot in providing cancer treatment recommendations that are concordant with National Comprehensive Cancer Network guidelines.

Natural Language Processing to Automatically Extract the Presence and Severity of Esophagitis in Notes of Patients Undergoing Radiotherapy.

PURPOSE: Radiotherapy (RT) toxicities can impair survival and quality of life, yet remain understudied. Real-world evidence holds potential to improve our understanding of toxicities, but toxicity information is often only in clinical notes. We developed natural language processing (NLP) models to identify the presence and severity of esophagitis from notes of patients treated with thoracic RT. METHODS: Our corpus consisted of a gold-labeled data set of 1,524 clinical notes from 124 patients with lung cancer treated with RT, manually annotated for Common Terminology Criteria for Adverse Events (CTCAE) v5.0 esophagitis grade, and a silver-labeled data set of 2,420 notes from 1,832 patients from whom toxicity grades had been collected as structured data during clinical care. We fine-tuned statistical and pretrained Bidirectional Encoder Representations from Transformers-based models for three esophagitis classification tasks: task 1, no esophagitis versus grade 1-3; task 2, grade ≤1 versus >1; and task 3, no esophagitis versus grade 1 versus grade 2-3. Transferability was tested on 345 notes from patients with esophageal cancer undergoing RT. RESULTS: Fine-tuning of PubMedBERT yielded the best performance. The best macro-F1 was 0.92, 0.82, and 0.74 for tasks 1, 2, and 3, respectively. Selecting the most informative note sections during fine-tuning improved macro-F1 by ≥2% for all tasks. Silver-labeled data improved the macro-F1 by ≥3% across all tasks. For the esophageal cancer notes, the best macro-F1 was 0.73, 0.74, and 0.65 for tasks 1, 2, and 3, respectively, without additional fine-tuning. CONCLUSION: To our knowledge, this is the first effort to automatically extract esophagitis toxicity severity according to CTCAE guidelines from clinical notes. This provides proof of concept for NLP-based automated detailed toxicity monitoring in expanded domains.

Natural Language Processing Methods to Empirically Explore Social Contexts and Needs in Cancer Patient Notes.

PURPOSE: There is an unmet need to empirically explore and understand drivers of cancer disparities, particularly social determinants of health. We explored natural language processing methods to automatically and empirically extract clinical documentation of social contexts and needs that may underlie disparities. METHODS: This was a retrospective analysis of 230,325 clinical notes from 5,285 patients treated with radiotherapy from 2007 to 2019. We compared linguistic features among White versus non-White, low-income insurance versus other insurance, and male versus female patients' notes. Log odds ratios with an informative Dirichlet prior were calculated to compare words over-represented in each group. A variational autoencoder topic model was applied, and topic probability was compared between groups. The presence of machine-learnable bias was explored by developing statistical and neural demographic group classifiers. RESULTS: Terms associated with varied social contexts and needs were identified for all demographic group comparisons. For example, notes of non-White and low-income insurance patients were over-represented with terms associated with housing and transportation, whereas notes of White and other insurance patients were over-represented with terms related to physical activity. Topic models identified a social history topic, and topic probability varied significantly between the demographic group comparisons. Classification models performed poorly at classifying notes of non-White and low-income insurance patients (F1 of 0.30 and 0.23, respectively). CONCLUSION: Exploration of linguistic differences in clinical notes between patients of different race/ethnicity, insurance status, and sex identified social contexts and needs in patients with cancer and revealed high-level differences in notes. Future work is needed to validate whether these findings may play a role in cancer disparities.

An End-to-End Natural Language Processing System for Automatically Extracting Radiation Therapy Events From Clinical Texts.

PURPOSE: Real-world evidence for radiation therapy (RT) is limited because it is often documented only in the clinical narrative. We developed a natural language processing system for automated extraction of detailed RT events from text to support clinical phenotyping. METHODS AND MATERIALS: A multi-institutional data set of 96 clinician notes, 129 North American Association of Central Cancer Registries cancer abstracts, and 270 RT prescriptions from HemOnc.org was used and divided into train, development, and test sets. Documents were annotated for RT events and associated properties: dose, fraction frequency, fraction number, date, treatment site, and boost. Named entity recognition models for properties were developed by fine-tuning BioClinicalBERT and RoBERTa transformer models. A multiclass RoBERTa-based relation extraction model was developed to link each dose mention with each property in the same event. Models were combined with symbolic rules to create a hybrid end-to-end pipeline for comprehensive RT event extraction. RESULTS: Named entity recognition models were evaluated on the held-out test set with F1 results of 0.96, 0.88, 0.94, 0.88, 0.67, and 0.94 for dose, fraction frequency, fraction number, date, treatment site, and boost, respectively. The relation model achieved an average F1 of 0.86 when the input was gold-labeled entities. The end-to-end system F1 result was 0.81. The end-to-end system performed best on North American Association of Central Cancer Registries abstracts (average F1 0.90), which are mostly copy-paste content from clinician notes. CONCLUSIONS: We developed methods and a hybrid end-to-end system for RT event extraction, which is the first natural language processing system for this task. This system provides proof-of-concept for real-world RT data collection for research and is promising for the potential of natural language processing methods to support clinical care.

Clinical Natural Language Processing for Radiation Oncology: A Review and Practical Primer.

Natural language processing (NLP), which aims to convert human language into expressions that can be analyzed by computers, is one of the most rapidly developing and widely used technologies in the field of artificial intelligence. Natural language processing algorithms convert unstructured free text data into structured data that can be extracted and analyzed at scale. In medicine, this unlocking of the rich, expressive data within clinical free text in electronic medical records will help untap the full potential of big data for research and clinical purposes. Recent major NLP algorithmic advances have significantly improved the performance of these algorithms, leading to a surge in academic and industry interest in developing tools to automate information extraction and phenotyping from clinical texts. Thus, these technologies are poised to transform medical research and alter clinical practices in the future. Radiation oncology stands to benefit from NLP algorithms if they are appropriately developed and deployed, as they may enable advances such as automated inclusion of radiation therapy details into cancer registries, discovery of novel insights about cancer care, and improved patient data curation and presentation at the point of care. However, challenges remain before the full value of NLP is realized, such as the plethora of jargon specific to radiation oncology, nonstandard nomenclature, a lack of publicly available labeled data for model development, and interoperability limitations between radiation oncology data silos. Successful development and implementation of high quality and high value NLP models for radiation oncology will require close collaboration between computer scientists and the radiation oncology community. Here, we present a primer on artificial intelligence algorithms in general and NLP algorithms in particular; provide guidance on how to assess the performance of such algorithms; review prior research on NLP algorithms for oncology; and describe future avenues for NLP in radiation oncology research and clinics.

Adverse drug event presentation and tracking (ADEPT): semiautomated, high throughput pharmacovigilance using real-world data.

OBJECTIVE: To advance use of real-world data (RWD) for pharmacovigilance, we sought to integrate a high-sensitivity natural language processing (NLP) pipeline for detecting potential adverse drug events (ADEs) with easily interpretable output for high-efficiency human review and adjudication of true ADEs. MATERIALS AND METHODS: The adverse drug event presentation and tracking (ADEPT) system employs an open source NLP pipeline to identify in clinical notes mentions of medications and signs and symptoms potentially indicative of ADEs. ADEPT presents the output to human reviewers by highlighting these drug-event pairs within the context of the clinical note. To measure incidence of seizures associated with sildenafil, we applied ADEPT to 149 029 notes for 982 patients with pediatric pulmonary hypertension. RESULTS: Of 416 patients identified as taking sildenafil, NLP found 72 [17%, 95% confidence interval (CI) 14-21] with seizures as a potential ADE. Upon human review and adjudication, only 4 (0.96%, 95% CI 0.37-2.4) patients with seizures were determined to have true ADEs. Reviewers using ADEPT required a median of 89 s (interquartile range 57-142 s) per patient to review potential ADEs. DISCUSSION: ADEPT combines high throughput NLP to increase sensitivity of ADE detection and human review, to increase specificity by differentiating true ADEs from signs and symptoms related to comorbidities, effects of other medications, or other confounders. CONCLUSION: ADEPT is a promising tool for creating gold standard, patient-level labels for advancing NLP-based pharmacovigilance. ADEPT is a potentially time savings platform for computer-assisted pharmacovigilance based on RWD.

Adverse drug event rates in pediatric pulmonary hypertension: a comparison of real-world data sources.

OBJECTIVE: Real-world data (RWD) are increasingly used for pharmacoepidemiology and regulatory innovation. Our objective was to compare adverse drug event (ADE) rates determined from two RWD sources, electronic health records and administrative claims data, among children treated with drugs for pulmonary hypertension. MATERIALS AND METHODS: Textual mentions of medications and signs/symptoms that may represent ADEs were identified in clinical notes using natural language processing. Diagnostic codes for the same signs/symptoms were identified in our electronic data warehouse for the patients with textual evidence of taking pulmonary hypertension-targeted drugs. We compared rates of ADEs identified in clinical notes to those identified from diagnostic code data. In addition, we compared putative ADE rates from clinical notes to those from a healthcare claims dataset from a large, national insurer. RESULTS: Analysis of clinical notes identified up to 7-fold higher ADE rates than those ascertained from diagnostic codes. However, certain ADEs (eg, hearing loss) were more often identified in diagnostic code data. Similar results were found when ADE rates ascertained from clinical notes and national claims data were compared. DISCUSSION: While administrative claims and clinical notes are both increasingly used for RWD-based pharmacovigilance, ADE rates substantially differ depending on data source. CONCLUSION: Pharmacovigilance based on RWD may lead to discrepant results depending on the data source analyzed. Further work is needed to confirm the validity of identified ADEs, to distinguish them from disease effects, and to understand tradeoffs in sensitivity and specificity between data sources.

Use of Natural Language Processing to Extract Clinical Cancer Phenotypes from Electronic Medical Records.

Current models for correlating electronic medical records with -omics data largely ignore clinical text, which is an important source of phenotype information for patients with cancer. This data convergence has the potential to reveal new insights about cancer initiation, progression, metastasis, and response to treatment. Insights from this real-world data will catalyze clinical care, research, and regulatory activities. Natural language processing (NLP) methods are needed to extract these rich cancer phenotypes from clinical text. Here, we review the advances of NLP and information extraction methods relevant to oncology based on publications from PubMed as well as NLP and machine learning conference proceedings in the last 3 years. Given the interdisciplinary nature of the fields of oncology and information extraction, this analysis serves as a critical trail marker on the path to higher fidelity oncology phenotypes from real-world data.

Supervised methods to extract clinical events from cardiology reports in Italian.

Clinical narratives are a valuable source of information for both patient care and biomedical research. Given the unstructured nature of medical reports, specific automatic techniques are required to extract relevant entities from such texts. In the natural language processing (NLP) community, this task is often addressed by using supervised methods. To develop such methods, both reliably-annotated corpora and elaborately designed features are needed. Despite the recent advances on corpora collection and annotation, research on multiple domains and languages is still limited. In addition, to compute the features required for supervised classification, suitable language- and domain-specific tools are needed. In this work, we propose a novel application of recurrent neural networks (RNNs) for event extraction from medical reports written in Italian. To train and evaluate the proposed approach, we annotated a corpus of 75 cardiology reports for a total of 4365 mentions of relevant events and their attributes (e.g., the polarity). For the annotation task, we developed specific annotation guidelines, which are provided together with this paper. The RNN-based classifier was trained on a training set including 3335 events (60 documents). The resulting model was integrated into an NLP pipeline that uses a dictionary lookup approach to search for relevant concepts inside the text. A test set of 1030 events (15 documents) was used to evaluate and compare different pipeline configurations. As a main result, using the RNN-based classifier instead of the dictionary lookup approach allowed increasing recall from 52.4% to 88.9%, and precision from 81.1% to 88.2%. Further, using the two methods in combination, we obtained final recall, precision, and F1 score of 91.7%, 88.6%, and 90.1%, respectively. These experiments indicate that integrating a well-performing RNN-based classifier with a standard knowledge-based approach can be a good strategy to extract information from clinical text in non-English languages.

Potential Impact of Initial Clinical Data on Adjustment of Pediatric Readmission Rates.

OBJECTIVE: Comparison of readmission rates requires adjustment for case-mix (ie, differences in patient populations), but previously only claims data were available for this purpose. We examined whether incorporation of relatively readily available clinical data improves prediction of pediatric readmissions and thus might enhance case-mix adjustment. METHODS: We examined 30-day readmissions using claims and electronic health record data for patients ≤18 years and 29 days of age who were admitted to 3 children's hospitals from February 2011 to February 2014. Using the Pediatric All-Condition Readmission Measure and starting with a model including age, gender, chronic conditions, and primary diagnosis, we examined whether the addition of initial vital sign and laboratory data improved model performance. We employed machine learning to evaluate the same variables, using the L2-regularized logistic regression with cost-sensitive learning and convolutional neural network. RESULTS: Controlling for the core model variables, low red blood cell count and mean corpuscular hemoglobin concentration and high red cell distribution width were associated with greater readmission risk, as were certain interactions between laboratory and chronic condition variables. However, the C-statistic (0.722 vs 0.713) and McFadden's pseudo R2 (0.085 vs 0.076) for this and the core model were similar, suggesting minimal improvement in performance. In machine learning analyses, the F-measure (harmonic mean of sensitivity and positive predictive value) was similar for the best-performing model (containing all variables) and core model (0.250 vs 0.243). CONCLUSIONS: Readily available clinical variables do not meaningfully improve the prediction of pediatric readmissions and would be unlikely to enhance case-mix adjustment unless their distributions varied widely across hospitals.

DeepPhe: A Natural Language Processing System for Extracting Cancer Phenotypes from Clinical Records.

Precise phenotype information is needed to understand the effects of genetic and epigenetic changes on tumor behavior and responsiveness. Extraction and representation of cancer phenotypes is currently mostly performed manually, making it difficult to correlate phenotypic data to genomic data. In addition, genomic data are being produced at an increasingly faster pace, exacerbating the problem. The DeepPhe software enables automated extraction of detailed phenotype information from electronic medical records of cancer patients. The system implements advanced Natural Language Processing and knowledge engineering methods within a flexible modular architecture, and was evaluated using a manually annotated dataset of the University of Pittsburgh Medical Center breast cancer patients. The resulting platform provides critical and missing computational methods for computational phenotyping. Working in tandem with advanced analysis of high-throughput sequencing, these approaches will further accelerate the transition to precision cancer treatment. Cancer Res; 77(21); e115-8. ©2017 AACR.

Suboptimal Clinical Documentation in Young Children with Severe Obesity at Tertiary Care Centers.

Background and Objectives. The prevalence of severe obesity in children has doubled in the past decade. The objective of this study is to identify the clinical documentation of obesity in young children with a BMI ≥ 99th percentile at two large tertiary care pediatric hospitals. Methods. We used a standardized algorithm utilizing data from electronic health records to identify children with severe early onset obesity (BMI ≥ 99th percentile at age <6 years). We extracted descriptive terms and ICD-9 codes to evaluate documentation of obesity at Boston Children's Hospital and Cincinnati Children's Hospital and Medical Center between 2007 and 2014. Results. A total of 9887 visit records of 2588 children with severe early onset obesity were identified. Based on predefined criteria for documentation of obesity, 21.5% of children (13.5% of visits) had positive documentation, which varied by institution. Documentation in children first seen under 2 years of age was lower than in older children (15% versus 26%). Documentation was significantly higher in girls (29% versus 17%, p < 0.001), African American children (27% versus 19% in whites, p < 0.001), and the obesity focused specialty clinics (70% versus 15% in primary care and 9% in other subspecialty clinics, p < 0.001). Conclusions. There is significant opportunity for improvement in documentation of obesity in young children, even years after the 2007 AAP guidelines for management of obesity.

Multilayered temporal modeling for the clinical domain.

OBJECTIVE: To develop an open-source temporal relation discovery system for the clinical domain. The system is capable of automatically inferring temporal relations between events and time expressions using a multilayered modeling strategy. It can operate at different levels of granularity--from rough temporality expressed as event relations to the document creation time (DCT) to temporal containment to fine-grained classic Allen-style relations. MATERIALS AND METHODS: We evaluated our systems on 2 clinical corpora. One is a subset of the Temporal Histories of Your Medical Events (THYME) corpus, which was used in SemEval 2015 Task 6: Clinical TempEval. The other is the 2012 Informatics for Integrating Biology and the Bedside (i2b2) challenge corpus. We designed multiple supervised machine learning models to compute the DCT relation and within-sentence temporal relations. For the i2b2 data, we also developed models and rule-based methods to recognize cross-sentence temporal relations. We used the official evaluation scripts of both challenges to make our results comparable with results of other participating systems. In addition, we conducted a feature ablation study to find out the contribution of various features to the system's performance. RESULTS: Our system achieved state-of-the-art performance on the Clinical TempEval corpus and was on par with the best systems on the i2b2 2012 corpus. Particularly, on the Clinical TempEval corpus, our system established a new F1 score benchmark, statistically significant as compared to the baseline and the best participating system. CONCLUSION: Presented here is the first open-source clinical temporal relation discovery system. It was built using a multilayered temporal modeling strategy and achieved top performance in 2 major shared tasks.

Methods to Develop an Electronic Medical Record Phenotype Algorithm to Compare the Risk of Coronary Artery Disease across 3 Chronic Disease Cohorts.

BACKGROUND: Typically, algorithms to classify phenotypes using electronic medical record (EMR) data were developed to perform well in a specific patient population. There is increasing interest in analyses which can allow study of a specific outcome across different diseases. Such a study in the EMR would require an algorithm that can be applied across different patient populations. Our objectives were: (1) to develop an algorithm that would enable the study of coronary artery disease (CAD) across diverse patient populations; (2) to study the impact of adding narrative data extracted using natural language processing (NLP) in the algorithm. Additionally, we demonstrate how to implement CAD algorithm to compare risk across 3 chronic diseases in a preliminary study. METHODS AND RESULTS: We studied 3 established EMR based patient cohorts: diabetes mellitus (DM, n = 65,099), inflammatory bowel disease (IBD, n = 10,974), and rheumatoid arthritis (RA, n = 4,453) from two large academic centers. We developed a CAD algorithm using NLP in addition to structured data (e.g. ICD9 codes) in the RA cohort and validated it in the DM and IBD cohorts. The CAD algorithm using NLP in addition to structured data achieved specificity >95% with a positive predictive value (PPV) 90% in the training (RA) and validation sets (IBD and DM). The addition of NLP data improved the sensitivity for all cohorts, classifying an additional 17% of CAD subjects in IBD and 10% in DM while maintaining PPV of 90%. The algorithm classified 16,488 DM (26.1%), 457 IBD (4.2%), and 245 RA (5.0%) with CAD. In a cross-sectional analysis, CAD risk was 63% lower in RA and 68% lower in IBD compared to DM (p<0.0001) after adjusting for traditional cardiovascular risk factors. CONCLUSIONS: We developed and validated a CAD algorithm that performed well across diverse patient populations. The addition of NLP into the CAD algorithm improved the sensitivity of the algorithm, particularly in cohorts where the prevalence of CAD was low. Preliminary data suggest that CAD risk was significantly lower in RA and IBD compared to DM.

Semi-supervised Learning for Phenotyping Tasks.

Supervised learning is the dominant approach to automatic electronic health records-based phenotyping, but it is expensive due to the cost of manual chart review. Semi-supervised learning takes advantage of both scarce labeled and plentiful unlabeled data. In this work, we study a family of semi-supervised learning algorithms based on Expectation Maximization (EM) in the context of several phenotyping tasks. We first experiment with the basic EM algorithm. When the modeling assumptions are violated, basic EM leads to inaccurate parameter estimation. Augmented EM attenuates this shortcoming by introducing a weighting factor that downweights the unlabeled data. Cross-validation does not always lead to the best setting of the weighting factor and other heuristic methods may be preferred. We show that accurate phenotyping models can be trained with only a few hundred labeled (and a large number of unlabeled) examples, potentially providing substantial savings in the amount of the required manual chart review.