Improving Large Language Models for Clinical Named Entity Recognition via Prompt Engineering

This study aims to assess the zero-shot capability of GPT-3.5 based ChatGPT (in the following discussion, we will use ChatGPT to refer to GPT-3.5 based ChatGPT) in the clinical NER task, as defined in the 2010 i2b2 challenge ³⁰. We compared the performance of ChatGPT and GPT-3 in a similar zero-shot setting and included a baseline model, BioClinicalBERT, which was trained on the 2010 i2b2 dataset detailed below. The primary workflow of our investigation is depicted in Figure LABEL:fig:workflow. Two different prompts were crafted to identify three types of clinical entities: Medical Problem, Treatment, and Test from clinical text using both ChatGPT and GPT-3. Additionally, we trained a supervised BioClinicalBERT model using an annotated corpus from the 2010 i2b2 challenge, as a baseline. All three models were then evaluated using an annotated corpus consisting of HPI sections from 100 discharge summaries in the MTSamples collection (see next section).

Two clinical NER datasets were used in our study, including (1) MTSamples, a set of 163 fully synthetic discharge summaries from MTSamples, which was annotated according to the annotation guidelines from the 2010 i2b2 challenge, which aims at extracting Medical Problem, Treatment, and Test ³⁰; (2) the VAERS corpus, a set of 91 publicly available safety reports in VAERS, aiming at extracting nervous system disorder-related events ³¹. The MTSamples dataset is fully synthetic, meaning that it has been artificially generated and contains no real patient information. The VAERS dataset, on the other hand, is derived from publicly available post-market safety reports that are anonymized and do not contain personally identifiable information. So, no sensitive data was sent to OpenAI API, making this study free of privacy concerns. Upon consultation, it was determined that our study did not require an IRB approval. The two datasets were split into training, validation, and test subsets. The training and validation subsets served the purpose of fine-tuning the BioClinicalBERT model. Annotated samples in prompts were randomly sampled from the training sets. The training sets were also used for error analysis to optimize our prompt strategies. The test subsets, however, were reserved exclusively for evaluating the final performance and for comparative analysis. A descriptive statistic of entities in these datasets is presented in Table 1.

We fine-tuned NER models using BioClinicalBERT ³², to serve as baselines of traditional supervised learning approaches. We present results for supervised learning on both the MTSamples test set and the VAERS test set. The model weights were initialized using the transformers package, available at huggingface¹¹1https://huggingface.co/emilyalsentzer/Bio_ClinicalBERT ³³. The hyperparameters employed during model training included a learning rate of 5e-5, a training batch size of 4, 20 epochs, and a weight decay of 0.01 using the AdamW optimizer ³⁴. In addition to fine-tuning NER models using BioClinicalBERT, we also employed a traditional machine learning approach for comparison. We utilized a Conditional Random Field (CRF) model with word features, including Bag-of-word, capitalization of letters in words, and prefixes and suffixes of words ³⁵.
Regarding the GPT models, we used the specific versions GPT-3.5-turbo-0301 and GPT-4-0314 for reproducibility. Temperature in a generative language model refers to a parameter that controls the randomness in the model’s predictions, typically ranging from 0 (completely deterministic) to 1 or higher (increasingly random and diverse outputs). The temperature parameter for GPT models was set to 0 to minimize randomness in response generation. A lower temperature value restricts the model’s tendency to take creative leaps, thereby ensuring more predictable and consistent outputs. This is crucial in clinical NER tasks where accuracy and reliability of information extraction are paramount. In our setup, the GPT models were interacted with in a ’user’ role. This role simulates a real-world user interaction with the model, where the ’user’ inputs prompts and the model generates responses accordingly. This approach reflects a typical use-case scenario for these models in practical applications. All input and output datasets along with prompt variants are included with Jypter notebooks that can interface with the OpenAI API in our GitHub repository. At the time of this study, costs of GPT-3.5 per 1k tokens were approximately $0.03 for input and $0.06 for output. Costs of GPT-4 per 1k tokens were approximately $ 0.001 for input and $ 0.002 for output. Because of privacy issues, notes containing Personal Identifiable Information (PII) could not be used in this experiment and should not be used with the GPT API.

For GPT models, we proposed a task-specific prompt including the following components:

(1) Baseline prompt with task description and format specification: This component provides the LLMs with basic information about the tasks we are instructing them to perform and in what format the LLMs should output results. We instructed the models to highlight the named entities within an HTML file using <span>tags with a class attribute indicating the entity types. This allows the output from GPT models to be easily converted into a traditional Inside-Outside-Beginning (IOB) format, which allows for a direct comparison of NER performance with findings from existing studies.

(2) Annotation guideline-based prompts: This component contains entity definitions and linguistic rules derived from annotation guidelines. Entity definitions offer comprehensive, unambiguous descriptions of an entity within the context of a given task. They play an instrumental role in steering the LLM toward the precise identification of entities within text documents. We noticed that the model’s predictions often differed substantially from the gold standard in terms of grammatical structure. For example, discrepancies may arise concerning what types of phrases to be included (e.g., noun phrases or adjective phrases). To enhance the model’s performance, we referred to and incorporated rules in the annotation guidelines to address these issues.

(3) Error analysis-based instructions: In addition to the original annotation guidelines, we also incorporated additional guidelines following error analysis of GPT outputs using the training data. For example, we noticed that GPT models often tend to annotate consultation procedures as test entities. To prevent this, we incorporated a specific rule stating, ”Consultation procedures should not be annotated as tests.”.

(4) Annotated samples: To further assist the LLMs in understanding the task and generating accurate results, we provided a set of annotated samples to improve its performance in a few-shot learning setting. We randomly selected either 1 or 5 annotated examples (1 or 5-shot learning) from the training set and formatted them according to the task description and entity markup guide. For instance, given a sentence ‘He had been diagnosed with osteoarthritis of the knees and had undergone arthroscopy years prior to admission .’ with ‘osteoarthritis of the knees’ and ‘arthroscopy’ annotated as medical problem and test entities , we incorporated this sentence into the prompt using the following format:
### Examples
Example Input: He had been diagnosed with osteoarthritis of the knees and had undergone arthroscopy years prior to admission .
Example Output: He had been diagnosed with <span class=”problem”>osteoarthritis of the knees</span>and had undergone <span class=”test”>arthroscopy</span>years prior to admission .

We compared the effectiveness of different prompt components by incrementally incorporating annotation guideline-based prompts, error analysis-based instructions and annotated samples as shown in Table 2 (see the complete prompts for two datasets in supplementary materials S1.1).

The performance of the models was evaluated using Precision (P), Recall (R), and F1 scores, following the same evaluation script in the 2010 i2b2 challenge ³⁰. These scores were computed based on both exact-match and relaxed-match criteria. In the context of an exact match, an extracted entity should have identical token boundary and entity type as that in the gold standard. For relaxed-match, an extracted entity that exhibits overlap in text and shares the same entity type with the gold standard is acceptable.

The performance evaluation of GPT-3.5 and GPT-4 in zero-shot settings using different prompts are detailed in Table 3 and Figure 2. Following the integration of annotation guideline-based prompts and error analysis-based instructions, we noticed an improvement in the performance metrics of both GPT models, across each dataset and under each evaluation criteria. Interestingly, we found these two components to have a more pronounced effect on the performance of GPT-3.5 than on GPT-4. More specifically, GPT-3.5 demonstrated an average increase of 0.09 in overall F1 scores, ranging from 0.04 to 0.14. Conversely, GPT-4 displayed a more restrained average improvement of 0.06, with a range of 0.01 to 0.10. Looking at the dataset-specific effects, these two components had a more substantial impact on the VARES dataset compared to the MTSamples dataset. For VARES, we saw an average increase of approximately 0.11, with a range from 0.09 to 0.14. In contrast, for MTSamples, we saw a more modest approximate average increase of 0.04, with the range extending from 0.01 to 0.08.

Table 4 and Figure 3 illustrate the performance comparison among different numbers of N-shot examples with all prompt components included. Generally, the inclusion of more examples leads to better model performance. A combination of 5-shot and all prompts produced the best results by GPT-4, achieving F1 0.593 and 0.861 for MTSamples and 0.542, 0.736 for VAERS under exact- and relaxed-match respectively.

Table 5 displays the performance of BioClinicalBERT, CRF, GPT-3.5, and GPT-4 models for comparison. Among the three models, BioClinicalBERT still demonstrated the highest performance. For MTSamples, it achieved overall F1 scores of 0.785 and 0.901 under exact-match and relaxed-match respectively. Its performance on the VAERS dataset also remained dominant, with overall F1 scores of 0.668 and 0.802 under exact-match and relaxed-match respectively. The CRF model achieved an F1 score of 0.584 and 0.525 in MTSamples and VAERS by exact match and surpassing GPT-3.5. In the relaxed-match criteria, the CRF model performed worse than GPT-4 and GPT-3.5 in the MTSamples and had comparable performance to GPT-3.5 in the VAERS dataset. Comparatively, GPT-3.5 lagged on two datasets with the lowest performance, yet still demonstrated a decent performance with scores of 0.794 and 0.676, as evaluated by relaxed-match criteria on the two datasets respectively. GPT-4 showcased highly competitive performance using the relaxed match criteria, accomplishing F1 scores of 0.861 and 0.736 on the MTSamples and VAERS datasets respectively. It is notable, however, that the performances of GPT-3.5 and GPT-4 as evaluated by the exact-match method were not as impressive as those by the relaxed-match. In addition to the test set results, we have provided the model’s performance on the validation sets in supplementary materials S1.2 to ensure the BioClinicalBERT is not overfitting.

A random sample of 20 sentences was selected from the outputs generated by each GPT model across the two datasets, post-processing. This selection included sentences with both false positives and false negatives. The error analysis was conducted based on exact match. The error statistics derived from this analysis are presented in Figure 5. When assessed on a dataset basis, GPT-3.5 and GPT-4 exhibited similar error patterns for the MTSamples dataset. Both models encountered challenges when it came to identifying correct entity boundaries. This typically involved making decisions on whether to include article words (such as ‘the’ in the phrase ‘the study drug’) or modifiers (such as ‘another large’ in the phrase ‘another large stroke’) that precede a noun phrase. In assessing model performance, we considered the exact-match criteria, which may present a different challenge for GPT models compared to BioClinicalBERT. While BioClinicalBERT is fine-tuned specifically on annotated entities with clear boundaries, the GPT models, being large language models, are trained on a broader and more diverse corpus. This distinction could impact their ability to adhere strictly to the exact boundaries of entities as defined in the training data, especially in the context of clinical NER where the linguistic structure and terminology are highly specialized. As for the VAERS dataset, several factors may contribute to its increased complexity. Firstly, inner-annotator agreement was lower compared to the MTSamples dataset (i.e., average F1 0.7707 ³¹ vs 0.8620), indicating less consistency in annotations. Additionally, the VAERS dataset contains more semantically specific annotation categories, such as distinguishing between different types of adverse events. This specificity demands a higher level of contextual understanding from the models. On the other hand, GPT-4’s major difficulties lie in determining the correct entity boundaries and accurately classifying the entity types. This discrepancy can be attributed to the unique characteristics of each dataset. The VAERS dataset contains more complex entities (i.e., Nervous adverse events vs Other adverse events) compared to the MTSamples dataset, leading to a higher error rate in entity type classification for the models. Another possible reason could be the inconsistency ³¹ in annotation, which needs further investigation.

Our study hints at the as-yet unrealized potential of LLMs in clinical NER tasks by proposing a clinical task-specific prompt framework that incorporates annotation guidelines, error analysis-based instructions, and few-shot examples. We found that the performance of GPT models improved with the task-specific prompts. The best performance achieved by GPT-4 shows a competitive performance as that of BioClinicalBERT in the relaxed-match criteria.
LLMs are making paradigm-shifting changes in NLP research and development. Our finding shows a quick and easy path to build more generalizable clinical NER systems by leveraging LLMs. This will significantly change our current practice in clinical NLP. Traditionally, to build a machine learning or deep learning-based NER system for specific types of clinical entities, we have to build an annotated corpus of clinical documents, which is time-consuming and costly, as it often requires medical domain experts. Remarkably, our research shows that LLMs, devoid of further model training or fine-tuning, have exhibited exceptional performance. With merely 1- or 5-shot annotated samples, these models can achieve performance that is close to the fine-tuned models that require hundreds of training samples. This suggests a potential reduction in some of the costs associated with clinical NER system development, particularly in the areas of data annotation. However, it is important to note that this does not eliminate the need for expert input in creating annotation guidelines and in the initial phases of model training. While our study demonstrates that GPT models can achieve competitive performance with fewer annotated examples compared to traditional NLP systems, the role of subject matter experts remains crucial. Experts are needed to write precise annotation guidelines, perform initial annotations for error analysis and example generation, and validate the model’s performance. Although the GPT models require fewer annotated instances, the costs associated with expert involvement, API usage, and running an LLM service should not be overlooked. A comprehensive comparison of resource requirements and costs between traditional NLP systems, word embedding models, and LLM-based systems would be valuable for future studies. This will provide a clearer understanding of the practical implications and feasibility of deploying LLMs in clinical NER tasks.
Moreover, our approach is generalizable – it shows consistent performance improvements across two different clinical NER tasks. The emergent abilities of LLMs ³⁶ have been further demonstrated in multiple clinical NER tasks here, indicating the feasibility of building one large model for diverse information extraction tasks in the medical domain, which is very appealing.
With those changes in mind, an urgent need will be to re-design the workflow for developing clinical NER systems using LLMs. The prompt framework for those two clinical NER tasks is the first step toward this direction and it sheds some lights for several aspects that are worth considering. The first aspect is how to clearly define an information extraction task. Our experiments show specific annotation guidelines are very helpful, which indicates medical knowledge (either in a knowledge base or from human experts) are still critical in LLMs-based NER systems and how to obtain and represent task-specific knowledge in prompts need further investigation. We also demonstrated that supplying annotated examples is effective for performance improvement. Nevertheless, how to select informative and representative samples have not been investigated in this study and other advanced few-shot learning algorithms could be explored.
Another important issue is evaluation. In this study, we instructed GPT models to output entities following traditional NER approaches so that we can evaluate them using the previous evaluation scripts. However, we would argue that the current evaluation schema for NER may not be ideal for LLMs-based systems. GPT models, due to their generative nature and extensive pre-training on diverse text corpora, exhibit a nuanced understanding of context and language structure. This enables them to interpret and generate text in a way that sometimes extends beyond the strict boundaries of predefined entity classes. For instance, GPT models often recognized lab tests with abnormal values (e.g., “a blood sugar level of 40” or “white blood cell count of 23,500”) as medical problems. While this interpretation is contextually relevant and clinically meaningful, it deviates from the strict entity definitions used in our evaluation, leading to apparent mismatches. Therefore, a better evaluation schema would be needed to assess LLMs performance more accurately.
Despite the promising results, our study has some limitations. First, we limited LLMs to GPT models in this study. In future, we will include other popular LLMs such as LLaMA and Falcon ^{37; 38; 39}. Second, our few-shot learning approaches were relatively simple, and we plan to investigate other approaches such as the chain-of-thoughts method ^{40; 41; 42}, hoping to yield better results.

This is one of the first studies that systematically investigated GPT models for clinical NER via prompt engineering. In this study, we proposed a clinical task-specific prompt framework by incorporating annotation guidelines, error analysis-based instructions, and annotated samples via few-shot learning, and our evaluation on two clinical NER tasks show that the GPT-4 model with our proposed prompts achieved close performance as the state-of-the-art BioClinicalBERT model. The best performance achieved by GPT-4 with 5-shot learning did not work as well as the BioClinicalBERT model on MTSamples and VAERS datasets. Nevertheless, considering almost no training data was used in GPT, its performance is impressive hints the potential of LLMs in clinical NER tasks. While the results demonstrate a promising direction, they also underscore the need for further refinement and development before LLMs can consistently outperform established models like BioClinicalBERT in these specific applications.

This work was supported by NIH grant number R21EB029575, R21AI164100, R01LM011934, 1K99LM01402, R01AG066749, R01AG066749-03S1, R01LM013712, and U01TR002062; NIA grant number 1RF1AG072799,
1R01AG080429; CPRIT grant number RR180012; NSF grant number 2124789.

Dr. Hua Xu and Dr. Jingcheng Du have research-related financial interests at Melax Technologies Inc.

Our code and datasets are available at Github ²²2https://github.com/BIDS-Xu-Lab/Clinical_Entity_Recognition_Using_GPT_models.