Two-Pronged Human Evaluation of ChatGPT Self-Correction in Radiology Report Simplification

Traditionally, text simplification referred to lexical simplification that paraphrases text Chen et al. (2018); Biran et al. (2011); Weng et al. (2018). More recently, it shifted towards semantic simplification that seeks to simplify grammatically complex text Shardlow (2014); Leroy et al. (2016). This paper adopts this more novel emphasis. Plain language summarization Guo et al. (2021); Devaraj et al. (2021) is an alternative term that reminds that the main objective is to enhance laypeople’s understanding of expert-written texts. We think that summarization is not the best term because it implies text compression, while text simplification is more interested in text understanding, which allows creation of text longer than the original.

Radiology text simplification using LLMs has recently drawn significant attention Ondov et al. (2022). A recent work used fine-tuned BART Lewis et al. (2019) to simplify 140 liver-related radiology sentences Yang et al. (2023). In Jeblick et al. (2022); Lyu et al. (2023), researchers explored the use of prompt learning with the GPT family Brown et al. (2020), including ChatGPT-3.5 and ChatGPT-4, to simplify radiology reports. Jeblick et al. (2022) focused on three artificial reports, while Lyu et al. (2023) considered over 100 reports. However, they did not provide fully coherent evaluation of the simplifications.

There are two related NLP problems that have been popular in radiology. Radiology report generation refers to automated creation of reports from X-ray or other radiographic images Liu et al. (2023a). This is an image-to-text task with a different set of objectives from text simplification. Radiology report summarization refers to condensing the detailed "Findings" section of radiology reports into a succinct "Impression" section Zhang et al. (2018). This involves creating a shorter version of the report that retains all critical information without a necessity to make it clearer to laypeople Chaves et al. (2022); Liang et al. (2022).

Assessing output of LLMs is integral to text simplification Van den Bercken et al. (2019); Cripwell et al. (2022) and related natural language generation tasks. In text simplification, automatic metrics such as ROUGE Lin (2004) and BERTScore Zhang et al. (2019) have been popular, which compare similarity between gold standard and generated sentences. SARI Xu et al. (2016) compares simplified text both with reference simplifications and the original sentences, thus assessing the operation of adding, deleting, and keeping words. Unfortunately, these metrics often correlate poorly with human evaluation of text simplification Alva-Manchego et al. (2021); Liu et al. (2023b); Guo et al. (2023). For readability assessment, the Flesch-Kincaid Grade Level (FKGL), Gunning Fog Index (GFI), and Automated Readability Index (ARI) are widely recognized metrics that estimate the text’s reading difficulty. More recently, Guo et al. (2023) proposed to assess readability by using difference in normalized perplexity scores from in-domain and out-of-domain language models.

Using human evaluators has been increasingly popular in text simplification, despite the significant associated costs. Researchers typically evaluate fluency, adequacy, factuality, and simplicity of the simplified texts Jiang et al. (2020); Cripwell et al. (2022). Very often, these measures are vaguely defined and subject to interpretation. Recently, factuality was formalized in terms of addition, substitution, and deletion of information Devaraj et al. (2022).

In radiology report simplifications, there is no clear standard for evaluation. Jeblick et al. (2022) enlisted 15 radiologists to assess simplified reports for factual correctness, completeness, and potential harm. Lyu et al. (2023) invited two radiologists to evaluate the simplified reports based on metrics such as information loss, misinterpretation, and an overall score. Interestingly, these studies did not evaluate the simplicity of the text. Lu et al. (2023) focused on simplicity, fluency, and factual accuracy. The study recruited students to assess factualness and simplicity and two medical experts to examine the factual consistency. However, it is unclear if the participating students possessed any medical expertise to represent laypeople and if they were qualified to assess the factualness.

Modern LLMs can solve various NLP tasks with high success through prompting and without necessitating fine-tuning Brown et al. (2020). The quality of output is very sensitive to prompting. While prompting is sometimes considered an art form, there are a few strategies that work more often than not. One is Chain-of-thought (CoT) Wei et al. (2022). Another is self-correction Chen et al. (2023); Madaan et al. (2023). Huang et al. (2022) self-improves an LLM through iterative fine-tuning. Bai et al. (2022) leverages AI-generated feedback through reinforcement learning. Li et al. (2023) allows LLMs to self-improve their generations without training. They instantiate multiple LLMs models as different agents and let them collaborate towards better generation.

Factuality refers to correctly maintaining the original information. Motivated by Devaraj et al. (2022) and Jeblick et al. (2022), we measure three aspects of factuality.

• •

  Correctness: (Factualness/Substitution) Evaluates whether the simplification correctly interprets the information in the original sentence.

• •

  Completeness: (Adequacy/Meaning preservation/Deletion) Evaluates if there is any significant information loss in the simplification compared to the original text. Simplifications should retain all critical information from the original text, but it might be permissible to ignore less important information.

• •

  Hallucination: (Addition) Evaluates if simplifications contain wrong statements or hallucinate new information that may misguide laypeople.

We introduce a new measure that is related to Completeness.

• •

  Structure: Refers to a desire that simplifications follow a certain structure. Specifically, a good radiology simplification should mention: body parts, findings, and consequences. Body parts specify the anatomies and organs referred to in the radiology sentence (such as kidneys). Findings refer to the key observations in the radiology sentence (such as injuries or masses). Consequences refer to what findings indicate, which might not be explicitly stated in the original sentence, such as severity, certainty, and follow up.

Only experts can adequately evaluate factuality and structure. Appendix A.1 shows the exact survey design for expert evaluation of simplifications we used in our experiments. Note that we also ask the radiologists to evaluate the simplicity of generated simplifications for our analysis.

In prior work, simplicity mostly refers to readability, which measures text fluency and complexity of terms and grammar correctness. However, LLMs typically generate very fluent text, so evaluating that aspect is not very informative. Instead, it is more relevant to measure how well laypeople comprehend the text.
Clarity: Instead of asking evaluators to provide a single score for the simplicity Jiang et al. (2020), we evaluate their understanding by devising a set of questions to measure the usefulness of simplifications. The critical objective of radiology report simplification is to improve the clarity about the severity of the described conditions. There are two important dimensions of clarity: how well people understand the text and how well they believe they understand. Different combinations of those two dimensions can have different consequences for patients. For example, being confident while misinterpreting the text might lead to being too concerned or relaxed. Uncertainty is a clear indication that simplification was not adequate. Our survey is sequenced as in Figure 1. We ask laypeople if they think they understand the original text (4 levels). Then, we ask them specifically if they think they understand the severity of the described condition (3 levels). This is followed by asking them to guess the severity, according to 5 severity levels defined in Appendix A.2. This allows us to compare with the actual severity provided by a radiologist. We repeat those questions by suplementing the original sentence with simplification. Finally, we also ask them about their subjective opinion about the helpfulness of the simplification.

We considered other ways to measure how well laypeople understand the text, such as quizzing them about the body parts and the meaning of the findings. We decided against it because it would be cumbersome to consistently convert the responses into numbers given a wide variety of radiology sentences. Also, asking this question would compound health literacy and simplicity. For example, even if a patient cannot fully understand the medical meaning, it could still be essential to hear that a condition impacting some part of their abdomen is not critical but requires a follow-up.

Preferences: Inspired by the design for evaluating text summarization Goyal et al. (2022), we also ask evaluators to choose the most and the least preferred simplifications among multiple choices. In addition, layperson evaluators are encouraged to provide justifications for their selections, as shown in the right panel of Figure 1. This free-text response can be used in qualitative analysis of laypeople’s simplification preferences.

Our preliminary results showed that ChatGPT can provide good simplifications of radiology sentences. Since we did not have a sufficiently large corpus of parallel text for radiology report simplification, we opted to use prompting without any fine tuning. Due to costs, we used ChatGPT-3.5 for all experiments in our study.

Prompt selection is partly an art form, so it was beyond the scope of this paper to comprehensively search for the best prompt for this application. Instead, we constructed two representative prompts after some trial and error – one very simple (Plain) and another that relies on the Chain-of-thought (CoT) strategy, which makes ChatGPT think aloud while generating a response. All designed prompts can be found in Appendix A.3.1.

Inspired by Madaan et al. (2023), we devised a Self-Correction mechanism for radiology report simplification. It relies on four differently instantiated ChatGPT agents: Generator, Radiologist, Patient, and Processor. The proposed workflow is shown in Figure 2. Given an original radiology sentence, Generator is asked to generate a simplification. Then, Radiologist and Patient provide feedback about the simplification. Finally, Processor summarizes the feedback and provides the summary to Generator who is asked to improve the simplifications. This process iterates among these four agents until Processor determines that no further improvement is needed. This self-correction mechanism can be applied to LLMs without any model training.

Inspired by Park et al. (2023), we instantiated Radiologist and Patient agents as distinct personas through distinct initial prompts shown in Appendix A.4. On the other hand, Generator and Processor agents are not prompted to become personas and are asked to provide an objective output. They are initialized using prompts that specify the task. Generator keeps the memory of conversation since it needs to refine the simplification based on the feedback from other agents. Generator is first provided a simple prompt for simplification.

Feedback generated by Radiologist and Patient agents is summarized by Processor to reduce the redundancy. We asked Processor agent to first decide if there is any critical comment or improvement suggestion in the generated feedback. If so, Processor summarizes the feedback and passes it back to Generator using a ’refine prompt’. Otherwise, Processor generates a string starting with "No". In this case, the last simplification is saved as the self-correct simplification. The prompts are shown in Appendix A.3.4.

The proposed variant of self-correction mechanism is designed to imitate a conversation that could occur between a real radiologist and a patient to generate a good simplification of a radiology report.

For our experimental evaluation, we identified 40 diverse, representative sentences from the radiology reports in the public database MIMIC III Johnson et al. (2016). To arrive at those 40 sentences, a radiologist read 100 randomly selected MIMIC III radiology reports and marked self-contained sentences about findings. The list of marked sentences was then narrowed down by removing redundancy and trying to maintain a diversity of findings in the reports. The selected sentences include a variety of findings (lesions, masses, obstructions, nodular surface, infection), conditions (being enlarged, abnormal, shrunken) of many anatomical parts in the abdomen (liver, kidney, pancreas, intestine, bones). Sentences were selected to range from relatively simple to relatively complex. Attention was paid to ensuring that the chosen sentences were self-contained and did not require reading the surrounding sentences

For each of the 40 radiology sentences, we produced four simplifications using ChatGPT-3.5. The first is Plain_BS, which uses the plain prompt, while the second is CoT_BS, which uses the CoT prompt, both introduced in Section 4.1. The remaining two use self-correction explained in Section 4.2. The initial Generator prompt in self-corrected Plain_SC is the plain prompt, while it is CoT prompt in CoT_SC. We used the same default temperature value for ChatGPT of 0.8 in all generations.

In our study we did not generate ground truth simplifications because there might be a variety of acceptable simplifications with different lengths and levels of detail. As a result, our automated metrics only include three reference-free measures of simplicity: Flesch-Kincaid Grade Level (FKGL), Gunning Fog Index (GFI), and Automated Readability Index (ARI). The FKGL gives a grade-level score, indicating the minimum education level needed to comprehend the text. The GFI estimates the years of formal education required for understanding by focusing on sentence length and the frequency of complex words. The ARI, on the other hand, calculates readability based on the number of characters per word and words per sentence.

As described in Section 3, we used two types of human evaluators to assess the quality of simplifications.

Radiologists. We recruited one radiologist to evaluate the factuality of all simplifications as described in Section 3.1. For further analysis, we also asked them to evaluate the simplicity via the question, "Do you think laypeople can understand the sentence?". Likert scores in the range 1-5 were used for all questions. In addition, the radiologist was encouraged to provide justifications for the ratings. Moreover, we asked the radiologist to estimate the severity of described medical condition in each sentence using the five levels of severity as described in the Appendix A.2. The distribution of severity scores on our data is: Critical: 4, Serious: 2, Moderate: 7, Mild: 20, Healthy: 7. The severity question is the same as Q3 in the survey for laypeople. This allowed us to evaluate the accuracy of laypeople’s guesses of severity.

Laypeople. We recruited eight laypeople to assess if the simplifications improve understanding. The participants were a mix of undergraduate and graduate students from a computer science department, none of whom had any training in medicine. Thus, they are representatives of highly-educated laypeople. For each of the 40 sentences, each layperson was given the questionnaire in Figure 1. First, they answered 3 questions about the original sentence (left panel). Then, we selected one of the four simplified sentences using the Latin square design and asked them 4 questions from the middle panel to see if it improved their understanding compared to the original sentence alone. As the layperson was already starting to understand the original sentence, we did not repeat the middle panel questions for the 3 remaining simplifications. Instead, we asked a layperson the right panel questions to find which simplifications they liked the most and least. The simplifications in the right panel were listed at random to prevent bias.

Selected sentences, generated simplifications, and evaluation answers from the radiologist and laypeople are released.¹¹1https://github.com/Ziyu-Yang/Human-Evaluation

Top half of Table 1 shows results from the evaluation conducted by a radiologist. The ’Original Sentence’ column denotes the scores assigned to the original radiology sentences. The factuality of the original radiology sentences was rated as five, by default. Simplicity score for the original sentences was very low (1.50), indicating that most of the original sentences are not expected to be understood by laypeople. Simplicity score was much larger for simplified sentences and was the largest for the self-correction with CoT (CoT_SC) approach. This result is consistent with the automated readability scores in the first three rows of Table 1. It can be seen that readability of all 4 simplifications is significantly smaller (freshmen high school level) than for the original sentences (college level).

Factuality scores for all four types of simplifications remained close to perfect. Hallucination and Structure scores were particularly high. Correctness scores were comparably lower, indicating occasional lack of precision in simplifications. Interestingly, factuality scores of Plain_BS are higher than for the other three simplification methods. This reflects the trade-off between simplicity and factuality. We consider CoT_SC the best approach because it achieved the highest simplicity with a very marginal decrease in factuality.

In the bottom half of Table 1, we evaluate laypeople responses about simplicity. Q1, Q2, and Q3 in Figure 1 were designed to assess laypeople understanding of both the original sentences and their simplified versions. Q4 directly evaluated the effectiveness of these simplifications. We converted the categorical responses into numerical values²²2Q1: ’Not at all’ -¿ 1; ’Completely’ -¿ 4. Q2: ’Not at all’ -¿ 1; ’High confidence’ -¿ 3. Q3: ’Critical’ -¿ 1; ’Healthy’ -¿ 5. Q4: ’Furthered confused’ -¿ -1; ’Much better’ -¿ 2. For the responses to Q3, we compared the participants’ severity level choices with those of the radiologist and computed the Mean Squared Error (MSE) and Accuracy (ACC).

All simplifications had significantly higher simplicity scores than the original sentences on all questions. Notably, CoT_SC achieved the highest scores accross all simplicity questions which is consistent with the radiologist’s rating.

Table 2 compares the correlation between the laypeople’s confidence and the actual understanding of the severity of described medical conditions. When laypeople report the lowest confidence (Not at all), they also achieve the lowest accuracy in predicting severity (30.7%), and when they report the highest confidence, the accuracy is the largest (55.5%). However, there is still a significant gap between confidence and actual understanding. Even when highly confident, laypeople could correctly predict severity in just over half (55.5%) of the sentences. We conclude that the simplifications might need to state the severity level explicitly.

To gain a deeper insight, in Figure 3, we show the distribution of confidence levels by laypeople for the original sentences and for each type of simplifications. We can see that all four types of simplifications are helpful, with CoT_SC being the most successful.

In this subsection, we report on the preferences of laypeople towards different types of simplifications (right panel in Figure 1). The findings are shown in Table 3, illustrating how often a specific simplification was deemed the most or least preferred based on the majority vote by the eight participants. The CoT_SC simplification was the clear favorite compared to the other three variants. On the other hand, Plain_BS simplification was the least favorite.

To further investigate laypeople’s preferences, we adopted the analysis technique outlined in Goyal et al. (2022). We calculated the inter-annotator agreement, applying Krippendorff’s alpha with MASI distance Passonneau (2006), to account for the possibility of multiple selections for the best or worst simplifications in our research design. The alpha scores for the most and least preferred options were $0.177$ and $0.132$ respectively.

To visualize these results more clearly, we plot the distribution of laypeople’s votes for each simplification version in Figure 4. The distribution of votes highlights that CoT_SC was chosen as the most preferred simplification by at least half of the participants in over 90% of the sentences. Additionally, both CoT_BS and Plain_SC received votes as the most preferred simplification from at least one participant for approximately 80% of the sentences. Unsurprisingly, based on the vote distribution, Plain_BS was the least preferred simplification version.

ChatGPT is providing impressive simplifications, but it is not perfect. We observed that the slight decrease in factuality scores was caused by a few outliers. We noticed that during the self-correction process, Generator agent can occasionally generate incorrect information and hallucinate as it is trying to addresses feedback from Radiologist and Patient agents. There were in total eight factual errors observed by the radiologist evaluator among simplifications by the two self-correction approaches.

In Table 4, we show and analyze four examples to point to typical errors. ChatGPT may generate hallucinations through over-interpreting the sentences. We can also see that ChatGPT may not use perfectly precise simple words to explain the complex medical jargon. However, ChatGPT is overall very impressive in producing high-quality text simplifications and is quite capable of self-correction.

In Table 5, we show five examples from laypeople responses and discuss why CoT_SC could enhance comprehension. These examples are selected because they are representatives of sentences with different severity levels. We observe that participants prefer a simplification that 1) explains the medical condition in detail, 2) uses simple language, 3) indicates the severity of the condition. The CoT_SC simplification in the second example implies a mild severity level, which is not explicitly stated in the original sentence.