Оценка и анализ < < Биаса > > крупных языковых моделей для подготовки синтетических электронных медицинских записей: сравнительное исследование.

Introduction

Background and Objectives

The integration of artificial intelligence (AI) in health care has opened up new possibilities, especially with the advent of large language models (LLMs) capable of generating synthetic electronic health records (EHRs) [ 1 ]. These synthetic records hold significant promise for clinical education and model training, offering a way to mitigate privacy concerns while still providing realistic, diverse patient data for medical research and training purposes [ 2 - 4 ]. However, the widespread deployment of synthetic EHRs in clinical practice is not without its challenges, particularly regarding the performance and biases inherent in these models. The main research question of this study is whether, despite their ability to generate high-quality synthetic EHRs, LLMs inadvertently introduce significant gender and racial biases into the data. In addition, it seeks to explore whether the size of the model and the real-world disease distribution have any correlation with these biases.

This study aimed to evaluate the performance of multiple LLMs in generating synthetic EHRs, with a focus on the completeness and demographic representativeness of the generated records. We hypothesized that larger models, while achieving better performance in EHR generation, are likely to exhibit greater gender and racial biases, reflecting the limitations of their training data and underlying architecture. In particular, we aimed to investigate the relationship between model size, real-world disease distribution, and demographic biases.

Literature Review and Research Gap

Previous research has explored the potential of synthetic EHRs for clinical education, highlighting their role in training medical professionals and improving health care delivery by offering access to large, diverse patient datasets [ 5 - 7 ]. However, little attention has been paid to the biases that these models may perpetuate, especially regarding gender and race [ 8 - 11 ]. These biases threaten not only the fairness and accuracy of case analyses but also risk exacerbating existing societal disparities [ 12 - 18 ]. While studies have demonstrated that LLMs can reproduce known societal biases, including gender and racial disparities, few have systematically investigated these biases in the context of synthetic EHR generation. For example, research has shown that models like GPT-4 often exhibit gender biases, such as overrepresenting male patients in certain medical scenarios, despite the real-world prevalence of conditions like HIV and multiple sclerosis, which have gender-specific distribution patterns [ 19 ]. This research sought to fill this gap by systematically analyzing the performance and biases of synthetic EHRs generated by LLM across a range of diseases with varying gender and racial prevalence. Specifically, we expanded upon existing literature by examining the following 3 key aspects: (1) whether racial and gender biases are widespread in synthetic EHRs generated by LLM; (2) the impact of model size on racial and gender biases within the synthetic EHRs; and (3) how the real-world distribution of diseases influences the racial and gender biases present in these synthetic records. By addressing these questions, we aimed to deepen our understanding of how the scale of LLMs and the inherent characteristics of disease prevalence may contribute to the perpetuation of demographic biases, ultimately influencing the utility and fairness of synthetic EHRs in clinical and research settings.

Significance and Practical Implications

The practical implications of this work are significant. If LLMs used to generate synthetic EHRs are biased, they may inadvertently perpetuate health care disparities, reinforcing inequities in medical education and patient care. By addressing gender and racial biases in LLMs, we can develop more equitable models that better serve diverse patient populations, ensuring that synthetic data reflects the demographic realities of real-world health care.

Methods

Model Selection and Setup

To evaluate the performance and biases of LLMs in generating synthetic EHRs, we used a framework illustrated in Figure 1 . This framework outlines a systematic process for generating EHRs and conducting information extraction and analysis. The process is divided into 3 key modules: prompt generation, EHR generation, and information extraction and analysis, as described in Textbox 1 .

Framework for evaluating the performance and bias of large language models (LLMs) in generating synthetic electronic health records (EHRs).

Systematic process for generating electronic health records (EHRs) and conducting information extraction and analysis.

Prompt generation

In total, 10 standardized template prompts were meticulously designed for this study to generate synthetic EHRs. Each template included a placeholder (disease), which was replaced by one of 20 distinct diseases spanning 5 categories: epidemics, chronic conditions, mental health disorders, rare diseases, and diseases associated with geographic or socioeconomic factors. These carefully crafted prompts ensured the generation of comprehensive and realistic EHRs containing patient demographics, clinical details, past medical history, and other relevant information. For detailed information on prompt construction and disease, please refer to Multimedia Appendix 1 [ 20 - 39 ].

EHR generation

To minimize the variability in outputs caused by differences in prompts and the inherent uncertainty of large language model responses, each model used the same set of 10 prompts per disease to generate synthetic EHRs. For each prompt, 100 cases were generated, resulting in a total of 1000 cases per disease. Across 20 diseases, each model generated 20,000 synthetic EHRs. With 7 models in total, the study produced a comprehensive dataset of 140,000 synthetic EHRs.

Information extraction and analysis

Once the synthetic EHRs were generated, patient attributes were extracted using custom-developed regular expressions. To ensure accuracy, the extracted attributes underwent a secondary round of manual verification. These attributes were then subjected to detailed performance analysis to evaluate the models’ ability to generate complete records, as well as bias analysis to examine gender and racial disparities in the generated outputs.

The 7 open-source LLMs selected for this study were chosen based on the considerations provided in Textbox 2 .

Considerations for the selection of open-source large language models.

Language capabilities reflecting diverse training corpora and cultural contexts

The selected models represented a variety of linguistic and cultural backgrounds. Monolingual English models, such as those from the Llama 2 series, provided a baseline for English-language performance. In contrast, multilingual models, such as those from the Yi and Qwen series, which are predominantly trained on Chinese data, enabled us to explore the influence of culturally and linguistically diverse corpora on generative capabilities and biases. This diversity was essential for understanding how language and culture affected the generation of synthetic EHRs.

Availability of multiple model sizes

The selected models spanned a range of sizes, from smaller models like Qwen-1.8B to larger ones like Yi-34B. This allowed for an examination of how model size influenced both performance and biases, providing insight into the trade-offs between computational efficiency and output quality.

Prominence in the open-source ecosystem

All selected models are widely recognized within the open-source community, ensuring transparency, reproducibility, and accessibility for future research. The performance of these models is often evaluated using well-established benchmarks, such as Massive Multitask Language Understanding (MMLU) and comprehensive Chinese evaluation suite for foundation models (C-Eval). MMLU scores, commonly referenced in English-speaking academic circles [ 40 ], and C-Eval, a widely used metric in Chinese-speaking domains [ 41 ], provide critical indicators of model performance.

By incorporating models with diverse linguistic capabilities, parameter sizes, and recognition in the open-source community, this study provides a comprehensive evaluation of LLM performance and biases in synthetic EHR generation. A detailed summary of the selected models, including their publishers, primary languages, sizes, and benchmark performances, is presented in Table 1 .

Table 1 Large language models’ information. Model Publisher Primary language Model size (billion) MMLU a score C-Eval b score Yi-6B 01.AI English and Chinese 6B 64.11 72 Yi-34B 01.AI English and Chinese 34B 79.4 81.4 Qwen-1.8B Alibaba Group English and Chinese 1.8B 45.3 56.1 Qwen-7B Alibaba Group English and Chinese 7B 58.2 63.5 Qwen-14B Alibaba Group English and Chinese 14B 66.3 72.1 Llama2-7B META English 7B 45.3 — c Llama2-13B META English 13B 54.8 — a MMLU: Massive Multitask Language Understanding. b C-Eval: comprehensive Chinese evaluation suite for foundation models. c Llama2-7B and Llama2-13B have not released their C-Eval test results, and no related records are found in the C-Eval rankings.

Large language models’ information.

a MMLU: Massive Multitask Language Understanding.

b C-Eval: comprehensive Chinese evaluation suite for foundation models.

c Llama2-7B and Llama2-13B have not released their C-Eval test results, and no related records are found in the C-Eval rankings.

The detailed explanation of the parameters in Table 1 is provided in Textbox 3 .

Detailed explanation of the large language model parameters.

Primary language

This indicates the main languages that the model is optimized for. For example, it specifies whether the model is primarily designed to perform well in English, Chinese, or both. It is essential to consider this when evaluating the model’s utility in multilingual or specific linguistic contexts.

Model size (billion)

It refers to the number of parameters (in billions) within the large language model. The parameter size reflects the model’s complexity and potential capacity for learning and generalization. Larger models typically perform better but require more computational resources.

Massive Multitask Language Understanding (MMLU) score

This benchmark is widely used in English-speaking academic circles to assess a model’s ability to perform well across a variety of tasks, including knowledge-intensive and reasoning tasks. A higher MMLU score indicates stronger performance in English-language tasks.

Comprehensive Chinese evaluation suite for foundation models (C-Eval) score

A benchmark designed to evaluate a model’s performance in Chinese tasks. It is commonly referenced in Chinese-speaking academic domains and includes tasks that test the model’s understanding and reasoning capabilities in the Chinese language. A higher C-Eval score reflects better proficiency in Chinese.

To support the operation of these models, we deployed 4 high-performance NVIDIA 3090 graphics cards and 2 A800 graphics cards and used the Python programming language (Python Software Foundation) to call the LLMs and obtain the generated results.

Ethical Considerations

The study was designed and conducted with strict adherence to ethical standards, without involving any human subjects, patient information, medical records, or observations of public behaviors. All data utilized in this research were synthetic and generated computationally, thereby eliminating the necessity for informed consent, privacy, confidentiality measures, or participant compensation. Accordingly, we confirmed that there was no requirement for institutional review board approval, in compliance with institutional policies regarding exemption from ethical review for research involving solely synthetic data.

Model Performance Evaluation

Synthetic EHR generation involves creating records that comprehensively cover key medical attributes (eg, sex, age, and medical history) to ensure their clinical relevance and usability. However, the generated outputs can vary in quality and completeness, leading to 3 distinct categories as follows:

Irrelevant Content ( N i ): Records that fail to align with the required medical attributes, often due to hallucinations, such as fabricating unrelated information or omitting critical details. These outputs are entirely unsuitable for clinical use.

Partially Relevant Content (Ns): Records that include some, but not all, of the specified attributes. While partially useful, these outputs lack the completeness required for fully reliable EHRs. A subset of these records (Ns–i) contains specific attributes Ai but still does not meet the full requirements.

Fully Relevant Content (Na): Records that accurately and completely include all specified attributes, representing the ideal output for synthetic EHR generation.

The total number of synthetic EHRs N produced is the sum of these scenarios:

To evaluate the performance of the EHR generation task in LLMs, we used 2 metrics: the electronic health record performance score (EPS) and the attribute-specific EPS (EPS i )

EPS measures the proportion of fully complete and accurate EHRs among all generated records. A higher EPS indicates the model’s ability to consistently generate reliable and clinically relevant outputs.

EPS i assesses the model’s capability to generate records that include a specific attribute A i This metric provides a more granular evaluation, helping to identify which attributes the model can reliably generate and which may require further optimization.

Due to constraints in scope and length, this study specifically examined the generative capabilities of LLMs concerning gender and racial attributes, hence defining A = { A gender , A race }. To assess the capabilities of LLMs in generating synthetic EHRs, we used the EPS, EPS gender , and EPS race .

Model Bias Evaluation

In this study, gender bias and racial bias are systematically defined and analyzed to assess their presence in synthetic EHRs generated by LLMs.

Gender bias is defined as the deviation in the distribution of male and female cases generated by LLMs compared to real-world gender prevalence for specific diseases. Similarly, racial bias refers to the deviation in the distribution of cases for different racial or ethnic groups compared to real-world racial prevalence. Both biases occur when the proportion of synthetic EHRs significantly diverges from actual epidemiological data.

To detect such discrepancies, chi-square tests were used, with a P value <.05 indicating statistically significant bias. Real-world prevalence data for 20 diseases in the United States, obtained through an exhaustive literature review [ 20 - 39 ], were used as benchmarks for evaluation.

When significant discrepancies were identified, statistical parity difference (SPD) was used to quantify bias and classify groups as either:

Overrepresented: SPD>+10%, where the generated proportion significantly exceeds the real-world prevalence.

Underrepresented: SPD<–10%, where the generated proportion falls significantly below the real-world prevalence.

SPD is calculated as:

Where P generated is the proportion of cases generated for a group, and P real is the real-world prevalence.

We further introduced the gender polarization effect and racial polarization effect to describe the tendency of LLMs to increasingly favor one gender or racial group as model size increases. This effect is characterized by a convergence toward a bias-polarized gender or bias-polarized race, where one group becomes disproportionately overrepresented while others are underrepresented.

For example, in diseases with balanced sex distributions (eg, hypertension: 72,926,629/152,466,669, 47.83% female vs 79,540,040/152,466,669, 52.17% male), larger models may disproportionately generate male cases, designating the male group as the bias-polarized sex. Similarly, in female-dominated diseases like lupus (29,578/33,145, 89.24% female vs 3567/33,145, 10.76% male), the overrepresentation of the female group may intensify as model size increases.

A similar pattern is observed for racial bias. In diseases with balanced racial distributions, larger models may overrepresent one racial group (eg, White individuals) while underrepresenting others (eg, Black or Hispanic populations), designating the overrepresented group as the bias-polarized race.

Results

Model Performance

There was a clear depiction of the performance metrics across different models in Table 2 , which quantitatively assessed each model’s capability to generate synthetic EHRs.

Table 2 Performance metrics across different models. Model Model size (billion) EPS a Race(EPS) b Gender(EPS) c MMLU d C-Eval e Yi-34B 34 96.8 96.84 98.82 74.9 81.4 Llama2-13B 13 93.37 93.48 94.63 54.8 — f Llama2-7B 7 92.72 92.9 93.56 45.3 — Qwen-7B 7 90.93 91.1 97.29 58.2 63.5 Qwen-14B 14 88.81 88.82 98.99 66.3 72.1 Yi-6B 6 77.31 78.12 85.37 64.11 72 Qwen-1.8B 1.8 63.35 65.05 83.31 45.3 56.1 a EPS: electronic health record performance score. b Race(EPS):EPS race . c Gender(EPS):EPS gender . d MMLU: Massive Multitask Language Understanding. e C-Eval: comprehensive Chinese evaluation suite for foundation models. f Llama2-7B and Llama2-13B have not released their C-Eval test results, and no related records are found in the C-Eval rankings.

Performance metrics across different models.

a EPS: electronic health record performance score.

b Race(EPS):EPS race .

c Gender(EPS):EPS gender .

d MMLU: Massive Multitask Language Understanding.

e C-Eval: comprehensive Chinese evaluation suite for foundation models.

f Llama2-7B and Llama2-13B have not released their C-Eval test results, and no related records are found in the C-Eval rankings.

The analysis of EPS scores across models indicated a clear correlation between the size of the models and their effectiveness in generating synthetic EHRs. Larger models consistently showed higher EPS values, suggesting enhanced performance in EHR generation tasks. For instance, the Yi-34B model, one of the largest models evaluated, boasted the highest EPS at 96.8, followed closely by its EPS race at 96.84 and EPS gender at 98.82. This was significantly higher compared to the smallest model, Qwen-1.8B, which scored only 63.35 in EPS, 65.05 in EPS race , and 83.31 in EPS gender . These figures explicitly demonstrated that larger models were more capable of producing comprehensive and detailed EHRs.

The EPS, along with its race and gender derivatives, offered detailed insights into each model’s precision in generating specific metrics within synthetic EHRs. When comparing models of similar sizes, it was evident that English LLMs, such as the Llama2 series, excelled in producing higher accuracy in racial data. The Llama2-7B, for instance, scored a 92.9 in EPS race , outperforming similarly sized models like Qwen-7B and Yi-6B, which recorded EPS race scores of 91.1 and 78.12, respectively. This pattern highlighted the capability of larger English LLMs to handle racial diversity effectively within synthetic EHRs, providing a nuanced and accurate portrayal of diverse patient demographics.

Traditional metrics, such as Massive Multitask Language Understanding (MMLU) and comprehensive Chinese evaluation suite for foundation models (C-Eval), were typically used to assess the general cognitive capabilities of models. However, these metrics did not always align with the specialized performance insights provided by EPS. For example, despite having a moderate MMLU score of 54.8, the Llama2-13B achieved a significantly higher EPS of 93.37, indicating that while the model may not have scored the highest on general tasks, it excelled in the specific task of EHR generation. Similarly, the Yi-6B model had a relatively high MMLU score of 64.11 but an EPS of only 77.31, suggesting that higher general cognitive scores did not necessarily translate to better performance.

Gender Bias

Gender bias was found to be pervasive in all models, regardless of their size. We used the chi-square test to assess whether the distribution of the generated clinical EHRs aligned with the objective distribution, using a significance level of 0.05. The detailed results, including P values, chi-square values, and df , are presented in Tables S1-S7 in Multimedia Appendix 2 (chi-square analysis results), with the key findings summarized in Table 3 In total, 17 diseases exhibited significant differences across the 5 models, while 14 diseases demonstrated significant differences across the 7 models ( P <.05) as shown in Table 3 . Only 2 gender-related diseases did not show significant differences.

Table 3 Gender bias results of 7 large language models for 20 diseases. Model EPS a Number of diseases with gender bias Diseases without gender bias Llama 2-7 B 92.72 18 Preeclampsia and prostate cancer Llama 2-13 B 93.37 18 Preeclampsia and prostate cancer Qwen-14B 88.81 18 Preeclampsia and prostate cancer Qwen-7B 90.925 18 Preeclampsia and prostate cancer Yi-34B 96.8 17 Preeclampsia, prostate cancer, and multiple sclerosis Yi-6B 77.31 14 Preeclampsia, prostate cancer, major depressive disorder, sarcoidosis, rheumatoid arthritis, and takotsubo cardiomyopathy Qwen-1.8B 63.35 14 Preeclampsia, prostate cancer, colon cancer, hypertension, lupus, and sarcoidosis a EPS: electronic health record performance score

Gender bias results of 7 large language models for 20 diseases.

a EPS: electronic health record performance score

Furthermore, a comparison of Table 3 and the EPS values revealed that the Yi-6B and Qwen-1.8B models exhibited fewer biases in the generated cases. This was primarily due to these models generating an excess of invalid data. Consequently, gender bias was a widespread issue in the task of generating synthetic EHRs.

As model size increased, gender bias became more pronounced. This trend was observed across multiple diseases, with larger models showing more extreme gender imbalances when compared to the real-world gender distributions as shown in Tables 4 - 6 . Although the trend was not statistically significant, the SPD, used to quantify gender bias, revealed a consistent pattern: larger models generated increasingly polarized gender distributions, especially when compared to the actual gender prevalence for each disease.

Table 4 Gender bias distribution in Qwen-1.8B (1.8 billion parameters), Qwen-7B (7 billion parameters), and Qwen-14B (14 billion parameters). Diseases and gender Qwen-1.8B Qwen-7B Qwen-14B Actual ratio (%)

Ratio (n=1000, %) SPD a (%) Ratio (n=1000, %) SPD (%) Ratio (n=1000, %) SPD (%)


Lupus


Not available 15.3 — b 3.3 — 1.6 — 0

Male 7.5 –3.2 1.1 –9.6 0.5 –10.2 c 10.7

Female 77.2 –12.1 c 95.6 +6.3 97.9 +8.6 89.3
Takotsubo cardiomyopathy


Not available 11.8 — 2.5 — 0.5 — 0

Male 23.8 +10.7 c 0.4 –12.7 c 0.6 –12.5 c 13.1

Female 64.4 –22.5 c 97.1 +10.2 c 98.9 +12.0 c 86.9
Multiple sclerosis


Not available 11.7 — 2.9 — 1.9 — 0

Male 16.6 –7.6 5.3 –18.9 c 0.8 –23.4 c 24.2

Female 71.7 –4.1 91.8 +16.0 c 97.3 +21.5 c 75.8
Rheumatoid arthritis


Not available 14.5 — 2.3 — 0.7 — 0

Male 16.1 –9.8 1.4 –24.5 c 0.6 –25.3 c 25.9

Female 69.4 –4.7 96.3 22.2 c 98.7 +24.6 c 74.1
Major depressive disorder


Not available 9 — 2.9 — 0.9 — 0

Male 24 –12.7 c 10.7 –26.0 c 0.6 –36.1 c 36.7

Female 67 +3.7 86.4 23.1 c 98.5 +35.2 c 63.3
Huntington disease


Not available 10 — 1.1 — 0.7 — 0

Male 62.1 +12.1 c 87.9 +37.9 c 75.1 +25.1 c 50

Female 27.9 +22.1 c 11.0 –39.0 c 24.2 –25.8 c 50
Bacterial pneumonia


Not available 14.2 — 2.4 — 1.1 — 0

Male 28.7 –22.3 c 88.6 +37.6 c 97.8 +46.8 c 51

Female 57.1 +8.1 9.0 –40.0 c 1.1 –47.9 c 49
Hypertension


Not available 8.8 — 2.4 — 1.0 — 0

Male 46.8 –5.4 92.6 +40.4 c 97.3 +45.1 c 52.2

Female 44.4 –3.4 5 –42.8 c 1.7 –46.1 c 47.8
Amyotrophic lateral sclerosis


Not available 18.5 — 4.3 — 1.6 — 0

Male 33.6 –20.2 c 80.5 +26.7 c 94.8 +41.0 c 53.8

Female 47.9 +1.7 15.2 –31.0 c 3.6 –42.6 c 46.2
COVID-19


Not available 11.5 — 3 — 0.3 — 0

Male 27.7 –27.1 c 79.4 +24.6 c 99.2 +44.4 c 54.8

Female 60.8 +15.6 c 17.6 –27.6 c 0.5 –44.7 c 45.2
Multiple myeloma


Not available 13.9 — 2 — 0.7 — 0

Male 32.9 –22.6 c 94.9 +39.4 c 99.3 +43.8 c 55.5

Female 53.2 +8.7 3.1 –41.4 c 0 –44.5 c 44.5
Colon cancer


Not available 8.2 — 2.1 — 1.4 — 0

Male 52.0 –5.1 93.4 +36.3 c 98.6 +41.5 c 57.1

Female 39.8 –3.1 4.5 –38.4 c 0.0 –42.9 c 42.9
Tricuspid endocarditis


Not available 10.4 — 0.8 — 0.2 — 0

Male 36.5 –21.5 c 92.4 +34.4 c 94.4 +36.4 c 58

Female 53.1 +11.1 c 6.8 –35.2 c 5.4 –36.6 c 42
Hepatitis B


Not available 15.3 — 3.3 — 0.9 — 0

Male 36.1 –23.7 c 96.3 +36.5 c 99 +39.2 c 59.8

Female 48.6 +8.4 0.4 –39.8 c 0.1 –40.1 c 40.2
Tuberculosis


Not available 19.7 — 2.7 — 0.8 — 0

Male 33.5 –27.7 c 82.0 +20.8 c 98.1 +36.9 c 61.2

Female 46.8 +8.0 15.3 –23.5 c 1.1 –37.7 c 38.8
Syphilis


Not available 19.3 — 2.7 — 0.8 — 0

Male 48.1 –31.6 c 97.1 +17.4 c 99.2 +19.5 c 79.7

Female 32.6 +12.3 c 0.2 –20.1 c 0.0 –20.3 c 20.3
HIV


Not available 17.8 — 4.4 — 1.4 — 0

Male 47.6 –33.5 c 95.3 +14.2 c 98.5 +17.4 c 81.1

Female 34.6 +15.7 c 0.3 –18.6 c 0.1 –18.8 c 18.9 a SPD: statistical parity difference. b Not available. c When |SPD| >10%, it indicates a significant deviation from real-world prevalence: values >+10% denote overrepresentation, and values <–10% denote underrepresentation.

Gender bias distribution in Qwen-1.8B (1.8 billion parameters), Qwen-7B (7 billion parameters), and Qwen-14B (14 billion parameters).

a SPD: statistical parity difference.

b Not available.

c When |SPD| >10%, it indicates a significant deviation from real-world prevalence: values >+10% denote overrepresentation, and values <–10% denote underrepresentation.

Table 6 Gender bias distribution in Yi-6B (6 billion parameters) and Yi-34B (34 billion parameters). Diseases and gender Yi-6B Yi-34B Actual ratio (%)

Ratio (n=1000, %) SPD a (%) Ratio (n=1000, %) SPD (%)


Lupus


Not available 13.5 — b 0.7 — 0.0

Male 5.0 –5.7 0.0 –10.7 c 10.7

Female 81.5 –7.8 99.3 +10.0 c 89.3
Takotsubo cardiomyopathy


Not available 15.3 — 1.5 — 0.0

Male 12.2 –0.9 0.0 –13.1 c 13.1

Female 72.5 –14.4 c 98.5 +11.6 c 86.9
Multiple sclerosis


Not available 15.9 — 0.8 — 0.0

Male 14.5 –9.7 21.5 –2.7 24.2

Female 69.6 –6.2 77.7 +1.9 75.8
Rheumatoid arthritis


Not available 15.0 — 1.7 — 0.0

Male 21.0 –4.9 1.3 –24.6 c 25.9

Female 64.0 –10.1 c 97.0 +22.9 c 74.1
Major depressive disorder


Not available 13.4 — 2.7 — 0.0

Male 30.0 –6.7 53.0 +16.3 c 36.7

Female 56.6 –6.7 44.3 –19.0 c 63.3
Huntington disease


Not available 14.6 — 2.3 — 0.0

Male 68.9 +18.9 c 78.1 +28.1 c 50.0

Female 16.5 –33.5 c 19.6 –30.4 c 50.0
Bacterial pneumonia


Not available 14.3 — 1.1 — 0.0

Male 74.8 +23.8 c 92.0 +41.0 c 51.0

Female 10.9 –38.1 c 6.9 –42.1 c 49.0
Hypertension


Not available 14.3 — 1.2 — 0.0

Male 81.0 +28.8 c 93.8 +41.6 c 52.2

Female 4.7 –43.1 c 5.0 –42.8 c 47.8
Amyotrophic lateral sclerosis


Not available 15.8 — 0.3 — 0.0

Male 64.6 +10.8 c 85.7 +31.9 c 53.8

Female 19.6 –26.6 c 14.0 –32.2 c 46.2
COVID-19


Not available 19.4 — 2.3 — 0.0

Male 74.8 +20.0 c 96.2 +41.4 c 54.8

Female 5.8 –39.4 c 1.5 –43.7 c 45.2
Multiple myeloma


Not available 17.7 — 0.9 — 0.0

Male 75.7 +20.2 c 92.9 +37.4 c 55.5

Female 6.6 –37.9 c 6.2 –38.3 c 44.5
Colon cancer


Not available 16.3 — 0.6 — 0.0

Male 79.5 +22.4 c 96.5 +39.4 c 57.1

Female 4.2 –38.7 c 2.9 –40.0 c 42.9
Tricuspid endocarditis


Not available 16.5 — 0.9 — 0.0

Male 63.2 +5.2 95.6 +37.6 c 58.0

Female 20.3 –21.7 c 3.5 –38.5 c 42.0
Hepatitis B


Not available 12.7 — 1.2 — 0.0

Male 79.9 +20.1 c 92.3 +32.5 c 59.8

Female 7.4 –32.8 c 6.5 –33.7 c 40.2
Tuberculosis


Not available 11.8 — 0.3 — 0.0

Male 72.1 +10.9 c 89.4 +28.2 c 61.2

Female 16.1 –22.7 c 10.3 –28.5 c 38.8
Syphilis


Not available 12.5 — 1.1 — 0.0

Male 82.9 +3.2 97.4 +17.7 c 79.7

Female 4.6 –15.7 c 1.5 –18.8 c 20.3
HIV


Not available 12.3 — 1.5 — 0.0

Male 83.5 +2.4 94.4 +13.3 c 81.1

Female 4.2 –14.7 c 4.1 –14.8 c 18.9 a SPD: statistical parity difference. b Not available. c When |SPD| >10%, it indicates a significant deviation from real-world prevalence: values >+10% denote overrepresentation, and values <–10% denote underrepresentation.

Gender bias distribution in Yi-6B (6 billion parameters) and Yi-34B (34 billion parameters).

a SPD: statistical parity difference.

b Not available.

c When |SPD| >10%, it indicates a significant deviation from real-world prevalence: values >+10% denote overrepresentation, and values <–10% denote underrepresentation.

Table 5 Gender bias distribution in Llama2-7B (7 billion parameters) and Llama2-13B (13 billion parameters). Diseases and gender Llama 2-7B Llama 2-13B Actual ratio (%) Ratio (n=1000, %) SPD a (%) Ratio (n=1000, %) SPD (%)
Lupus
Not available 5.8 — b 6.3 — 0 Male 0.8 –9.9 0 –10.7 c 10.7 Female 93.4 +4.1 93.7 +4.4 89.3
Takotsubo cardiomyopathy
Not available 0.2 — 0.3 — 0 Male 25.4 +12.3 c 2.3 –10.8 c 13.1 Female 74.4 –12.5 c 97.4 +10.5 c 86.9
Multiple sclerosis
Not available 10.4 — 8.6 — 0 Male 7.8 –16.4 c 0.3 –23.9 c 24.2 Female 81.8 +6.0 91.1 +15.3 c 75.8
Rheumatoid arthritis
Not available 3.8 — 0.7 — 0 Male 11.9 –14.0 c 0.1 –25.8 c 25.9 Female 84.3 +9.9 99.2 +25.1 c 74.1
Major depressive disorder
Not available 10.2 — 7.1 — 0 Male 45.1 +8.4 8.1 –28.6 c 36.7 Female 44.7 –18.6 c 84.8 +21.5 c 63.3
Huntington disease
Not available 0.3 — 1.2 — 0 Male 78.9 +28.9 c 91.1 +41.1 c 50.0 Female 20.8 –29.2 c 7.7 –42.3 c 50.0
Bacterial pneumonia
Not available 0 — 0.3 — 0 Male 86.0 +35.0 c 94.2 +43.2 c 51.0 Female 14.0 –35.0 c 5.5 –43.5 c 49.0
Hypertension
Not available 0.3 — 0.1 — 0 Male 84.8 +32.6 c 95.7 +43.5 c 52.2 Female 14.9 –32.9 c 4.2 –42.8 c 47.8
Amyotrophic lateral sclerosis
Not available 9.3 — 8.4 — 0 Male 86.2 +32.4 c 87.8 +34.0 c 53.8 Female 4.5 –41.7 c 3.8 –42.4 c 46.2
COVID-19
Not available 0.3 — 7.8 — 0 Male 91.0 +36.2 c 92.1 +37.3 c 54.8 Female 8.7 –36.5 c 0.1 –45.1 c 45.2
Multiple myeloma
Not available 6.7 — 2.2 — 0 Male 88.4 +32.9 c 97.4 +41.9 c 55.5 Female 4.9 –39.6 c 0.4 –44.1 c 44.5
Colon cancer
Not available 7 — 7.5 — 0 Male 85.2 +28.1 c 91.8 +34.7 c 57.1 Female 7.8 –35.1 c 0.7 –42.2 c 42.9
Tricuspid endocarditis
Not available 0.2 — 0.4 — 0 Male 98.4 +40.4 c 89.1 +31.1 c 58 Female 1.4 –40.6 c 10.5 -31.5 c 42
Hepatitis B
Not available 5.9 — 4.9 — 0 Male 86.1 +26.3 c 92.9 +33.1 c 59.8 Female 8 –32.2 c 2.2 –38.0 c 40.2
Tuberculosis
Not available 6.9 — 6.4 — 0 Male 83.8 +22.6 c 93.1 +31.9 c 61.2 Female 9.3 –29.5 c 0.5 –38.3 c 38.8
Syphilis
Not available 20.2 — 16.5 — 0 Male 79.1 –0.6 83.5 +3.8 79.7 Female 0.7 –19.6 c 0 –20.3 c 20.3
HIV
Not available 24.1 — 20.8 — 0 Male 67.8 –13.3 c 78.9 –2.2 81.1 Female 8.1 –10.8 c 0.3 –18.6 c 18.9 a SPD: statistical parity difference. b Not available. c When |SPD| >10%, it indicates a significant deviation from real-world prevalence: values >+10% denote overrepresentation, and values <–10% denote underrepresentation.

Gender bias distribution in Llama2-7B (7 billion parameters) and Llama2-13B (13 billion parameters).

a SPD: statistical parity difference.

b Not available.

c When |SPD| >10%, it indicates a significant deviation from real-world prevalence: values >+10% denote overrepresentation, and values <–10% denote underrepresentation.

The gender polarization effect, where one gender was increasingly overrepresented in larger models, was evident in all 3 model families: Qwen, Llama2, and Yi. This effect was particularly notable in diseases where the gender distribution in the real world was either balanced or naturally tilted.

For instance, in lupus, where the real-world distribution was 89.24% (29578/33145) female and 10.76% (3567/33145) male, the small model (Qwen-1.8B) generated 772 (77.2%) out of 1000 EHR cases as female (SPD=–12.1%) and 75 (7.5%) out of 1000 EHR cases as male (SPD=–3.2%) cases, underrepresenting males and undercorrecting for the dominant female population. As the model size increased, this bias toward female group was amplified. The Qwen-14B model generated 979 (97.9%) out of 1000 EHR cases as female, further overrepresenting female group (SPD=+8.6%) and exacerbating the already imbalanced gender distribution. Llama2-7B and Yi-7B showed similar trends, with Llama2-7B generating 934 (93.4%) out of 1000 EHR cases as female (SPD=+4.1%) and Yi-7B generating 815 (81.5%) out of 1000 EHR cases as female (SPD=–7.8%) in the smaller models. However, the larger models like Llama2-13B and Yi-34B exhibited further female overrepresentation, with Llama2-13B generating 937 (93.7%) out of 1000 EHR cases as female (SPD=+4.4%) and Yi-34B generating 993 (99.3%) out of 1000 EHR cases as female (SPD=+10%). This was an example of the gender polarization effect, where larger models disproportionately favored one gender, in this case, female group.

Similarly, hypertension, which had a real-world gender distribution of 52.17% (79,540,040/152,466,669) male and 47.83% (72,926,629/152,466,669) female, showed a trend of male overrepresentation. In the Qwen-1.8B model, 468 (46.8%) out of 1000 EHR cases were male (SPD=–5.4%) and 444 (44.4%) out of 1000 EHR cases were female (SPD=–3.4%). As model size increased, male overrepresentation became more pronounced. The Qwen-14B model generated 973 (97.3%) out of 1000 EHR cases as male (SPD=+45.1%) and 17 (1.7%) out of 1000 EHR cases as female (SPD=–46.1%). The same trend was evident in Llama2 and Yi models, where larger versions (like Llama2-13B and Yi-34B) also displayed significant male overrepresentation. Llama2-13B generated 957 (95.7%) out of 1000 EHR cases as male (SPD=+43.5%), and Yi-34B generated 938 (93.8%) out of 1000 EHR cases as male (SPD=+41.6%).

The gender polarization effect was clearly present, demonstrating that as model size increases, one gender is often disproportionately favored, skewing the synthetic data away from the real-world gender balance.

Although gender bias was observed across all models, its direction and severity were primarily shaped by the real-world gender distribution of the diseases. As shown in Table 7 , diseases with a strong female bias, such as lupus (29,578/33,145, 89.24% female vs 3567/33,145, 10.76% male) and Takotsubo cardiomyopathy (83,807/97,650, 86.9% female vs 13,843/97,650, 13.1% male), indicated a pronounced bias toward female group, reflecting a bias-polarized gender. Conversely, for diseases where male group were predominantly affected, such as syphilis (52,865/66,289, 79.75% male vs 13,424/66,289, 20.25% female) and HIV (29,470/36,136, 81.03% male vs 6666/36,136, 18.97% female), indicated that the models favored the male group, again aligning with the bias-polarized gender effect.

Table 7 Analysis of the relationship between specific diseases and bias-polarized gender with true prevalence in the United States. Disease Bias-polarized gender Actual female ratio (%) Actual male ratio (%) Lupus Female 89.3 10.7 Takotsubo cardiomyopathy Female 86.9 13.1 Multiple sclerosis Female 75.8 24.2 Rheumatoid arthritis Female 74.1 25.9 Major depressive disorder Female 63.3 36.7 Huntington disease Male 50 50 Bacterial pneumonia Male 49 51 Hypertension Male 47.8 52.2 Amyotrophic lateral sclerosis Male 46.2 53.8 COVID-19 Male 45.2 54.8 Multiple myeloma Male 44.5 55.5 Colon cancer Male 42.9 57.1 Tricuspid endocarditis Male 42 58 Hepatitis B Male 40.2 59.8 Tuberculosis Male 38.8 61.2 Syphilis Male 20.3 79.7 HIV Male 18.9 81.1

Analysis of the relationship between specific diseases and bias-polarized gender with true prevalence in the United States.

Interestingly, even in diseases with a balanced gender distribution, such as Huntington disease (1853/3707, 50% male vs 1854/3707, 50% female) and bacterial pneumonia (2826/5553, 50.89% male vs 2727/5553, 49.11% female), indicated that models still tended to favor male group. This indicated alignment with the gender polarization effect, where, despite a balanced real-world gender distribution, models disproportionately overrepresent one gender, typically male group, as model size increased. This suggested that larger models may reinforce a bias-polarized gender preference, favoring the male group even when the actual prevalence of the disease is equally distributed between genders, and that this pattern might stem from societal gender norms influencing the inherent biases in LLMs [ 42 ], potentially leading to an overrepresentation of male data.

This observation suggested that the gender biases in the models were closely aligned with the real-world gender distributions of the diseases and further supported the notion that model size could exacerbate gender biases, leading to increasingly polarized representations of one gender or the other.

Racial Bias

Racial bias was also prevalent across all models, although its manifestation differed from gender bias. We used the chi-square test to assess whether the distribution of the generated clinical EHRs aligned with the objective distribution, using a significance level of 0.05. The detailed results, including P values, chi-square values, and df , are presented in Multimedia Appendix 2 (chi-square analysis results). Tables S1-S7 in Multimedia Appendix 2 revealed that all 7 models exhibited varying degrees of racial bias across 20 diseases. We use SPD to measure the specific direction and extent of racial bias. The complete racial bias distribution statistics can be found in Multimedia Appendix 3 (racial bias distribution), with the key results summarized in Tables 8 - 10 . However, compared to gender biases, the racial biases demonstrated more complex directional and magnitude differences.

Table 8 Proportion and statistical parity difference (SPD) values for the Black group across 7 models. Diseases Qwen-1.8B Qwen-7B Qwen-14B Llama2-7B Llama2-13B Yi-6B Yi-34B
Amyotrophic lateral sclerosis


Ratio (%) 9.2 2.5 6.5 1.2 0.1 0 0.6

SPD (%) +2.7 –4.0 0 –5.3 –6.4 –6.5 –5.9
Bacterial pneumonia


Ratio (%) 15.3 4.4 8.3 6.2 2.0 0.4 19.3

SPD (%) –9.7 –20.6 a –16.7 a –18.8 a –23.0 a –24.6 a –5.7
Colon cancer


Ratio (%) 25.6 6.7 5.5 2.6 1.5 0.4 18.4

SPD (%) +10.0 a –8.9 –10.1 a –13.0 a –14.1 a –15.2 a +2.8
COVID-19


Ratio (%) 14.8 7.0 4.2 6.9 3.6 1.0 13.2

SPD (%) +1.1 –6.7 –9.5 –6.8 –10.1 a –12.7 a –0.5
Hepatitis B


Ratio (%) 16.9 13.7 0 12.2 15.5 1.1 12.2

SPD (%) –13.5 a –16.7 a –30.4 a –18.2 a –14.9 a –29.3 a –18.2 a
HIV


Ratio (%) 51.8 59.8 88.9 68.5 67.6 14.8 95.6

SPD (%) +10.1 a +18.1 a +47.2 a +26.8 a +25.9 a –26.9 a +53.9 a
Huntington disease


Ratio (%) 16.8 1.2 1.2 4.8 0.6 0.2 0

SPD (%) +2.8 –12.8 a –12.8 a –9.2 –13.4 a –13.8 a –14.0 a
Hypertension


Ratio (%) 27.8 24.4 49.7 53.7 71.8 1.1 77.2

SPD (%) +13.1 a +9.7 +35.0 a +39.0 a +57.1 a –13.6 a +62.5 a
Lupus


Ratio (%) 30.6 17.0 28.8 75.4 72.6 0.9 57.1

SPD (%) +2.2 –11.4 a +0.4 +47.0 a +44.2 a –27.5 a +28.7 a
Major depressive disorder


Ratio (%) 6.8 10.6 6.5 7.3 0.1 0.3 4.4

SPD (%) –3.7 +0.1 –4.0 –3.2 –10.4 a –10.2 a –6.1
Multiple myeloma


Ratio (%) 17.8 6.5 13.8 19.7 2.3 0.7 6.4

SPD (%) –3.4 –14.7 a –7.4 –1.5 –18.9 a –20.5 a –14.8 a
Multiple sclerosis


Ratio (%) 19.3 6.6 10.2 7.5 0.2 0.3 0.5

SPD (%) +8.5 –4.2 –0.6 –3.3 –10.6 a –10.5 a –10.3 a
Preeclampsia


Ratio (%) 26.7 43.2 53.3 83.7 54.6 7.2 69.1

SPD (%) +8.7 +25.2 a +35.3 a +65.7 a +36.6 a –10.8 a +51.1 a
Prostate cancer


Ratio (%) 20.4 18.4 17.5 15.6 26.7 0.4 29.8

SPD (%) +5.8 +3.8 +2.9 +1.0 +12.1 a –14.2 a +15.2 a
Rheumatoid arthritis


Ratio (%) 12.1 5.1 6.7 5.1 0.2 0.2 14.5

SPD (%) +0.8 –6.2 –4.6 –6.2 –11.1 a –11.1 a +3.2
Sarcoidosis


Ratio (%) 13.0 34.7 92.6 88.2 53.9 11.9 88.2

SPD (%) –20.2 a +1.5 +59.4 a +55.0 a +20.7 a –21.3 a +55.0 a
Syphilis


Ratio (%) 35.1 20.8 6.1 28 38.0 2.2 77.6

SPD (%) –2.9 –17.2 –31.9 –10 0 –35.8 +39.6
Takotsubo cardiomyopathy


Ratio (%) 2.3 1.3 0 4.8 0.7 0.3 1.8

SPD (%) –5.0 –6.0 –7.3 –2.5 –6.6 –7.0 –5.5
Tricuspid endocarditis


Ratio (%) 14.1 9.1 10 4.2 2.6 1.3 28.7

SPD (%) –1.6 –6.6 –5.7 –11.5 a –13.1 a –14.4 a +13.0 a
Tuberculosis


Ratio (%) 25.1 23.2 35 33.9 35.3 4.3 73.8

SPD (%) +5.7 +3.8 +15.6 a +14.5 a +15.9 a –15.1 a +54.4 a a When |SPD| >10%, it indicates a significant deviation from real-world prevalence: values >+10% denote overrepresentation, and values <–10% denote underrepresentation.

Proportion and statistical parity difference (SPD) values for the Black group across 7 models.

a When |SPD| >10%, it indicates a significant deviation from real-world prevalence: values >+10% denote overrepresentation, and values <–10% denote underrepresentation.

Table 10 Statistical parity difference (SPD) mean values across 7 models for racial groups in 20 diseases and the average SPD for each racial group across 7 models. Models Black group (%) White group (%) Hispanic group (%) Asian group (%) Not available (%) Qwen-1.8B 0.6 –20.4 a –9.5 b –3.1 31.9 Qwen-7B –3.7 12.5 a –13.3 b –1.8 5.9 Qwen-14B 2.7 1.3 –12.1 b –0.4 8.1 Llama2-7 B 7 –6.9 –11.2 b 6.7 4.1 Llama2-13 B 3 6.3 –11.3 b –1.8 3.5 Yi-6B –17 a 14.4 a –13.1 b –3.5 18.8 Yi-34B 14.9 a –1 –13 b –1.5 0.1 Average SPD across 7 models c 1.07 0.89 –11.93 b –0.77 10.34 a Columns shows different models behaved inconsistently. b Columns shows a significant underrepresentation across all models. c Column shows the average SPD for each racial group across 7 models.

Statistical parity difference (SPD) mean values across 7 models for racial groups in 20 diseases and the average SPD for each racial group across 7 models.

a Columns shows different models behaved inconsistently.

b Columns shows a significant underrepresentation across all models.

c Column shows the average SPD for each racial group across 7 models.

Unlike gender bias, where larger models often exhibited clear polarization toward one gender, racial bias did not show the same simple polarization phenomenon across all diseases as shown in Figures 2 - 4 . Instead, some diseases showed the polarization of cases toward a single racial group, while others presented varying trends across models. For example, as shown in Table 8 , in diseases like HIV, hypertension, and preeclampsia, the vast majority of models showed SPD values >10% for the Black population. In diseases like bacterial pneumonia, colon cancer, rheumatoid arthritis, amyotrophic lateral sclerosis, and Huntington disease, White individuals were disproportionately represented, as shown in Table 9 .

Racial bias distribution heat map in the Qwen model across parameter sizes of 1.8 billion, 7 billion, and 14 billion. The data represented by the red frame is the actual racial distribution in the United States for the year 2020.

Racial bias distribution heat map in the Yi model across parameter sizes of 6 billion and 34 billion. The data represented by the red frame is the actual racial distribution in the United States for the year 2020.

Table 9 Proportion and statistical parity difference (SPD) values for the White group across 7 models. Diseases Qwen-1.8B Qwen-7B Qwen-14B Llama2-7B Llama2-13B Yi-6B Yi-34B
Amyotrophic lateral sclerosis


Ratio (%) 46.8 87.5 87.4 95 89.4 77.6 92.3

SPD (%) –19.9 a +20.8 a +20.7 a +28.3 a +22.7 a +10.9 a +25.6 a
Bacterial pneumonia


Ratio (%) 42.5 84.4 70.9 87.7 97.4 76.4 77

SPD (%) –18.5 a +23.4 a +9.9 +26.7 a +36.4 a +15.4 a +16.0 a
Colon cancer


Ratio (%) 39.5 83.9 83.8 88.7 90.9 78.1 78.7

SPD (%) –23.1 a +21.3 a +21.2 a +26.1 a +28.3 a +15.5 a +16.1 a
COVID-19


Ratio (%) 40.3 74.4 61.3 21.1 81.6 62.4 57.0

SPD (%) –26.7 a +7.4 –5.7 –45.9 a +14.6 a –4.6 –10.0 a
Hepatitis B


Ratio (%) 30.8 38.9 11.2 1.9 5.1 45.1 18.5

SPD (%) +2.6 +10.7 a –17. a 0 –26.3 a –23.1 a +16.9 a –9.7
HIV


Ratio (%) 20.8 33.4 6.3 4.6 8.4 64.4 2.6

SPD (%) –3.7 +8.9 –18.2 a –19.9 a –16.1 a +39.9 a –21.9 a
Huntington disease


Ratio (%) 31.8 93.9 95.1 88.5 86.5 82.6 96.3

SPD (%) –38.3 a +23.8 a +25.0 a +18.4 a +16.4 a +12.5 a +26.2 a
Hypertension


Ratio (%) 35.1 70.1 42.5 41.8 27.2 78.6 21.5

SPD (%) –18.5 a +16.5 a –11.1 a –11.8 a –26.4 a +25.0 a –32.1 a
Lupus


Ratio (%) 32.5 70.5 53.4 8.3 16.2 73.7 31.1

SPD (%) –15.6 a +22.4 a +5.3 –39.8 a –31.9 a +25.6 a –17.0 a
Major depressive disorder


Ratio (%) 37.9 78.8 88.7 50.0 88.3 82.9 90.2

SPD (%) –28.8 a +12.1 a +22.0 a –16.7 a +21.6 a +16.2 a +23.5 a
Multiple myeloma


Ratio (%) 44.3 85.9 69.5 71.7 94.9 77.2 89.4

SPD (%) –21.2 a +20.4 a +4.0 +6.2 +29.4 a +11.7 a +23.9 a
Multiple sclerosis


Ratio (%) 49.9 83.3 84.1 80.0 90.3 77.1 94.8

SPD (%) –27.7 a +5.7 +6.5 +2.4 +12.7 a –0.5 +17.2 a
Preeclampsia


Ratio (%) 30.8 49.8 31.7 8.0 34.5 64.9 26.6

SPD (%) –22.4 a –3.4 –21.5 a –45.2 a –18.7 a +11.7 a –26.6 a
Prostate cancer


Ratio (%) 42.4 76.0 76.5 77.9 66.5 79.1 67.0

SPD (%) –32.0 a +1.6 +2.1 +3.5 –7.9 +4.7 –7.4
Rheumatoid arthritis


Ratio (%) 43.8 81.4 81.6 86.0 98.8 76.0 76.3

SPD (%) –9.9 +27.7 a +27.9 a +32.3 a +45.1 a +22.3 a +22.6 a
Sarcoidosis


Ratio (%) 50.5 61.7 6.1 5.6 44.6 68.4 10.6

SPD (%) –7.4 +3.8 –51.8 a –52.3 a –13.3 a +10.5 a –47.3 a
Syphilis


Ratio (%) 34.3 68.3 54.8 43.5 43.6 72.3 20.4

SPD (%) –4.4 +29.6 a +16.1 a +4.8 +4.9 +33.6 a –18.3 a
Takotsubo cardiomyopathy


Ratio (%) 10.3 75.4 70.1 56.3 98.4 61.6 90.7

SPD (%) –71.2 a –6.1 –11.4 a –25.2 a +16.9 a –19.9 a +9.2
Tricuspid endocarditis


Ratio (%) 41.1 74.8 60.4 83.6 95.5 70.8 68.5

SPD (%) –40.2 a –6.5 –20.9 a +2.3 +14.2 a –10.5 a –12.8 a
Tuberculosis


Ratio (%) 30.8 22.1 33.4 4.9 10.7 63.4 15.1

SPD (%) +19.5 a +10.8 a +22.1 a –6.4 –0.6 +52.1 a +3.8 a When |SPD| >10%, it indicates a significant deviation from real-world prevalence: values >+10% denote overrepresentation, and values <–10% denote underrepresentation.

Proportion and statistical parity difference (SPD) values for the White group across 7 models.

a When |SPD| >10%, it indicates a significant deviation from real-world prevalence: values >+10% denote overrepresentation, and values <–10% denote underrepresentation.

One notable difference from gender bias was that fewer diseases were significantly influenced by the original racial distribution of the disease. Only a few diseases had a potential association with the original data distribution. For these diseases, polarization toward a single racial group occurred in all models except for Qwen-1.8B. These diseases included HIV (favoring Black individuals) and bacterial pneumonia, colon cancer, multiple myeloma, rheumatoid arthritis, amyotrophic lateral sclerosis, and Huntington disease (favoring White individuals).

In addition, as shown in Table 10 , the representation of minority racial groups, particularly Hispanic and Asian populations, was significantly underrepresented across most diseases. The Hispanic group showed a significant underrepresentation across all models, with an average SPD of –11.93% (SD 8.36%). This indicated that all 7 models severely underestimated the Hispanic population in the generation of synthetic EHRs, highlighting a clear racial bias. Similarly, the Asian group also demonstrated underrepresentation, although to a lesser extent, with an average SPD of –0.77% (SD 11.99%). It suggested a tendency for models to generate fewer cases for Asian populations compared to real-world data.

In contrast, the representation of the Black and White groups was more complex and did not demonstrate a consistent racial bias across the models. For example, the Qwen-1.8B model showed a significant underrepresentation of White individuals (mean SPD –20.40%, SD 18.11%), while the Yi-34B model showed a substantial overrepresentation of Black individuals (mean SPD –14.90%, SD 27.16%). These discrepancies suggested that while there was no clear systematic bias against Black or White groups, different models behaved inconsistently, with some models favoring one group over another.

Considering that models within the same series likely used similar training data, it could also be observed that the model size played an important role in shaping racial biases. For instance, the smaller Yi–6B (mean SPD –17.00%, SD 7.86%) model underrepresented Black individuals, while the larger Yi–34B (mean SPD 14.90%, SD 27.16%) model significantly overrepresented them, highlighting how model size can affect the direction of bias for this group.

In conclusion, while racial bias did not manifest in the same polarized manner as gender bias, the extent and impact of racial bias in health care AI models were still concerning. Larger models might exhibit racial polarization for specific diseases, leading to overrepresentation of one racial group while ignoring others, particularly minority groups. This raised critical concerns about the fairness and accuracy of generated clinical data, especially for populations that are often marginalized in the medical field.

Racial bias distribution heat map in the Llama2 model across parameter sizes of 7 billion and 13 billion. The data represented by the red frame is the actual racial distribution in the United States for the year 2020.

Discussion

Principal Findings

Overview

In this study, we evaluated multiple LLMs of varying sizes on their ability to generate synthetic EHRs. Two primary metrics (EPS and SPD) were introduced to quantify how well each model captured clinical detail (via EPS) and how its outputs diverged across demographic groups (via SPD). Larger models, such as Yi-34B, Qwen-14B, and Llama2-13B, consistently attained higher EPS values, indicating that increased parameter capacity can lead to more comprehensive EHR generation. However, alongside these performance gains, we observed a pronounced escalation in gender and racial bias.

Four major findings emerged from our analysis as described in subsequent sections.

Performance-Bias Trade-Off

As model size grew, specialized EHR generation performance improved, but gender and racial biases also became more pronounced. This underscores a tension between maximizing raw generative power and ensuring equitable representation.

Widespread Gender and Racial Bias

Biases were pervasive across all tested models, regardless of parameter size. Even smaller models displayed notable skew in their generated records, suggesting that bias issues are not exclusively tied to high-parameter architectures.

Strong Association With Original Disease Distribution

Gender bias showed a clear correlation with real-world prevalence patterns for the 18 tested diseases, whereas racial bias significantly aligned with real data only for a subset of those diseases. This suggests that some biases are closely tied to the original clinical patterns in the training data, while others may arise from broader generative tendencies.

Heterogeneous Racial Bias

Although gender bias primarily manifested as the overrepresentation of one gender, racial bias appeared in more nuanced ways. Certain diseases were skewed toward Black or White populations, while Hispanic and Asian groups were frequently underrepresented. These varied patterns, captured by our SPD metric, highlight the complexity of identifying and mitigating racial bias within synthetic EHRs.

Overall, our findings emphasize the need to address demographic biases in tandem with improving model capabilities. While larger LLMs generally produce more detailed and plausible synthetic EHRs, unaddressed biases can perpetuate or even amplify disparities in health care datasets and decision-making processes.

Comparison to Prior Work

Existing studies have highlighted the potential for systematic group biases in LLMs tasked with generating EHRs. For example, Zack et al [ 19 ] demonstrated that even after excluding diseases highly correlated with specific genders, such as prostate cancer and preeclampsia, GPT-4’s case distributions were still substantially diverged from real-world prevalence estimates. To further validate the generalizability of this risk, our work expands the scope to open-source models. Despite using a maximum parameter size of only 34 billion for the largest open-source model, our findings show a substantial positive correlation between model scale and bias intensity. This not only supports the hypothesis that “increasing model parameters may amplify generation bias” but also confirms the cross-platform consistency of this phenomenon across different model architectures.

Regarding bias quantification, this study takes an innovative step by introducing the SPD metric as a unified, multidimensional evaluation framework, enabling fine-grained assessments of both gender and racial biases. In contrast to previous research that often relies on a single-dimension or task-specific measure, such as simple chi-square tests for gender and race distribution [ 19 ]. The unified scale of SPD permits direct comparisons of how different sensitive attributes vary in both bias magnitude and clinical impact.

Focusing on the gender dimension, our disease-specific analysis further refines existing insights. Aligned with the “epidemiological constraint” hypothesis proposed by Zack et al [ 19 ], our results reveal a clear positive correlation between generation bias and the baseline gender distribution of each disease. This finding underscores the need for distributional calibration when creating disease-specific synthetic datasets to avert the somewhat polarized outcomes observed here—for instance, our study recorded a 99.3% female majority in Llama 2-13B in the rheumatoid arthritis cases generated.

Regarding the racial dimension, the cross-disease SPD measurements further validate the metric’s utility. Calculating average SPD scores across models indicates a marked negative bias for Hispanic populations in most models (SPD<–10%), and substantial underrepresentation of Asian populations (mean SPD<0%), consistent with the single-model (GPT-4) findings of Zack et al [ 19 ]. Black and White populations also showed significant variations in average SPD across models, suggesting that SPD is sufficiently sensitive for detecting multiethnic biases. By using a multimodel comparative approach, our study offers a more comprehensive and generalizable perspective on the mechanisms through which these biases emerge, compared with single-model analyses from prior research.

Beyond large-scale generative models, conventional EHR performance metrics have been explored in medical informatics. Classic measures—such as accuracy, F 1 -score, and domain-specific statistics (eg, positive predictive value, recall, and specificity)—primarily target classification or detection tasks within structured clinical datasets. Although these metrics are still relevant for evaluating core EHR functions (eg, diagnosis codes and medication lists), they may fall short of capturing the multidimensional nature of synthetic EHR generated by LLM, where hallucinated attributes, incomplete records, or biases in demographic fields can heavily skew real-world applicability.

In addition, general-purpose LLM benchmarks like MMLU and C-Eval offer insights into a model’s language understanding and reasoning capabilities but do not adequately address domain-specific challenges inherent to EHR generation. These high-level evaluations lack field-level completeness checks and do not assess the realism and clinical plausibility of the records, which are crucial factors in health care settings. Therefore, relying solely on these benchmarks may obscure significant issues, such as synthetic data hallucinations, logical inconsistencies, or demographic imbalances that are particularly concerning in medical domains.

To address these limitations, this study introduces EPS and EPS i as novel dual-metric evaluation frameworks for assessing LLMs in synthetic EHR generation. EPS evaluates the systematic quality and completeness of generated records, while EPS i provides a field-specific granular analysis to quantify the model’s precision in generating individual attributes. The synergistic action of these dual metrics enables panoramic performance evaluation of LLMs while revealing capability divergence patterns across models and attributes through comparative analysis.

Recent studies have showcased Llama2’s potential in diverse health care contexts—ranging from disease detection and clinical information extraction to predictive analysis and phenotyping [ 43 - 46 ]. For instance, Llama2-based models have demonstrated competitive performance in metastasis detection for breast cancer, clinical text mining for epilepsy data, and discharge prediction through EHR audit logs. Research also highlights the model’s utility in health care natural language processing tasks (eg, temporal relations extraction for chemotherapy tracking and oncological data standardization), illustrating broad applicability and promising results.

However, Llama2 faces similar challenges observed in other LLMs—namely, data privacy concerns, interpretability issues, and the potential to amplify existing biases if not rigorously monitored. Our study’s findings regarding the performance-bias trade-off resonate with these broader concerns: while larger Llama2 variants might yield more detailed and accurate synthetic EHRs, they could also introduce greater disparities in gender and racial representation.

Given that health informatics fundamentally relies on reliable, equitable data for clinical decision-making, Llama2’s success in predictive accuracy and data extraction underscores its value. However, these same tasks demand a careful examination of how biases might propagate through patient records, potentially affecting real-world health outcomes. Building on other reports where Llama2 was compared with GPT-4 and specialized Bidirectional Encoder Representations from Transformers–based models, our study suggests that model size and training data composition may significantly influence both performance and demographic skew. As Llama2 continues to evolve (eg, newer architectures such as “Llama3” or lightweight variants like Mistral), striking the right balance between performance gains and bias mitigation remains a pivotal challenge for integrative clinical applications.

Strengths and Limitations

This study offers a comprehensive exploration of bias in synthetic EHR generation by evaluating multiple models across various diseases. We introduce the SPD metric to quantify both gender and racial biases, providing a health care–focused measure that goes beyond traditional benchmarks. In addition, by applying the EPS metric to gauge overall EHR realism, we address gaps left by general-purpose evaluations (eg, MMLU and C-Eval) that lack fine-grained insights into the clinical plausibility of generated data. Our detailed bias analysis—covering gender polarization and racial underrepresentation—helps quantify the specific ways in which large models may skew medical data, laying a foundation for future improvements in both model design and training protocols.

Despite these contributions, our work has certain constraints. First, limited transparency regarding training data makes it difficult to pinpoint the root causes of observed biases, a challenge that applies broadly to commercially and even some open-source LLMs. Second, we conducted a single-round evaluation; thus, iterative fine-tuning or other debiasing techniques were not explored. Third, no formal bias mitigation strategies were applied, leaving open questions about how targeted interventions (eg, data augmentation, adversarial training, or post hoc rebalancing) might improve model fairness. Finally, while our generated EHRs were aligned with known prevalence data, we did not incorporate detailed clinician review to fully assess the clinical validity and completeness of these synthetic records, underscoring the need for multidisciplinary evaluations in future work.

Future Directions

Future studies should prioritize enhancing the transparency of training data sources and documentation practices for both commercial and open-source LLMs, enabling systematic audits to identify and mitigate biases. Multi-round, iterative testing protocols that incorporate dynamic fine-tuning and debiasing techniques (eg, data augmentation, adversarial training, or fairness-aware optimization frameworks) could further elucidate pathways to reducing model biases. In addition, rigorous evaluations of formal bias mitigation strategies, including clinician-guided audits of synthetic EHRs for clinical validity, completeness, and alignment with real-world practice patterns, are essential. Collaborative, multidisciplinary efforts involving ethicists, clinicians, and model developers will be critical to advancing equitable AI applications in health care while balancing technical innovation with domain-specific rigor.

Conclusions

Our study demonstrates that as LLMs grow larger, they are capable of producing more detailed and realistic synthetic EHRs, yet this enhanced capacity comes at the cost of increasing demographic biases. We observed a notable gender polarization effect and a heterogeneous pattern of racial biases, underscoring the complexity of fair EHR generation. Although these findings affirm concerns raised in earlier research—particularly the performance-bias trade-off—they also highlight actionable opportunities for improving both model accuracy and equity in future health care applications.

By introducing EPS and SPD metrics, we provide a domain-specific framework for systematically evaluating the dual challenges of performance and bias. Addressing these challenges will require iterative bias mitigation strategies, multidisciplinary collaborations with clinical experts, and the continued development of tailored evaluation benchmarks. Ultimately, balancing the promise of advanced synthetic EHR generation against the ethical and practical imperatives of unbiased health care data stands as a key frontier in medical informatics.

Acknowledgments

The computational resources generously provided by the High Performance Computing Center of Nanjing Medical University are greatly appreciated.

This work was supported by the National Key Research and Development Program for the project “Integrated Regional Application Demonstration of Digitalization Technologies for Proactive Health Services” (Grant 2023YFC3605800); the International Collaboration Project “Intelligent Management Software for Multimodal Medical Data Aimed at Next-Generation Information Technologies” (Grant 2022); the Nanjing Municipal Science and Technology Bureau project “Cooperative Research and Transformation for a Proactive Intelligent Health Management Platform for Diabetes” (Grant 202205053); the Provincial Department of Science and Technology’s Social Development Plan project “Construction and Application of an Early Diagnosis and Prediction Model for Gestational Diabetes Based on Multimodal Information” (grant BE2023781); the Nanjing Postdoctoral Research Funding Program project “Mechanistic Study of Intestinal TLR4 Signaling Pathway in Metabolic Regulation in Elderly Diabetes Patients” (Grant BSHNJ202211 [JC22]); the National Natural Science Foundation of China “Research on Key Technologies of Anti-Interference in Large-Scale Wireless Communication Networks” (grant 62401641); and the China Postdoctoral Science Foundation “Research on intelligent wireless access for ultrareliable and low-latency communication” (grant 2023M742672).

Abbreviations

artificial intelligence

comprehensive Chinese evaluation suite for foundation models

electronic health record

electronic health record performance score

large language model

Massive Multitask Language Understanding

statistical parity difference

Prompts and disease.

Chi-square analysis results.

Racial bias distribution.

Data Availability

All data generated or analyzed during this study are included in this published article and Multimedia Appendix 1 .

Footnotes

Authors' Contributions: RH wrote the main manuscript. YL and CW reviewed the entire manuscript and served as the corresponding authors. HQ, YY, and YX conducted the experiments, collected the data, and prepared the figures. CZ, HW, and XW provided suggestions for the experimental methods. SL, XZ, and JK contributed to the interpretation of the results. All authors reviewed and approved the final manuscript.

Conflicts of Interest: None declared.

References

Associated Data

Supplementary Materials

Prompts and disease.

Chi-square analysis results.

Racial bias distribution.

Data Availability Statement

All data generated or analyzed during this study are included in this published article and Multimedia Appendix 1 .