The study flow is illustrated in Figure 1. A total of 19,983 clinical notes were randomly collected from multiple large-scale tertiary hospitals, encompassing records from the years 2007 through 2023. These records covered a variety of document types, including ward round notes, initial clinical summaries, transfusion logs, handover reports, and routine progress documentation. A rigorous data preprocessing pipeline was implemented to protect patient confidentiality and improve data integrity. During the deidentification phase, all personally identifiable information relating to patients and physicians was uniformly replaced with generic terms, such as “Patient” and “Physician.” Similarly, health care institution names were anonymized.

To ensure data quality, a multitiered filtration protocol was implemented: (1) removal of records manifesting abnormal, nonsensical, or noncompliant content with established medical documentation standards; (2) exclusion of entries shorter than 100 characters; and (3) elimination of excessively long records exceeding 3000 characters that contained abundant irrelevant content or extraneous tags. Following this systematic preprocessing, 18,450 records meeting the quality criteria were initially retained. To prevent potential patient-level data leakage, we further verified all records using a double-deidentified patient master index. This verification showed that the 18,450 notes corresponded to 18,432 unique patients, with only 14 patients contributing multiple notes. After retaining 1 note per patient and removing 18 duplicate-patient records, all of which were ADE-negative, 18,432 patient-level clinical records were retained.

The curated dataset was randomly partitioned into 2 distinct subsets with a 3:7 ratio: (1) a standard reference dataset comprising 5535 entries, designated as the gold standard to rigorously validate the efficacy of ADE identification methodologies; (2) an ADE case knowledge base consisting of 12,897 entries, intended for providing domain-specific knowledge. Both datasets were processed using a standardized annotation protocol. To eliminate potential automation bias, initial ADE preannotations were generated using Qwen-turbo (model version updated April 28, 2025; temperature=1.0, top_P=.9, penalty_score=1.05), a model explicitly distinct from the 3 evaluated LLMs. These preannotations were subsequently reviewed and revised by clinical pharmacists with 2‐10 years of pharmacovigilance experience. All annotators were fully blinded to the preannotation model and the evaluation design, receiving only raw clinical notes and machine-generated labels. Interrater reliability was assessed using Cohen κ based on 5535 clinical notes independently reviewed by 2 annotators, yielding a high agreement (Cohen κ=0.86). Any disagreements between annotators were resolved through consensus discussion, thereby confirming the high consistency and robustness of the gold-standard reference dataset.

During the annotation and review process, we observed a high degree of homogeneity in both clinical notes and the ADE annotations within the positive samples. Such redundancy has the potential to introduce bias in subsequent analyses, particularly as cases sharing identical ADE categories may collectively succeed or fail in recognition, thereby compromising the robustness of evaluative conclusions. To address this issue, we implemented a strict deduplication and balancing procedure to derive the final datasets. First, within the initial standard reference dataset of 5535 entries (comprising 2789 positive and 2746 negative instances), we deduplicated the positive reference set by limiting each drug entity to no more than 3 appearances. Deduplication was performed strictly according to the annotated drug entity name, without collapsing specific drugs into a generic category. Following deduplication, the positive reference set was refined to 1228 records. Correspondingly, an equivalent sample of 1282 instances was randomly selected from the 2746 negative cases, resulting in a balanced, final standard reference dataset of 2510 records. Subsequently, a similar deduplication and balancing process was applied to the initial dataset comprising 12,897 clinical records. This resulted in the creation of a final ADE knowledge base comprising 5144 records (2273 positive cases and 2871 negative cases).

To standardize the annotation process and accurately capture the subtle characteristics of ADE descriptions in clinical notes, a JSON schema was adopted for ADE annotation and recognition tasks (Figure S1 in Multimedia Appendix 1). Specifically, the JSON schema encapsulates a structured array wherein each ADE incident is described as an object comprising three pivotal fields: (1) sentence: the original textual fragment characterizing the ADE event; (2) drugs: an array listing the implicated pharmaceutical agents, accommodating polypharmacy interactions; and (3) reactions: an array detailing the corresponding ADE entities, supporting multiple concurrent reactions. If no ADE is identified within a record, an empty array “[]” is returned.

Using a JSON-based model for ADE information has several distinct advantages: (1) structured storage facilitates efficient data organization through standardized fields, thereby providing a solid foundation for downstream tasks such as causal inference and association analysis; (2) stringent field constraints (eg, requiring nonempty reaction arrays) enforce rigorous accuracy requirements for ADE entity recognition; and (3) compatibility with LLM input-output paradigms enhances the efficacy of generative models in extracting and completing ADE-related information.

The study used 3 state-of-the-art LLMs: DeepSeek-V3 (DeepSeek), ERNIE 3.5-8K (Baidu), and GPT-4o (OpenAI). Specifically, DeepSeek-V3 offers cutting-edge open-source performance with privacy-preserving local deployment; ERNIE 3.5-8K provides a commercial baseline optimized for Chinese clinical natural language processing; and GPT-4o serves as a top-tier, general-purpose international benchmark. Together, they represent a broad spectrum of open-source versus closed-source, Chinese-specialized versus general-purpose, and local versus cloud paradigms. The key to leveraging these LLMs was prompt engineering. To systematically assess the impact of prompt design and RAG strategy on ADE recognition performance, a progressive optimization framework was implemented: (1) nonaugmented generation (NAG)—a zero-shot prompting baseline strategy supplying the LLM with ADE identification instructions without any supplementary knowledge, thereby delineating the model’s innate capabilities; (2) static-augmented generation (SAG)—building upon NAG, this zero-shot approach incorporates static clinical knowledge descriptions of ADE concepts to formulate semantically constrained recognition prompts, evaluating the influence of conceptual knowledge injection; and (3) RAG—further advancing SAG into a dynamic few-shot prompting strategy with a dynamic retrieval mechanism that fetches analogous cases from the ADE knowledge base in real time, yielding context-aware prompts optimized for the complexity of clinical notes (Figure 2 and Figure S3 in Multimedia Appendix 1).

According to the retrieval schema of the RAG framework proposed in this paper, each case is structured as a key-value pair: the key is the original ADE text snippet, and the value is a standardized JSON string. Keys were vectorized using the shaw/dmeta-embedding-zh model (vector dimension=768) and indexed in a Milvus (v2.5.2; Zilliz) database, with corresponding JSONs stored as structured metadata. During dynamic retrieval, input clinical notes are split into sentences using punctuation marks and processed via a sliding window (size=3, step=2). These contextual segments are vectorized to retrieve the top 5 most similar ADE entries from the Milvus database via cosine similarity. The retrieval module achieved a Recall@5 of 0.908 and a mean reciprocal rank of 0.874, empirically verifying its ability to accurately fetch relevant ADE cases and reduce model hallucinations.

This prompt architecture follows a logical progression: from baseline capability verification to knowledge-enhanced refinement and adaptive contextualization, establishing a rigorous and reusable experimental paradigm for quantitatively evaluating knowledge augmentation strategies. All LLM inference, prompting, and ADE extraction tasks were conducted natively in Chinese. The English prompt text and translations shown in figures and tables are provided solely for reader comprehension and were not used as model inputs. The NAG instruction comprises 3 sections: “System Instruction,” “Output Requirement,” and “Clinical Course Record.” The “System Instruction” guides ADE identification, while the “Output Requirement” specifies adherence to the ADE JSON schema. The “Clinical Course Record” represents the data to be analyzed, with “record” acting as a placeholder for later use. The SAG instruction mirrors NAG but elaborates on ADE criteria, including temporal associations, drug exposure history, clinical manifestations, and causality terms. Exclusion criteria are also defined, including mentions of drugs without adverse reactions, non–drug-related reactions, prophylactic medications, vigilant use, postdiscontinuation reactions, and prior ADEs. The RAG instruction extends SAG by adding a “Reference Knowledge” section. Specifically, the retrieved top 5 ADE reference cases are explicitly injected into this section of the structured RAG prompt template (Figure 2), providing contextual domain knowledge to enhance the LLMs’ accurate and interpretable ADE identification.

Our entire experimental pipeline was implemented natively in Java (JDK 1.8; Oracle Corporation), using the HttpClient library to establish seamless communication with cloud-based model services. Although DeepSeek-V3 permits privacy-preserving local deployment, the present experiments used its cloud API version, together with the cloud APIs of ERNIE 3.5-8K and GPT-4o, to ensure consistent batch inference and experimental efficiency. All API calls used deidentified Chinese clinical-note inputs and Chinese prompt templates. To guarantee complete study reproducibility, the exact versions and configurations of the invoked models are specified as follows: DeepSeek-V3 (updated March 24, 2025; temperature=0.3, top_P=.70, penalty_score=1.0), ERNIE 3.5-8K (updated December 22, 2024; temperature=0.3, top_P=.70, penalty_score=1.0), and GPT-4o (updated November 20, 2024; temperature=0.3, top_P=.70, frequency_penalty=0, presence_penalty=0).

The experimental outcomes were evaluated using precision, recall, the F₁-score, specificity, the F₂-score, G-mean, and the Cohen κ metric. Within the reference dataset, clinical notes containing ADEs were designated as positive instances, whereas those devoid of ADEs were classified as negative instances. The metrics were computed as follows:

TP denotes the count of true positives where the model correctly identifies an instance as positive; FP represents false positives where the model identifies an instance as positive but the reference set labels it as negative; FN signifies false negatives where the model fails to identify an instance marked positive in the reference set. P_o represents the actual observed accuracy and P_e represents the chance-expected accuracy. All reported metrics were computed at the document level using a clinical note, rather than an individual ADE entity, as the unit of evaluation.

It is important to recognize that identifying ADEs is a complex task. For any given sample in the standard reference set, a more nuanced determination of recognition accuracy is required. To this end, we have defined 3 levels of recognition precision: L1, L2, and L3 (Table 1). The L1, L2, and L3 rules were used to determine TP versus FN status only among ADE-positive notes. L1 (exact match) strictly requires complete alignment of the predicted sentence, drug array, and adverse reaction array with the gold standard. Under L1, partial matches (eg, correct drug but incomplete reactions) were rejected as TP and counted as FN at the document level. Conversely, L2 (sentence-level match) counted a note as TP when the predicted and gold-standard sentences aligned, even if the drug or reaction entities were incomplete or partially discrepant. L3 (overlap match) further relaxes this criterion, counting a TP if at least one drug or adverse reaction entity overlaps, regardless of overall array completeness. Figure S4 in Multimedia Appendix 1 provides illustrative examples to further elucidate these definitions. Paired bootstrap tests were applied to the 1000 bootstrap resamples to evaluate the statistical significance of performance differences across models, prompts, and matching levels. P values were adjusted using the Benjamini-Hochberg false discovery rate (FDR) procedure, and statistical significance was defined as FDR P<.05.

To evaluate model robustness under real-world clinical conditions, we additionally assessed DeepSeek-V3 using uncurated clinical progress notes sampled from the original pool of 19,983 raw records. Prior to calculating the metrics, the remaining records that were excluded from the curated dataset were also annotated with supplementary ADE labels using the same JSON schema and a consistent manual review protocol. In total, 1000 resampling iterations were performed, with 1000 notes randomly selected in each iteration and evaluated under the NAG, SAG, and RAG prompting strategies. The average ADE-positive prevalence was approximately 4.6%. Given this imbalance, specificity, F₂-score, G-mean, and Cohen κ were calculated in addition to precision, recall, and F₁-score to assess false-positive control, recall-oriented detection, overall discrimination, and agreement with the gold standard. To further assess performance on challenging clinical language, we constructed a targeted 100-note subset containing ADE negation, abbreviations, temporal and causal contexts, and evaluated it using DeepSeek-V3 with RAG. All reported metrics, including precision, recall, F₁-score, specificity, F₂-score, G-mean, and Cohen κ, were computed at the document level using a clinical note.

Statistical analyses were performed in Python (version 3.10; Python Software Foundation) using NumPy, pandas, SciPy, and scikit-learn packages for bootstrap resampling, percentile-based 95% CIs, paired bootstrap tests, Cohen κ, and classification metrics.

This retrospective study was approved by the Ethics Review Board of Xiangya Hospital of Central South University (2025091288). The requirement for informed consent was waived because this study involved a secondary analysis of existing clinical notes. All data used were anonymized to ensure participant privacy. No personally identifiable information was included in the study, nor was any such information disclosed to any LLMs.

After a thorough manual review and annotation process, an initial analysis was conducted prior to data deduplication to understand the distribution characteristics of the data. A frequency analysis of drug entities and ADEs was conducted within the positive reference set. This dataset encompassed 748 unique drug entities, each of which appeared on average 3 times, with 28 entities exceeding 10 occurrences. Notably, chemotherapeutic agents represented the most common drug entity category, appearing 775 times (Figure 3A). This count included only instances in which the generic term itself was recorded in the annotation, whereas specific chemotherapy drugs were retained as distinct entities rather than aggregated under this term. Concurrently, 2668 unique ADE types were annotated, 22 of which appeared 10 or more times (Figure 3B). As detailed in the “Methods” section, to mitigate potential bias caused by this high frequency of redundant cases, both datasets underwent a rigorous deduplication and balancing process. This procedure ultimately yielded a refined standard reference dataset of 2510 records and an ADE knowledge base of 5144 records, which were subsequently used for all downstream model evaluations. A baseline RAG experiment using the full nondeduplicated knowledge base of 12,897 records showed slightly lower performance than the deduplicated knowledge base (Table S1 in Multimedia Appendix 1), supporting that limiting each drug entity to no more than 3 appearances improved retrieval diversity and reduced retrieval homogenization.

This work marks the first public release of a Chinese clinical note-based ADE research dataset, which is intended to be an open resource for the scientific community. The dataset is accessible at a GitHub repository [23]. Each record is characterized by 3 attributes: ID (a unique dataset identifier), Content (the clinical progress note text), and ADEs (annotations conforming to the ADE JSON schema specification).

To ensure the reliability of the experimental results and the stability of the methodology, we used 1000 bootstrap resamples. In each round, 1000 samples were randomly selected with replacement from the standard reference dataset of 2510 samples, and the ADE identification method was evaluated on the resampled subset. The statistical results are summarized in Table 2 as bootstrap means with percentile-based 95% CI. Under the NAG prompt, all LLMs demonstrated strong ADE recognition performance, with both DeepSeek-V3 and ERNIE 3.5-8K achieving an overall F₁-score of over 0.9. At the L2 and L3 matching levels, adoption of the SAG prompt improved the recognition performance of all models compared to the NAG prompt, indicating that optimizing prompt design and incorporating domain-specific knowledge generally enhances the model’s ability to recognize ADEs. Notably, applying the RAG prompt, which integrates retrieval information from a large-scale ADE case knowledge base, resulted in even greater performance improvements for all models compared to the SAG prompt. DeepSeek-V3 achieved the most outstanding recognition performance with an impressive overall F₁-score of 0.9638 (95% CI 0.9541‐0.9727).

Initially, an analysis of the ADE recognition instruction framework at the L3 level is presented. As illustrated in Figure 4, the comparison of recognition performance across various prompts at the L3 level is demonstrated for different LLMs. Compared with NAG, the SAG prompt increased the F₁-score of DeepSeek-V3 from 0.9022 to 0.9441 (FDR P=.003) and GPT-4o from 0.7698 to 0.8173 (FDR P=.006). For ERNIE 3.5-8K, SAG produced only a modest increase from 0.9453 to 0.9571, which did not reach statistical significance in the paired bootstrap test (FDR P=.09). Building on the SAG prompt, the RAG prompt further augmented the comprehension capacity of the LLMs, yielding remarkable performance. DeepSeek-V3 and GPT-4o showed improvements of 1.97% (FDR P=.003) and 13.06% (FDR P=.003), respectively. These findings affirm the crucial role of augmenting domain knowledge in improving ADE recognition performance within mainstream LLMs and validate the effectiveness of the ADE case knowledge base designed in this study. The knowledge retrieval enhancement method was successful in significantly boosting ADE recognition performance. Ultimately, under the RAG prompt, mainstream LLMs exhibited remarkable improvement in ADE recognition, with average F₁-scores exceeding 0.95.

Regarding precision (Figure 4), all models achieved a high level of precision under the NAG prompt, with the lowest precision reaching 95.59%. Minor fluctuations were subsequently observed under the SAG and RAG prompts, yet an overall upward trend in performance was evident. In terms of recall (Figure 4), ERNIE 3.5-8K and DeepSeek-V3 similarly achieved high recall rates under the NAG prompt, reaching 92.47% and 85.44%, respectively. In contrast, GPT-4o exhibited a significantly lower recall rate of only 64.19%, indicating a substantial proportion of ADEs went unrecognized. However, under the SAG prompt, GPT-4o’s recall rate increased to 69.93%, and under the RAG prompt, it surged to 92.41%, representing an improvement of almost 30% (FDR P=.003). This corroborates the effectiveness of the knowledge retrieval enhancement method used in this experiment. It also highlights that while GPT-4o has considerable generative capabilities, its lack of domain-specific understanding limits its performance. However, upon incorporating contextual knowledge, its performance was substantially enhanced, ultimately bringing its overall F₁-score in line with those of ERNIE 3.5-8K and DeepSeek-V3.

We further analyzed the variations in ADE recognition performance across different models. Taking the RAG prompt as an example, the comparative analysis at different matching levels is illustrated in Figure 5. With regard to precision, all 3 models demonstrated consistently stable performance across the L1, L2, and L3 levels, with each achieving a precision rate of over 94%. In terms of recall, ERNIE 3.5-8K slightly outperformed the other 2 models. Regarding the F₁-score, there was minimal variation in the models’ recognition performance: at the L1 level, ERNIE 3.5-8K marginally surpassed the others with an F₁-score of 72.43%, though this advantage was not statistically significant compared to GPT-4o (FDR P=.86); at the L2 level, ERNIE 3.5-8K led with a score of 94.74%, significantly outperforming both DeepSeek-V3 and GPT-4o (both FDR P=.003). At the L3 level, DeepSeek-V3 slightly outperformed the others with an F₁-score of 96.38% (FDR P=.01 compared to GPT-4o, but no significant difference compared to ERNIE 3.5-8K, FDR P=.68). Thus, under the RAG prompt, the overall performance of LLMs in ADE recognition for Chinese clinical progress notes is relatively consistent, with each model displaying robust semantic comprehension and generation capabilities.

Taking the RAG prompt analysis as an example, we examined the models’ performance disparities across different matching levels (Figure 6). The performance of each model showed an upward trend from L1 to L2, and subsequently to L3 (all FDR P<.05), with the most significant improvements in recall value. For instance, at L2, ERNIE 3.5-8K demonstrated a 34.38% improvement over L1, reaching 93.16%. DeepSeek-V3 showed a 30.55% increase, achieving 85.37%; and GPT-4o exhibited a 27.68% enhancement, attaining 85.65%. These results emphasize the important role of accurate entity recognition in ADE detection. Regarding the F₁-score, the substantial increase in recall at L2 led to a similar upward trend across all matching levels.

Given that the most substantial performance variance occurred at the L2 level, we undertook a granular analysis of these cases. Focusing on the DeepSeek-V3 model as a representative example, we identified 376 instances of L2-level matches within a standard reference corpus of 2510 entries. Our analysis of these instances revealed 2 main error categories: anomalous recognition of drug entities and anomalous recognition of adverse reaction entities (Table 3). The former category was further stratified into four subtypes: (1) omission of the correct drug entity, (2) commission of an incorrect drug entity, (3) misidentification of the drug entity, and (4) lack of precision in boundary detection. Similarly, anomalies in adverse reaction entity recognition comprised: (1) entity omission, (2) entity commission, (3) failure to disaggregate compound entities, and (4) imprecise boundary detection. Quantitative analysis of these error subtypes highlighted the omission of adverse reaction entities as the predominant failure mode, accounting for around 40.7% of the L2-level discrepancies.

A statistical analysis of error instances revealed that ERNIE 3.5-8K incorrectly handled 66 case records alone, DeepSeek-V3 incorrectly handled 38 case records alone, and GPT-4o incorrectly handled 73 case records alone (Figure 7). Notably, 15 critical records posed recognition challenges for all 3 models. This suggests that, despite the integration of 3 LLMs in a composite approach, these ADE remain unrecognized. A detailed etiological analysis of these 15 universally misidentified cases partitioned the errors into 2 principal categories: failures of omission (failure to identify a true ADE) and failures of commission (identification of an erroneous ADE; Table 4). The latter constituted the majority of errors, accounting for approximately 80% of instances. The primary drivers of these commission errors were identified as (1) non–drug-related adverse events; (2) the incorrect classification of prophylactic, precautionary, or advisory medication mentions; and (3) the misinterpretation of historical ADE descriptions as contemporary events. Further investigation demonstrated that these erroneous identifications could be rectified by augmenting the models with a domain-specific ADE knowledge base via a RAG framework. This underscores the profound efficacy of incorporating a curated knowledge repository and RAG methods to enhance the precision of ADE detection.

To empirically evaluate the model’s ability to handle complex clinical language, we constructed a targeted subset of 100 clinical notes containing 3 challenging ADE-related contexts: negation, abbreviations, and temporal and causal descriptions (Table S2 in Multimedia Appendix 1). Using DeepSeek-V3 under the RAG prompting strategy, the model achieved a precision of 0.9167, recall of 0.9429, and F₁-score of 0.9296 on this subset. These findings indicate that the RAG framework can effectively process negated ADE descriptions, abbreviated drug expressions, and historical or risk-monitoring contexts, supporting its robustness in complex clinical narratives.

To validate clinical utility under real-world conditions, DeepSeek-V3 was evaluated using 1000 random resampling iterations from 19,983 uncurated clinical progress notes, with 1000 notes sampled in each iteration (Table 5). At the L3 level, RAG achieved an F₁-score of 0.8159 (95% CI 0.7736‐0.8598), compared with 0.6380 (95% CI 0.5759‐0.6942) for NAG and 0.8134 (95% CI 0.7600‐0.8627) for SAG. Despite relatively low precision under this low-prevalence setting, RAG demonstrated strong clinical utility, achieving a recall of 0.9448, F₂-score of 0.8885, G-mean of 0.9631, Cohen κ of 0.8057, and specificity of 0.9821. These findings confirm that while absolute F₁-scores slightly adjust due to the natural data imbalance, integrating a curated knowledge base via RAG remains a highly effective and practically viable approach for clinical ADE identification.

This study investigates the identification of ADEs from clinical course records by synergistically integrating LLMs with RAG technology. Using a progressive research methodology, we demonstrate that the incorporation of a RAG framework substantially enhances the performance of LLMs in the ADE recognition task. Comparative experimental analysis reveals that our proposed RAG solution, which leverages a curated ADE knowledge repository, achieves superior performance across precision, recall, and F₁-score metrics, thereby validating its considerable utility in the domain of medical text analysis. Furthermore, this work contributes a novel benchmark corpus, addressing the existing scarcity of annotated data for ADE extraction in the Chinese language.

In the absence of knowledge augmentation (using NAG prompt), the various LLMs exhibited significant performance disparities in ADE recognition. This variability likely stems from differences in their foundational training corpora, including the breadth of medical data coverage, divergent parameter optimization strategies, and the nuanced depth of their comprehension of Chinese clinical terminology [24]. Among the models evaluated, DeepSeek-V3 demonstrated optimal recognition performance under the RAG prompt. DeepSeek-V3, which claimed performance competitive with GPT-4o shortly after its release in January 2025, has rapidly gained global prominence [25]. Unlike the proprietary GPT-4o, DeepSeek-V3 is an open-source model that permits local deployment. In a local deployment, its parameter weights can remain insulated from alterations in cloud infrastructure or API updates, thereby supporting long-term consistency and reproducibility while protecting patient data privacy. These advantages create a favorable pathway for the future deployment and external validation of this architectural framework within local health care institutions. However, this study used the cloud API version of DeepSeek-V3 to improve experimental efficiency and standardize batch inference. Local deployment therefore remains a future pathway for privacy-preserving implementation, rather than the deployment mode used in this study.

In the real-world low-prevalence evaluation, the L3 F₁-score differed only marginally between SAG and RAG (0.8134 vs 0.8159), likely because the 2 evaluation settings differed substantially in data distribution. In the same uncurated setting, SAG slightly exceeded RAG in F₁-score at L1 (0.5801 vs 0.5687) and L2 (0.7799 vs 0.7663). The curated reference set was relatively balanced and enriched for ADE-positive cases, allowing dynamic retrieval to provide closely matched examples and entity-level contextual cues that improved drug and reaction recognition. By contrast, routine clinical notes had a low ADE prevalence and were dominated by negative records. In this setting, the static clinical rules embedded in SAG already captured common negative contexts, including historical ADE descriptions, prophylactic medication use, and routine monitoring statements. This strong baseline specificity limited the incremental F₁-score gain from dynamic retrieval. However, RAG improved recall at the L3 level from 0.9062 to 0.9448, indicating better detection of rare positive ADE cases. Thus, in low-prevalence clinical data, RAG primarily enhances positive-case recall, underscoring its capacity to effectively mitigate the risk of underreporting ADEs.

Our empirical findings demonstrate that the performance of a conventional, pure language model is markedly inferior to that of a variant augmented with a RAG framework. This outcome aligns with the established theoretical premise that the integration of domain-specific knowledge substantially enhances the capabilities of large models. In contrast to fine-tuning methodologies, RAG obviates the need for voluminous training corpora and protracted training cycles, and circumvents the laborious process of creating contemporary annotated datasets or engaging in repetitive model retraining to assimilate updated, customized knowledge [26]. The paramount advantage of the RAG model resides in its profound adaptability [19], enabling the continuous assimilation of the latest information on ADE cases—a feature of particular salience within the dynamic landscape of medicine and pharmacology, where novel therapeutics are developed at a rapid pace. Consequently, when information regarding an ADE for a newly marketed drug becomes available, it can be seamlessly incorporated into the ADE knowledge repository, facilitating dynamic updates without necessitating extensive modifications to the core model. Furthermore, RAG’s capacity to reference and cite antecedent ADE cases enhances the verifiability and perceived relevance of its outputs, thereby effectively mitigating the phenomenon of “confabulation,” or artifactual generation, which frequently plagues the application of LLMs in medical contexts, while concurrently improving model transparency [27]. An error analysis revealed 15 common errors that were uniformly misidentified by all tested models, exposing the inherent limitations of LLMs in processing specific complexities within clinical texts. Critically, we demonstrated that these identification failures could be rectified by supplementing the model’s context with pertinent ADE case knowledge via the RAG architecture. This not only furnishes a definitive pathway for augmenting the precision of LLMs in specialized domains but also underscores the intrinsic value of the ADE knowledge repository. Future work should therefore be directed towards the construction of a more comprehensive repository that encompasses not only ADE case reports but also multifaceted information from product information, clinical practice guidelines, and disease manifestations.

To date, systematic investigations into the application of LLMs for general ADE information recognition remain nascent, with most research focusing on narrowly defined ADEs. For instance, Cheligeer et al [28] evaluated 4 open-source LLMs in the detection of pulmonary embolism from narrative electronic medical records, while Li et al [29] demonstrated the efficacy and robustness of LLMs in extracting postvaccination adverse events from diverse data sources, including the VAERS (US Centers for Disease Control and Prevention and US Food and Drug Administration), Twitter (Twitter, Inc), and Reddit (Reddit, Inc). Our research, in contrast, underscores the broader potential of LLMs to identify all types of ADEs within unstructured Chinese clinical texts. The best-performing RAG-enhanced LLM achieved an F₁-score of 96.38%, which is higher than the performance reported in some previous deep learning–based ADE extraction studies [30]. However, because these studies used different datasets, annotation criteria, and evaluation settings, this comparison should be interpreted only as indirect evidence rather than a head-to-head demonstration of superiority. Nevertheless, our findings suggest that RAG-enhanced LLMs represent a promising framework for ADE extraction, with potential advantages in contextual reasoning, prompt-based adaptation, and flexible knowledge updating [31]. Moreover, LLMs obviate the need for extensive labeled datasets and are amenable to continuous refinement as new data becomes available [32].

The methodologies and results presented herein possess considerable potential for translation into clinical practice, where they could elevate the standards of ADE surveillance and reporting and advance pharmacovigilance and drug safety research. Moreover, our approach can be embedded within Clinical Decision Support Systems. Such integration would synergistically combine ADE information with other relevant patient data streams (such as laboratory results, imaging data, and demographic information) to provide clinicians with a more holistic and precise patient assessment, thereby guiding more timely and judicious clinical decision-making. To further enhance the practical utility of our model, we are currently developing a web-based platform, which will feature a user-friendly interface allowing clinicians to query for ADE information within texts through simple conversational input.

This study has several limitations. First, the knowledge repository was constructed from a single data source, which may result in an incomplete representation of all potential ADEs. Crucially, evaluating models on the same data distribution as the knowledge base risks overfitting and implicit data leakage within the RAG framework, limiting external validity. Multi-institutional validation, using independent data and varied clinical note formats, is therefore essential to verify real-world generalizability. In the future, the integration of multisource information, including pharmaceutical package inserts and expert-derived knowledge, would likely yield a substantial improvement in model performance.

Second, given the rapid pace of technological iteration in LLMs, the models used in this research may not represent the latest state-of-the-art (eg, the most recent Gemini [Google LLC] or Claude architectures [Anthropic PBC]). While larger-parameter models may offer performance gains, such enhancements must be carefully weighed against the concomitant increase in computational resource expenditure and memory requirements, which presents a significant challenge in resource-constrained settings.

Third, although the RAG-enhanced LLMs achieved promising performance in ADE identification, this study did not include a direct head-to-head comparison with standard clinical natural language processing baselines, such as fine-tuned clinical Bidirectional Encoder Representations from Transformers (BERT)–based models. Therefore, our findings should be interpreted as evidence of the feasibility and potential utility of the proposed framework, rather than definitive evidence of superiority over traditional deep learning approaches. Future work will include direct comparisons with conventional clinical natural language processing models on the same curated dataset and further explore LLM fine-tuning strategies to rigorously evaluate and optimize ADE identification performance.

Fourth, all clinical notes were strictly deidentified, preventing the extraction and reporting of baseline patient demographics. Despite this limitation, the generalizability of our findings is preserved through the random sampling of heterogeneous data.

Fifth, although the RAG model achieved high specificity (0.9821) in the real-world evaluation, this corresponds to an approximate false-positive rate of 1.79%. At the observed ADE prevalence of about 4.6%, this would translate into roughly 17 falsely flagged ADE-negative notes per 1000 screened notes, which could increase manual review burden or alert fatigue if deployed without clinician oversight. Future work should calibrate decision thresholds, retrieval filtering, human-in-the-loop review, and iterative knowledge-base expansion to balance sensitivity against false-alert burden while further improving RAG effectiveness and reducing hallucination-related false positives.

Sixth, all performance metrics in this study were computed at the document level. This evaluation strategy is clinically meaningful for screening whether a note contains ADE information, but document-level precision does not capture entity-level overextraction within otherwise ADE-positive notes. Future work should therefore supplement document-level evaluation to more comprehensively assess extraction granularity and hallucinated entities within positive notes.

In light of these constraints, our future research will explore the incorporation of knowledge graph embedding techniques [33]. By leveraging graph-based structural reasoning, this approach could facilitate the logical inference of ADE knowledge not explicitly contained within the existing database, thereby empowering users to identify potential ADEs with greater accuracy. To further enhance the clinical utility of the proposed system, future investigations should focus on expanding the ADE knowledge base framework and evaluating the system’s performance in diverse, real-world clinical settings. To facilitate this translation, we recently developed ADESys [34]. Future work will leverage this system for multicenter external validation to enhance our method’s robustness and translational value. Such efforts are essential to reducing ADE underreporting and increasing the overall efficiency of pharmacovigilance research.

This study evaluated the effectiveness of a knowledge base-driven RAG framework in identifying ADE information from Chinese clinical narratives. The experimental results across 3 LLMs (DeepSeek-V3, ERNIE 3.5-8K, and GPT-4o) indicate that the RAG strategy consistently improves identification performance compared to NAG and SAG approaches. Furthermore, this research establishes a valuable benchmark, addressing the critical gap in ADE corpora within the Chinese language domain.