Recent advances in artificial intelligence, particularly in large language models (LLMs), have demonstrated remarkable capabilities of automated data extraction and organization from complex clinical documents [1-3]. These artificial intelligence–driven approaches can be used to process large volumes of clinical data with a consistent methodology, potentially reducing human bias and improving the collection efficiency of research data. Although LLMs show promise in health care applications, few studies on their practical efficacy compared with that of traditional, manual processing by physicians have been published, particularly in complex areas such as the extraction of surgical oncology data [4].

In the field of breast cancer surgery, retrospective data analysis presents unique challenges owing to the complexity of unstructured clinical data. Electronic health records (EHRs) contain diverse information across clinical charts, operation records, and pathology reports, often in a free-text format. The complexity is compounded by breast cancer–specific characteristics, including the bilateral nature of the organs, concurrent malignant and benign lesions, and multiple radiological features. This complicates automated data curation and has traditionally necessitated manual review by physicians for accurate data interpretation and collection.

However, this manual approach has several limitations. As the volume of clinical data increases, consistency in physician reviews becomes increasingly difficult to maintain, potentially leading to discrepancies in data interpretation [5,6]. The time-intensive nature of manual review and the risk of errors in the processing of large volumes of clinical data present considerable challenges in retrospective research [7-10]. In addition, direct EHR access for manual data extraction raises privacy concerns when sensitive patient information is handled [11-13]. Although LLM-based automation of information extraction from anonymized EHR data may address these challenges, its effectiveness compared with that of traditional physician reviews remains to be evaluated.

Although LLMs have been validated for the extraction of specific medical data, their potential for the curation of comprehensive data of patients with cancer remains largely unexplored [14]. In this study, we compared traditional physician reviews with LLM-based processing of anonymized clinical data in the field of breast cancer, focusing on the development of a practical approach for surgical oncologists. We hypothesized that LLM-based analysis would yield comparable results to manual review in handling large volumes of clinical data while reducing processing time and resource use.

This retrospective comparative study included patients with breast cancer who underwent surgery at 5 academic hospitals from January 1, 2019, to December 31, 2019. This study was designed to compare 2 practical data curation pathways available in real-world clinical research settings. The manual review pathway represents the traditional method of direct EHR access. The LLM processing pathway uses preextracted, deidentified data. This study adheres to the Transparent Reporting of a Multivariable Prediction Model for Individual Prognosis or Diagnosis–Large Language Model (TRIPOD-LLM) reporting guidelines (Checklist 1).

We compared 2 data extraction methods: manual physician review (group 1) and LLM-based processing (group 2). In group 1, 1 dedicated breast-surgical oncologist from each hospital reviewed data spanning 2 years (2019‐2020) over 7 months (May 2021–November 2021) using a standardized data collection form (Multimedia Appendix 1). The data encompassed 89 clinical variables across 3 domains: patient demographics (basic information, medical history, and family history), treatment information (surgical details, neoadjuvant or adjuvant therapy, complications, and follow-up treatment), and pathological information (tumor characteristics, tumor stage, biomarker status, and margin status). Follow-up observations regarding recurrence and mortality were updated until January 2024.

Patients in group 2 were initially identified using the clinical data warehouse (CDW) of Catholic Medical Center, an integrated data platform of 8 affiliated academic hospitals in Korea [15,16]. The CDW supports this research by providing anonymous clinical data to investigators following institutional review board approval [15]. The LLM structured 31 clinical factors from the raw data, including patient demographics (basic information, survival data, and diagnostic data), treatment information (surgery types and neoadjuvant or adjuvant therapy), pathological information (tumor characteristics, tumor stage, biomarker status, and nodal status), and imaging features. Data extraction and curation were performed from October 20, 2024, to November 1, 2024.

The CDW query identified 17,317 patients diagnosed with invasive breast cancer or ductal carcinoma in situ between July 2018 and July 2021. From this cohort, we selected patients diagnosed during the study period (January 2019-December 2019) who underwent breast cancer surgery. CDW extraction included unstructured EHR reports containing clinical information, operation records, and pathology reports (Multimedia Appendices 2-4). Follow-up data through October 31, 2023, were used.

Unstructured data extracted from the CDW were processed using Claude 3.5 Sonnet (Anthropic) to extract and structure the required factors into predefined categories.

The final prompt focused on accurate extraction from 4 distinct data categories: diagnostic information, clinical measurements, surgical procedures, and pathological findings.

Prompt execution of tasks in a specific sequence to ensure data integrity:

The prompt instructed the LLM to generate outputs directly in .CSV format with predefined column structures, automated date formatting (YYYY-MM-DD), and standardized missing data codes (999,999), eliminating the need for extensive after processing. The structured output was validated through the methodology described in the Data Quality Assessment and Validation section.

This study aimed to assess the feasibility of replacing manual physician reviews with LLM-based processing of breast cancer–related clinical data. We compared the demographic characteristics, clinical parameters, treatment patterns, disease characteristics, and survival outcomes between the 2 groups.

Categorical variables were compared using chi-square or Fisher exact tests, with agreement assessed using Cohen κ coefficient (κ<0.20=poor, 0.21‐0.40=fair, 0.41‐0.60=moderate, 0.61‐0.80=good, >0.80=very good). Continuous variables were analyzed using the Student t test and the intraclass correlation coefficient. Effect sizes were calculated using Cohen d (continuous) and Cramér V (categorical).

Overall survival was analyzed using the Kaplan-Meier method and compared using the log-rank test. Both approaches were validated using the Korean Breast Cancer Society (KBCS) 2019 national registry data by comparing age, tumor stage, surgical procedures, molecular subtypes, and survival trends [17].

For validation, 50 cases from each group were selected using proportionate stratified random sampling. Stratification was based on the cancer stage (0-IV) and type of surgical intervention (breast-conserving surgery vs mastectomy) to ensure representative sampling across key clinical categories. Random selection was performed using Python (version 3.8; Python Software Foundation) with the NumPy (v.1.21.0) and pandas (v.1.3.0) libraries and a fixed random seed of 2,02,41,201 for reproducibility. Four breast-surgical oncologists (SJO, JPY, HK, and SL) independently evaluated 18 predefined clinical factors in each case (900 data points per group). Accuracy rates were calculated as the percentage of correctly extracted factors relative to the total number of factors. A dual-reference validation approach was implemented: group 1 was validated against the EHR, whereas group 2 was compared to the CDW raw data. The evaluation included both present and missing values. The accuracy threshold was set at 90% based on previous validation studies of clinical data extraction systems [18].

This study was approved by the institutional review board of the Catholic Medical Center (approval: OC24WIDI0138). As this was a retrospective analysis of existing clinical data, the requirement for informed consent was waived by the institutional review board. All patient data were deidentified prior to analysis, with personal identifiers replaced by anonymized codes. The CDW platform ensures privacy protection through automated deidentification. No compensation was provided to participants, as this study involved retrospective data analysis only.

For comparative analysis, 18 key clinical factors were selected from both groups (Figure 1). The manual review (group 1) and LLM processing (group 2) groups comprised 1366 and 1734 cases, respectively. Although both groups completely captured age data, they exhibited different patterns of missing data for the other parameters. Group 2 had higher missing rates in terms of cancer stage (12.2% vs 3.1%) and HER2 status (15.1% vs 11.0%), whereas group 1 had more missing data for lesion size (20.5% vs 5.9%) and lymph node assessment (21.5% vs 8.8%). Both groups maintained high documentation rates (>90%) for hormone receptor status (Figure 2).

The validation analysis encompassed 1800 data points (900 per group) across clinical factors. Group 1 demonstrated perfect accuracy with no discrepancies. Group 2 exhibited 83 discordant factors out of 900 data points with an accuracy rate of (817/900, 90.8%). Among 1734 patients in group 2, 260 (15%) underwent multiple surgical procedures. The LLM successfully integrated data from sequential operations in 53% (138/260) of these cases, while missing data integration in 47% (122/260).

LLM-based processing demonstrated efficiency gains compared to manual review (Table 1). Group 1 required 7 months (May 2021-November 2021) with 5 dedicated breast surgical oncologists. In contrast, group 2 processing was completed in 12 days (from October 20, 2024, to November 1, 2024) by 2 physicians, with data extraction taking 10 days and validation requiring 2 additional days.

Manual review required approximately 1025 physician hours total across 7 months. LLM processing required approximately 96 physician hours in total over 12 days, a 91% reduction in time investment.

Resource use differed substantially between the methods. The manual review required 5 breast surgical oncologists with a direct EHR access infrastructure and dedicated data entry personnel. LLM processing required 2 physicians using a standard workstation without EHR access. The API cost for LLM processing was approximately $0.15 per case, totaling $260 for 1734 cases. Although direct personnel costs were not calculated due to institutional variations in physician compensation, the 91% reduction in physician hours represents substantial resource savings.

The baseline characteristics of both groups are summarized in Table 2. The mean age differed slightly between groups 1 and 2 (55, SD 11.5 vs 53.5, SD 11.4 y; P<.001; Cohen d=0.13). For breast surgery, total mastectomy was performed in 19.2% of cases in group 1 and 26.5% in group 2, while nipple (skin)-sparing mastectomy rates were 15.7% and 9.6%, respectively (χ^2₃=164.3; P<.001; Cramér V=0.29). When these procedures were combined, groups 1 and 2 exhibited similar proportions of breast-conserving surgery (63.5% vs 63.9%) and mastectomy (34.8% vs 36%), with a small effect size (Cramér V=0.10).

In terms of axillary surgery, group 1 had a higher rate of SLNB than group 2 (68.4% vs 59.7%), whereas the rates of ALND were similar (χ²₂=47.2; P<.001; Cramér V=0.13). The mean (SD) value of the harvested lymph nodes was similar between the groups (7.79, SD 7.03 vs 7.11, SD 7.20; P=.016).

The stage distribution differed between groups (χ²₈=68.9; P<.001), but only slightly (Cramér V=0.16), with group 1 identifying more cancers as advanced. Hormone receptor status was similar between the groups (estrogen receptor: 78.3% vs 76.2%, P=.172; progesterone receptor: 68.7% vs 67.5%, P=.525). HER2 status differed negligibly, and Ki67 expression was similar between the groups (HER2: P=.003, Cramér V=0.003; Ki67: P=.391, Cramér V was approximately 0.00). Histological grade distributions were similar between the groups (P=.764).

ICC analysis of continuous variables demonstrated a consistently low agreement: age (ICC 0.013, 95% CI –0.035 to 0.060), tumor size (ICC 0.029, 95% CI –0.021 to 0.078), number of metastatic lymph nodes (ICC 0.031, 95% CI –0.019 to 0.081), number of harvested lymph nodes (ICC 0.025, 95% CI –0.025 to 0.075), and Ki67 expression (ICC 0.027, 95% CI –0.023 to 0.077). All ICC values were negligible, and all CIs included zero.

Survival analysis revealed significant differences between the groups (hazard ratio 2.917, 95% CI 1.496 to 5.688; P=.002). Group 2 captured more events (11 vs 41). The proportional hazards assumption was met (χ²_₁=2.37; P=.120), and the log-rank test confirmed a difference in survival distributions (χ²_₁=10.9; P=.001).

Comparison with the KBCS 2019 registry data (N=9447) revealed small differences in breast surgery patterns for both groups (Cramér V=0.03‐0.04; P≤.018; Table 3). For axillary surgery, both groups had lower SLNB rates (group 1: 68.30% and group 2: 59.75% vs KBCS: 73.18%) and similar ALND rates (20.50% vs 20.88% vs 18.60%). Group 2 had a higher rate of no axillary surgery than group 1 (19.37% vs 11.20%).

Stage distribution analysis revealed significant but small differences from the national data (group 1: Cramér V=0.076, P<.001; group 2: Cramér V=0.038, P=.003). Regarding biomarker subtypes, both groups had slightly higher proportions of hormone receptor–positive with HER2-negative (group 1: 67.02% and group 2: 66.57% vs KBCS: 63.14%) and triple-negative cases (12.70% and 13.59% vs 11.98%) with minimal effect sizes (Cramér V=0.03‐0.04).

This comparative study demonstrated that LLM-based processing achieved 90.8% accuracy in extracting clinical data from breast cancer records, with significant reductions in processing time (12 d vs 7 mo) and resource requirements (2 vs 5 physicians). The LLM approach captured substantially more survival events (41 vs 11; P=.002) and showed better documentation of lymph node assessment (91.2% vs 78.5%), although it had higher missing rates for integrated assessments such as cancer staging (12.2% vs 3.1%). Both methods yielded similar patterns in key clinical parameters, including breast-conserving surgery rates (63.5% vs 63.9%) and biomarker distributions.

A particular finding was the substantial difference in captured survival events between LLM processing (41 events) and manual review (11 events). This discrepancy may reflect fundamental differences in how humans and LLMs approach large-scale data extraction. Manual reviewers processing many charts may inadvertently adopt a mechanical approach, focusing on the most obvious data fields, while potentially overlooking mortality information scattered across multiple sections of the medical record. In contrast, LLM maintained consistent thoroughness throughout the extraction process, systematically examining all available data sources for each case without the cognitive fatigue that affects human reviewers during repetitive tasks.

This finding challenges the assumption that manual review is the gold standard for all types of clinical data extraction. While human expertise remains essential for complex clinical interpretation, our results suggest that LLM processing may provide a more complete capture of certain objective outcomes, particularly those requiring synthesis across multiple data fields. However, this interpretation requires further research to confirm whether the additional events captured by the LLM represent true positives or extraction errors.

Although LLMs have shown favorable results in extracting specific medical data from radiology and pathology reports [19-22], our study represents the first comprehensive evaluation of surgical oncology data curation. Previous studies have focused on the extraction of single data points or specific types of reports. For instance, Park et al [20] demonstrated the effectiveness of LLM in extracting pulmonary disease information from radiology reports, whereas Cheng [22] reviewed its applications in pathology. Our study extends these findings by demonstrating that LLMs can handle the complex integration required for comprehensive oncological data including surgical procedures, pathological findings, and survival outcomes.

The observed statistical differences between groups, while significant, were mostly clinically negligible (Cramér V<0.30 for all comparisons), suggesting that LLM processing maintains clinical validity. The observed accuracy of 90.8% is comparable to recent studies, where LLMs achieved similar performance in extracting structured information from clinical notes, including the social determinants of health [23]. Although comprehensive systematic reviews of LLM applications in healthcare are still emerging [3], individual studies have consistently demonstrated their potential for automated clinical data extraction. However, unlike previous studies that focused on extracting discrete clinical variables or single-domain information [19-23], surgical oncology data require integration of multiple interconnected factors, presenting unique challenges in automated extraction.

Our study had several strengths. The large sample size provided robust statistical power. The use of real-world clinical data from multiple institutions enhanced the generalizability of the results. The head-to-head comparison with a manual physician review rather than against a reference standard alone provides practical insights for clinical research applications. Additionally, the demonstrated efficiency gains (12 d vs 7 mo) highlight the potential for scaling oncological research.

However, several limitations of this study must be acknowledged.

First, LLMs showed limitations in integrating multiple clinical events. Although the model performed well in extracting explicit data points, it struggled to synthesize information across sequential surgical procedures. This was evidenced by the higher rate of missing surgical data (12.2% vs 3.1%) in patients who underwent multiple operations. Manual reviewers could identify and integrate multiple surgical steps, such as lymph node assessment after initial diagnostic excision, while the LLM typically captured data from a single representative operation.

Second, the validation sample size of 50 cases per group represented only 2.9% and 2.6% of the respective cohorts. While this sample was stratified to ensure representation across cancer stages and surgical types and achieved adequate power for detecting clinically meaningful differences, a larger validation set would strengthen confidence in the accuracy estimates.

Third, differences in data sources between the groups may have affected direct comparability. The manual review group accessed complete EHRs directly, whereas the LLM group processed CDW-extracted data. This methodological design was unavoidable; manual reviewers needed direct EHR access for a comprehensive review, while the LLM required deidentified extracted data for processing. However, this means we cannot isolate whether performance differences stemmed from the extraction approach itself or from the inherent differences in available data. Future studies comparing both methods using identical raw data would overcome this limitation.

Fourth, generalizability to other LLM models requires evaluation. We used Claude 3.5 Sonnet, but the performance may vary across models, versions, and prompting strategies. The rapid evolution of LLM capabilities suggests that our findings represent a snapshot of current technology rather than definitive limits.

Future research should address several key issues. The first is the development of improved prompting strategies to handle complex clinical scenarios that require integrated assessment, particularly for sequential surgical procedures and temporal relationships. Second, LLM processing is compared with manual review using identical raw data sources to isolate the actual performance differences between the methods. Third, feasibility studies of integrated data curation across multiple clinical events are needed to overcome the current limitations in synthesizing longitudinal patient data. Fourth, a systematic examination of the performance characteristics of different LLM models is required to identify optimal models for specific clinical data types.

LLM-based processing demonstrated comparable effectiveness to manual review by physicians for breast cancer clinical data extraction, while significantly reducing processing time and resource utilization. Despite the limitations of integrated assessments requiring synthesis across multiple clinical events, this approach offers a solution for efficient clinical data extraction in oncology research. The ability to process large volumes of data consistently and rapidly while maintaining an accuracy above 90% suggests that LLM-based methods can accelerate retrospective clinical research.