This study used a structured pipeline that systematically progresses from raw medical report processing to final clinical staging output, as illustrated in Figure 1. This end-to-end framework facilitates reproducible model development through a structured transformation of unstructured reports into staging conclusions that adhere to clinical standards.

This study was approved by the Shanghai Ethics Committee for Clinical Research (SECCR2025-260) as a retrospective analysis. The requirement for informed consent was waived due to the retrospective nature of the study and the use of deidentified data. All patient information was anonymized prior to analysis, with all identifiers, medical record number/visit number, hospital name, attending physician’s name, examining doctor’s name, and other personally identifiable information such as ID number and contact details removed, to ensure privacy and confidentiality. No compensation was provided to participants as this was a retrospective study using existing clinical data. The study did not involve any images that could potentially identify individual participants. This research was conducted in accordance with the principles of the Declaration of Helsinki and adhered to all relevant regional and national research ethics guidelines.

In this study, to enhance the practical application of AJCC 8th Edition guidelines within Chinese clinical contexts, we developed supplementary TNM staging rules through expert clinical consultation. These rules address specific scenarios either not explicitly covered in the official guidelines or requiring disambiguation for Chinese medical terminology. A comprehensive mapping table (Table 1) documents each rule alongside its corresponding AJCC guideline reference. The supplementary rules primarily address: (1) Clinical interpretations for anatomically complex scenarios (eg, trans-lobar growth), and (2) Disambiguation of Chinese medical terms with overlapping semantics (eg, distinguishing between local invasion vs metastatic spread).

Note that all supplementary logic maintains consistency with AJCC principles while improving alignment with Chinese clinical documentation practices. These enhancements ensure more accurate staging outcomes without contradicting established guideline definitions.

In accordance with our supplementary staging rules, we adopted a conservative interpretation for M staging: lesions are classified as metastatic only when the report contains terminology with high specificity for metastasis (eg, “转移” and “考虑转移”). Descriptions such as “小结节” (small nodule) or “高密度影” (high-density shadow), among others, which are common but nonspecific in Chinese radiology reports, are not considered sufficient evidence on their own, though they may contribute to a metastatic inference when supported by additional contextual findings. This strategy is designed to minimize false positives. While this approach could theoretically increase false negatives, we conducted a comparative analysis using a more lenient interpretation strategy. This alternative approach removed key restrictive rules, including:

The results of the comparison of 2 strategies on M staging are shown and discussed in the Clinical Impact Assessment subsection of the Model Evaluation section in Results and Discussion, respectively.

Finally, the overall clinical staging (Stage IA/IB, IIA/IIB, IIIA/IIIB/IIIC, or IVA/IVB) was algorithmically derived from the predicted T, N, and M staging results. This derivation strictly adhered to the stage grouping rules defined in the AJCC 8th Edition Cancer Staging Manual. The definitive mapping table from TNM combinations to clinical stages, as provided by the AJCC guidelines, is illustrated in Figure 3.

The GLM series, developed by Tsinghua University and Zhipu AI, has been pretrained on approximately 10 trillion tokens of multilingual data, predominantly in Chinese and English [25]. GLM-4-Air, an optimized variant of GLM-4 within the ChatGLM family, was selected for this study. This 32-billion-parameter model preserves the robust performance of its predecessor while offering reduced processing latency and lower inference costs, making it a more viable candidate for local deployment scenarios where data privacy is paramount, and the prohibitive hardware requirements of much larger models are a practical constraint.

The SFT process adjusts the model parameters using supervised learning techniques. This allows the model to adapt to specific data distributions and task requirements, leading to improved performance. In this study, we used low-rank adaptation (LoRA) technology for SFT on the GLM-4-Air model. LoRA is a parameter-efficient fine-tuning technique. It reduces computational complexity and memory requirements by decomposing the weight update matrix while maintaining fine-tuning effectiveness [26].

Based on our annotated training dataset, we reprocessed the stage labels. To be specific, the stage label was reprocessed to the reasoning process consistent with the optimized prompt instructions by professional physicians (Textbox 5). This made the training data more comprehensible to the GLM-4-Air model, providing a solid foundation for model training.

To deal with the potential risks of overfitting and catastrophic forgetting associated with the limited sample size, we introduced a key design in the SFT stage for the core task. Specifically, we constructed a heterogeneous instruction-tuning dataset by blending the 292 target-domain cases with approximately 2000 general-purpose instruction samples covering diverse tasks such as mathematics, summarization, and translation. The details of the categories, exact counts, proportions, and specific examples for each category of the instruction samples are provided in Multimedia Appendix 3. Based on our empirical assessment, these components were mixed at a ratio of approximately 1:7. This strategy was designed to leverage a multi-task learning framework, compelling the model to adapt to the specific medical task while simultaneously preserving and reinforcing its general instruction-following and reasoning capabilities, thereby enhancing its generalization stability.

In the fine-tuning phase, we used the LoRA method for SFT with the following parameters (Table 2), and the GPU we used in this task was NVIDIA A800.

These parameters were carefully tuned to ensure efficient training and stable performance of the model.

During the SFT process, we used white-box testing to identify and analyze deficiencies in the model’s medical knowledge and capabilities. Based on these findings, we conducted targeted parameter optimization and fine-tuning to enhance its TNM staging performance. The results showed that this refined tuning strategy, based on continuous feedback and optimization, improved the model’s accuracy and stability in NSCLC TNM staging.

To determine if the observed performance differences between models were statistically significant, we applied paired significance testing. Specifically, for the key metric of classification accuracy, we performed McNemar’s test to compare the paired categorical outcomes of our final model (GLM-4-Air) and the baseline model (GPT-4o).

To calculate a standardized measure of the strength of association between categorical variables, we used the statistic Cohen ω [27,28], defined by the following formula. This metric provides a standardized effect size that is independent of sample size, facilitating comparisons across studies.

The formula for Cohen ω is:

where χ² is the chi-square test statistic, reflecting the overall deviation of the observed data from the expected distribution, and n is the total sample size.

According to Cohen's conventional benchmarks, the effect size can be interpreted as follows: ω of 0.1 indicates a small effect, ω of 0.3 a medium effect, and ω of 0.5 a large effect. This statistic provides an objective assessment of the practical importance of observed differences or associations, moving beyond reliance solely on the statistical significance of P values.

In clinical practice, the consequences of a misclassification in TNM staging are not uniform; an error can range from being inconsequential to leading to significant deviations in treatment planning. Therefore, beyond conventional performance metrics, we conducted a clinical impact analysis to evaluate the practical implications of model errors.

Based on the confusion matrices, all misclassifications were systematically categorized into 3 tiers according to their potential impact on clinical decision-making as follows:

All erroneous predictions made by the models on the test set were retrospectively reviewed, classified according to this 3-tiered system, and quantified. This analysis provides a crucial safety-centric perspective on model performance, assessing its reliability and potential for adoption in real-world clinical workflows.

In addition, a detailed breakdown of error examples and their categorization rationale is provided in Table 4.

To assess the model’s deployment feasibility on accessible hardware, we performed a local performance benchmark using consumer-grade GPUs (Four NVIDIA GeForce RTX 4090, 24 GB VRAM each). The GLM-4-Air model was loaded at full precision with an 8K token context length, and its inference latency for each TNM staging component was measured on the black-box test set.

The Chinese character error rates on the randomly selected 27 reports from white-box and black-box settings were 0.26% and 0.22%, respectively, which is consistent with the data provided by the supplier [22]. Manual review confirmed that no serious errors were found in the extraction of core information critical for TNM staging, such as tumor size, nodal status, or metastasis presence, in all cases. Specifically, the primary errors in this study were confined to issues like erroneous line breaks, spurious spaces between characters, omitted/misplaced punctuation, and poor paragraph segmentation, none of which impacted the model’s TNM staging decisions. The detailed results of the post-OCR Chinese character error rate test and examples of common OCR error types can be found in Multimedia Appendix 4.

The result of the calculation of interannotator agreement using the Cohen kappa coefficient among the clinical annotators is shown in Table 5. For the white-box set, Kappa values were 0.714 (T stage), 0.755 (N stage), and 0.778 (M stage). In the black-box set, agreement levels were 0.851 (T stage), 0.796 (N stage), and 0.796 (M stage). All Kappa values exceeded 0.70, indicating substantial agreement across all staging categories in both test sets. The T stage demonstrated the most pronounced improvement in interrater reliability between white-box and black-box evaluations, while N and M stages maintained consistently high agreement levels across both settings.

The results of this study demonstrate a general and substantial improvement in staging accuracy across successive prompt iterations. While an isolated decrease was observed in M staging from Version 1 to Version 2, all T and N stages, as well as the overall clinical stage, exhibited a consistent upward trend in accuracy from Version 1 through Version 4. Figure 4 and Table 6 compare the accuracy rates across 4 versions of prompts (0705 v, 0708 v, 0715 v, and 0801 v) for T, N, M, and clinical staging. The accuracy rates improved across all staging categories: T staging rose from 18.7% to 65%, N staging from 52.5% to 89%, and M staging from 56.2% to 90%. These improvements led to an increase in overall clinical staging accuracy from 44% to 82%.

The notable improvements occurred in 2 instances: first, between version 1 and version 2 for T staging, with accuracy increasing by 30.2% following the implementation of structured information extraction; second, between Version 2 and Version 3 for M staging, through the implementation of diverse prompts for independent T, N, and M staging. The data also showed that M and N staging maintained consistently higher accuracy than T staging across all versions.

Analysis of the LoRA fine-tuning results (Figures 5-7) led to the selection of optimal models for T and N staging, with learning rates of 4.5e-4 and 2.5e-4, respectively. For M staging, as no fine-tuned models exceeded the original model’s performance, the baseline GLM-4-Air model was retained, maintaining its accuracy of 90%.

The results demonstrated the effectiveness of the training process, particularly for T and N staging tasks. The accuracy of T staging improved from 65% to 91%, representing a substantial increase of 26%.

The principal finding of this study is that a mid-scale LLM, when optimized through a systematic framework of clinical prompt engineering and SFT, can be developed into a highly effective and more practical tool in this specific clinical task, demonstrating the efficacy and potential of domain-specific adaptation. Our optimized GLM-4-Air model achieved higher accuracy in NSCLC TNM staging relative to the general-purpose GPT-4o, alongside a notable reduction in clinically critical errors and practical deployability on consumer-grade hardware with latencies acceptable for clinical workflows. These performance outcomes, particularly evident in the semantically complex N and M staging tasks, suggest that our hybrid optimization strategy can effectively enhance domain-specific reasoning and guideline adherence required for clinical interpretation. Furthermore, these results provide evidence that a specialized, moderately sized model can achieve high accuracy in complex clinical tasks, offering a complementary pathway to the scale-driven paradigm and highlighting the critical role of domain-specific optimization.

One of the keys to our research was the collection of high-quality, real-world data. We therefore invested considerable effort in data preparation, working closely with medical experts for rigorous validation and cleaning, which provides high-quality source material for the OCR process. We believe OCR is a mature technology now, and this study relies fundamentally on the LLM’s capacity for deep semantic understanding and information extraction from medical report text, not on exact string matching in a traditional way. Pretrained on massive text corpora, the LLM possesses inherent and robust capabilities for tolerating and correcting common OCR-induced noise by leveraging contextual information.

Our analysis also revealed that data quality, especially annotation quality, significantly impacts model performance, with even a small proportion of mislabeled samples potentially undermining the overall results. Given our limited dataset size, maintaining annotation quality became crucial. To ensure annotation quality and consistency, we developed standardized protocols aligned with AJCC TNM staging guidelines and implemented a multi-expert review mechanism. Annotators were instructed to make decisions strictly based on the medical reports’ textual content, minimizing subjective bias.

From the results of calculating and reporting the interannotator agreement using the Cohen kappa coefficient in Table 5, we observed an enhancement in T-stage agreement within the black-box set. This improvement may be attributed to the annotation process for the white-box set itself, possibly acting as a calibration phase. Through challenging cases, the annotators likely developed a more consistent internal standard for interpreting ambiguous T-stage descriptors (eg, invasion vs abutment). This refined, shared understanding was then applied during the black-box annotation, leading to higher immediate agreement. The sustained high agreement for N and M stages across both conditions reflects the more binary nature of nodal and metastatic assessment compared with the finer gradations required for T-stage determination. These findings underscore the context-dependent nature of staging reliability and highlight the value of multi-environment validation in assessing clinical annotation consistency.

Building upon the high-quality dataset, we realized three key strategies in prompt optimization were important to enhance the model’s performance: (1) structured output and CoT design, (2) independent modular architecture design, and (3) background knowledge injection.

First, structured output and CoT design were adopted to improve the readability and interpretability of the model’s output. In each staging task, the model not only provided the final conclusion but also presented the logical reasoning path. For example, in T staging, the model might state, “The tumor diameter exceeds 7 cm and invades the chest wall structures, thus meeting the criteria for T4 staging.” This approach ensures transparency, facilitating manual verification and error adjustment. Collaborating with medical professionals for white-box testing allowed us to analyze prediction errors and optimize prompts effectively.

Second, we implemented an independent modular architecture. We treated T (tumor size and location), N (lymph node metastasis), and M (distant metastasis) staging as separate tasks to reduce interference between staging judgments. To be specific, earlier versions had a 35% error rate in N staging due to mixed criteria. Modularization allowed the model to concentrate on specific tasks, increasing accuracy by approximately 30% and reducing cross-interference.

For example, in a case with the description “T9 vertebra suspected metastatic bone tumor,” the model initially misclassified it as M1a. After prompt optimization, it correctly identified it as T4 staging. This improvement was achieved through a specialized T staging prompt that detailed specific invasion criteria. The prompt listed T4 invasion sites, including diaphragm, mediastinum, heart, major vessels, trachea, recurrent laryngeal nerve, esophagus, vertebral bodies, and carina. This guided the model to correctly extract “T4 Invasion Sites”: “vertebrae.”

The model’s reasoning demonstrated clear compliance with the instructions: “The report indicates T9 vertebra involvement, which aligns with T4 staging criteria. While the tumor diameter is less than 50 mm and no T3 invasion sites are present, the vertebral involvement confirms T4 staging.”

Finally, to address the issue of misjudgment due to the lack of professional medical knowledge in the model’s practical application, we applied a Background Knowledge Injection strategy. This strategy involved incorporating common error patterns and detailed professional criteria from medical imaging analysis.

The background knowledge serves as a reference for each decision-making process. For instance, we observed that the model tended to incorrectly interpret small lesions in Chinese, such as cysts and thickening, as evidence of distant metastasis. This led to significant errors in M staging.

To address this, we made an additional injection of specific judgment criteria and medical terminology explanations, according to supplementary rules shown in Table 1, such as:

Additionally, we clarified the criteria for key terms such as “tumor invasion,” “cancerous nodules,” and “lymph node metastasis.”

The integration of professional knowledge improved the model’s ability to distinguish between minor abnormalities and actual metastatic lesions. As a result, the misjudgment rate for minor abnormalities decreased from 29% to 6%, while M staging accuracy increased from 70% to 90%.

The prompt optimization framework central to this study, which involves establishing a reasoning baseline, extracting key information, separately addressing T/N/M components, and injecting domain knowledge, was designed as a model-agnostic and universally applicable framework. To ensure a controlled and fair comparison, the iterative prompt optimization was informed by bad-case analysis from both models, and the final prompt set was applied uniformly to all evaluated models. This approach underscores that our primary objective was to validate a generalizable optimization pathway for the clinical task, not to orchestrate a maximally tailored benchmark for any specific model.

It should be acknowledged that the performance of any LLM, including generalist models, such as GPT-4o, could potentially be further enhanced with dedicated, model-specific prompt tuning. This is an inherent consideration when comparing a domain-optimized pipeline with a general-purpose model. Our findings and the proposed framework are positioned within the context of this shared, task-centric optimization objective.

After the prompt optimization process, white-box testing results indicated that the model still faces challenges in precise numerical calculations and prompt adherence. Especially for precise numerical comparisons, it is difficult for further prompt optimization to solve this problem due to the lack of related abilities in the original model itself. Therefore, we introduced SFT into the training framework to further enhance model capabilities.

The SFT results showed significant improvements in domain-specific tasks, especially in tumor size measurements and classification. This demonstrated that SFT was an effective strategy for enhancing the model’s numerical computation abilities. Interestingly, we also observed significant differences in the fine-tuning results of different staging models: T staging accuracy improved markedly, while N and M staging did not. The results for M staging (Figure 7) show that the fine-tuned models achieved accuracies comparable to the baseline, with the highest-performing configuration matching the baseline accuracy of 90%, while others were marginally lower. We first clarify that, based on experimental review, neither label noise in the M-staging training data nor the learning rate configuration was the primary cause of this performance plateau. We attribute the observed results primarily to the inherent characteristics of the task and the interplay within our multi-task SFT framework.

First, N and M staging tasks have fewer subcategories compared with T staging. The baseline model, powered by extensive pretraining, may already be near a performance ceiling for tasks heavily reliant on key information retrieval. Consequently, the baseline accuracies for N and M staging were high (89% and 90%, respectively), leaving limited absolute room for improvement via SFT on our dataset. This is particularly relevant for M staging, a binary task (M0 vs M1) with relatively definitive criteria that depend on detecting explicit descriptions of distant metastasis.

Building on this, a more granular technical analysis reveals underlying mechanisms. Beyond the task’s retrieval-intensive nature, the limited number of M1-positive cases in our dataset may constrain the model’s ability to learn the subtle semantic distinctions required (eg, between M1a “contralateral lung nodules” and benign findings). More importantly, within our single-stage, multi-task (T, N, M) joint SFT framework, we hypothesize that conflicting optimization signals may have arisen. T and N staging, as complex tasks demanding deep semantic reorganization, likely generated stronger gradients that dominated the model’s optimization trajectory. In contrast, M staging performance relies heavily on the precise pattern-matching capabilities inherent in the base model. The gradient updates aimed at learning complex semantic mappings for T/N tasks may have inadvertently perturbed these foundational retrieval mechanisms. When the training signal for the M task is relatively limited, this “interference cost” could offset any task-specific gains, resulting in a neutral net performance benefit. It is important to note that we present this “interference cost” as a theoretical hypothesis for the observed results, and it remains to be validated through controlled single-task experiments.

Therefore, the observed plateau in M staging primarily reflects a misalignment between the task’s needs and the full-parameter SFT approach in a multi-task setting, rather than a model flaw. This analysis directly corroborates the rationale for our final hybrid strategy. For M staging, preserving the base model’s strong inherent retrieval ability and guiding it precisely via prompt engineering proved to be a more robust and parameter-efficient solution than full semantic remapping via SFT. Conversely, the significant accuracy gains in T and N staging validate that SFT is highly effective for tasks that require the deep semantic comprehension and reasoning it is designed to enhance.

In addition, based on the comparative analysis of strict versus lenient interpretation strategies for M staging, the strict strategy proved highly effective in minimizing false positives while not significantly increasing false negatives. In contrast, the lenient strategy resulted in a substantially higher rate of false positives.

The adoption of the strict interpretation strategy for M staging aims to minimize false positives without compromising the need to maintain a low rate of false negatives and is grounded in specific clinical imperatives. From a clinical perspective, the high false positive rate observed under the lenient strategy is unacceptable. For instance, benign findings commonly described in reports, such as cysts in the kidneys or liver in older patients, could be misclassified as M1 (Stage IV) metastases under lenient rules. This could immediately induce unnecessary patient anxiety and trigger a cascade of inappropriate clinical actions. More critically, a false-positive call at this diagnostic juncture can fundamentally misdirect the primary treatment pathway, potentially steering a patient eligible for curative-intent therapy (eg, surgery or definitive chemoradiation) toward an erroneous palliative systemic approach. Furthermore, it compels a series of costly, invasive, and burdensome confirmatory investigations (eg, advanced imaging, biopsies) without clinical merit.

This strategic choice is also aligned with the linguistic reality of Chinese radiology reports, which often use broad or cautious terminology. The strict interpretation acts as a necessary filter against this inherent ambiguity, serving as a “localized calibration” technique in this context. Its core function is to correct the tendency of general-purpose LLMs to over-interpret ambiguous expressions in the absence of domain-specific fine-tuning, an enhancement generally applicable within Chinese clinical practice. It is crucial to acknowledge that this calibrated approach, while effective, is not a universal reasoning logic. It could theoretically introduce a risk of false negatives in scenarios where conservatively phrased reports use descriptive terminology. This technique complements the AJCC-based rule supplements we implemented.

Therefore, while sensitivity remains important, our strategy is designed to ensure that the model’s outputs provide highly credible and actionable decision support. A crucial caveat is that using such a strategy requires preliminary testing and calibration based on the target language or dataset. The decision to apply a strict strategy (to correct LLM over-interpretation) versus a lenient one (to correct LLM under-interpretation) should be guided by initial validation, as the latter scenario is also possible in other linguistic contexts or specific datasets. The strict strategy for M staging in this study thus represents a calibrated balance, optimizing the model for reliable integration into high-stakes clinical workflows.

This evaluation examines the operational efficiency and deployment feasibility of our optimized GLM-4-Air model. Through domain-specific optimization, the 32B-parameter GLM-4-Air model can achieve high staging accuracy while demonstrating a practical efficiency profile. A key practical implication of this study is the demonstration that a specialized, moderately sized language model can perform complex TNM staging with high accuracy using cost-effective, consumer-grade hardware. Specifically, our local deployment benchmark achieved low per-component latencies on 4 RTX 4090 GPUs, confirming that the model’s performance profile is acceptable and suitable for integration into real-world clinical workflows without requiring prohibitive computational infrastructure. While this study demonstrates the cost-effectiveness of the 32B-parameter GLM-4-Air model within our hybrid framework, we acknowledge that the broader claim of hardware accessibility is not exclusive to this specific model. Other competitively performant “lightweight” models (eg, GPT-4o-mini, Llama-3-70B), which share a similar order-of-magnitude parameter scale, could potentially offer comparable reductions in computational resource demand and deployment cost. While the assessment of other lightweight models was beyond the scope of this study, a systematic comparison among them represents a direction for future research.

Furthermore, the Cost-effective Evaluation suggests a trend of complementary strengths between large general-purpose models and smaller, domain-adapted models in the medical TNM staging task. On one hand, GPT-4o, a general-purpose LLM, demonstrates high conciseness in tasks with shorter reasoning chains, such as N staging. Its powerful linguistic generalization capabilities enable it to rapidly extract key information. On the other hand, for complex tasks like T staging, which require in-depth parsing of anatomical relationships and multimodal clinical descriptions, GLM-4-Air, enhanced by targeted medical SFT, establishes more stable and interpretable reasoning pathways. While large-parameter models excel in cross-domain generalization and flexible response for open scenarios, smaller-parameter models like GLM-4-Air are potentially better suited for high-reliability specialized tasks. They can potentially deliver lower overall latency and higher deployment cost-effectiveness while ensuring reasoning accuracy and completeness, with a proper training framework.

While the performance of our model is robust on the held-out test sets, we acknowledge that the generalizability of any AI model trained on a dataset of this scale (292 fine-tuning cases) warrants careful discussion. The primary consideration is its performance on data from wider distributions, such as more institutions with variations in reporting styles. Furthermore, our dataset was sourced exclusively from the proprietary medical record-management platform reliant on user-initiated uploads. This data acquisition method may introduce potential selection biases: the user population likely possesses higher digital literacy and health awareness, potentially limiting representativeness in terms of age or socioeconomic status; the act of voluntary uploading may also over-represent patients with more complex conditions, unresolved diagnostic concerns, a willingness to seek second opinions, or experiences with multi-institutional care in the dataset. These specific characteristics suggest that the patient profile in our study may differ from that of a general hospital cohort, which would typically include a higher proportion of patients with lower digital literacy (often associated with advanced age or lower socioeconomic status) and earlier-stage disease. The model’s staging accuracy for these specific subgroups requires further validation. Additionally, as uploaded data typically consists of single or nonconsecutive medical records, it lacks the temporal continuity and completeness in documenting disease progression compared with the structured, longitudinal records within a hospital information system. This difference in information presentation implies that the model may face challenges in inference and integration when processing richer contextual reports. Consequently, while the model demonstrates robust performance within the represented data distribution, its generalizability to passively collected, consecutive hospital-wide cohorts could be further explored in future work. Caution is advised when interpreting performance metrics for subclasses with a low number of cases. The high recall rates observed in some specific subgroups primarily reflect performance on that particular test subset and are subject to statistical variability due to limited sample size. It is also crucial to emphasize that an awareness of these potential biases directly informed our study design. To proactively enhance model robustness and generalizability within this context, we integrated 2 core methodological choices: the curation of rigorously balanced test sets (Figure 2) to ensure reliable evaluation, and the use of a heterogeneous data regimen during SFT to significantly reduce the risk of overfitting.

Another topic we would like to discuss is language generalizability. In this study, the model was exclusively trained and validated on medical reports in Chinese, and we acknowledge its potential limitations in generalizing across languages and health care systems. The unique expression logic, terminology, and writing structure inherent to Chinese clinical texts may restrict the model’s direct applicability to other languages, such as English, or different medical environments. A noteworthy point for discussion is the systematic difference in linguistic precision between Chinese and English medical language. Chinese reports often use broad-meaning or imprecise terms; for instance, the word “侵犯” can encompass various pathological states from adhesion to infiltration, with its specific interpretation highly dependent on context and the radiologist’s personal expression habits. In contrast, the mature system of English medical literature, shaped by long-term standardization, tends to use more discriminative terminology, such as “abutment,” “invasion,” and “encasement” to describe different degrees of involvement. These terms often have more direct correspondence with definitions in the AJCC staging guidelines. This linguistic characteristic suggests that while our current model is specialized for the Chinese context, adapting its core competency to the potentially more standardized terminology of English environments might present relatively manageable semantic understanding challenges, and the model may need recalibration for health care systems with more precise reporting standards.

Building upon the generalizable foundation established in this work, our future research will focus on extending the model’s robustness to multi-institutional environments. The strategies outlined below are presented as logical and powerful extensions of this study:

In terms of language generalizability, to systematically enable effective deployment in broader scenarios, we propose the following three actionable technical pathways:

In summary, by implementing this technical roadmap centered on global collaboration, modular design, and cross-lingual terminology alignment, we can systematically enhance the model’s generalizability, ultimately developing it into a practical tool for global, multi-lingual clinical environments.

Given the inherent complexity and subjective nature of cancer staging, we recognize that a prospective comparison between the model’s performance and the independent interpretations of multiple clinicians on new, unseen cases would provide a more ecologically valid assessment of its clinical utility. Such a study design, where the model’s outputs and individual expert annotations are collected in parallel without a pre-established consensus, would more accurately simulate real-world deployment scenarios. This approach would not only allow for a robust benchmarking against the natural variation inherent in clinical practice but also help identify specific contexts where the model’s reasoning aligns with or diverges from human experts. We plan to implement this in future research, potentially involving a larger and more diverse panel of annotators from multiple institutions. This direction is particularly significant for validating AI assistants in complex, subjective tasks like cancer staging, and its methodology could be extended to other high-stakes clinical decision support applications where interexpert variability is a key consideration.

In conclusion, this study establishes a robust and efficient framework for automating TNM staging in NSCLC by leveraging the GLM-4-Air model enhanced through advanced prompt engineering and selectively applies SFT to reasoning-heavy tasks (T/N) while leveraging the baseline model for retrieval-centric tasks (M). This task-dependent optimization strategy highlights that the value of fine-tuning is not universal but varies with the nature of the clinical subtask, which is a key insight for future clinical AI development. The finalized model demonstrated superior performance over a leading commercial model, GPT-4o, achieving higher accuracy, particularly in complex staging decisions, while simultaneously reducing the incidence of clinically critical errors. Its design, which emphasizes reasoning transparency and adherence to clinical guidelines, ensures reliability. The exceptional cost efficiency of the solution further underscores its viability for scalable deployment.

This work also provides a validated pathway toward standardizing and augmenting cancer staging processes, with the immediate potential to improve consistency in treatment planning and to expand access to expert-level staging support in diverse health care settings.