Abstract

Objectives

This study aimed to evaluate the concordance between artificial intelligence (AI)-generated treatment recommendations and real-world multidisciplinary tumor board (MDT) decisions in postoperative breast cancer management.

Material and Methods

This retrospective validation study included 30 female patients with breast cancer who were previously discussed in a MDT. Demographic, clinical, pathological, and molecular data were retrospectively collected from institutional records and anonymized before analysis. Structured clinical scenarios were created for each patient and entered into an AI-based decision-support system, developed using ChatGPT version 5.2, in accordance with current ESMO, NCCN, and ASCO guidelines. AI-generated recommendations for systemic therapy, radiotherapy, and genomic testing were compared with corresponding MDT decisions. Concordance was assessed for each treatment domain.

Results

The mean age was 52±13.1 years, and the median interval from pathology reporting to MDT discussion was 7 days (interquartile range, 4-11). Neoadjuvant therapy was administered to 11 patients (36.7%), and 4 (36.6%) achieved a pathological complete response. Breast-conserving surgery was performed in 22 patients (73.3%). Concordance between AI and MDT varied by treatment modality. Agreement was highest for radiotherapy (90.0%), followed by anti-human epidermal growth factor receptor 2 (HER2) therapy (73.3%) and endocrine therapy (70.0%). Lower concordance rates were observed for genomic testing (53.3%) and chemotherapy (33.3%). Chemotherapy showed the highest discordance rate (66.7%).

Conclusion

AI demonstrated high concordance with MDT decisions in treatment domains guided by standardized criteria, particularly radiotherapy, endocrine therapy, and anti-HER2 therapy. Lower agreement was observed for chemotherapy and genomic testing, suggesting that these decisions require more individualized clinical judgment. AI may serve as a useful decision-support tool within multidisciplinary breast cancer care.

Introduction

Breast cancer is the most common cancer among women (1). Over the past decades, advances in screening, surgical techniques, and oncological therapies have significantly improved survival outcomes. However, these advances have also made treatment strategies more complex, requiring a more individualized and multidisciplinary approach to patient care. This multidisciplinary approach requires a team built on the collaboration of breast surgeons, medical oncologists, radiation oncologists, pathologists, and radiologists (2).

Decision-making in cancer care relies on both structured frameworks from international guidelines and individualized patient evaluation. Multidisciplinary tumor boards (MDTs) are central to combining these elements for clinicians and patients (3, 4). While MDTs are necessary, their meetings also have practical limitations. Because coordinating multiple specialists means meetings are typically held at set intervals, this can limit flexibility and possibly cause delays. Furthermore, the increasing patient load and the complexity of treatment algorithms add to the burden on multidisciplinary teams (4).

The use of artificial intelligence (AI) and, more recently large language models (LLMs), has rapidly expanded in healthcare (5). These systems can interpret clinical data and guideline-based recommendations. They can also generate structured treatment suggestions when given appropriate inputs (6). In oncology, decision-making often relies on integrating multiple variables. AI-based tools can support clinicians with consistent, guideline-informed recommendations (7).

Given these developments, new AI and LLM technologies can be integrated into daily practice to support clinicians’ decision-making. In this study, we aim to evaluate the concordance between an AI-based decision-support system and real-world MDTs in postoperative breast cancer treatment recommendations.

Materials and Methods

This retrospective validation study evaluated the concordance between an AI-based decision support system and real-world MDT decisions in postoperative breast cancer management.

The AI-based decision support system was developed using ChatGPT, version 5.2. It followed current international clinical guidelines, including those of ESMO, NCCN, and ASCO. The tool is guideline-informed, not an autonomous decision-making system. For each patient, it generated structured recommendations for systemic therapy, radiotherapy, and genomic testing. The system operated in a locked configuration, with no changes to the model version, architecture, or prompt strategy during the study.

Patients evaluated by a multidisciplinary breast cancer tumor board were identified retrospectively. Demographic, clinical, pathological, and molecular data were extracted from institutional records. All data were anonymized before analysis. For each patient, structured clinical scenarios were created. These included age, tumor stage, family history, BRCA mutation, histopathological characteristics, hormone receptor status, human epidermal growth factor receptor 2 (HER2) status, Ki-67 index, molecular subtype, breast-conserving surgery status, axillary management, and postoperative tumor-node-metastasis staging. No personally identifiable information was provided to the AI system.

Anonymized patient scenarios were entered into the AI-based decision support system. The generated recommendations were recorded under predefined categories: systemic therapy strategy, radiotherapy indication, and genomic testing recommendation. Outputs were then compared with the corresponding real-world MDT decisions for the same patients. The level of agreement was classified as concordant or discordant for each parameter.

The AI system was not used in real-time clinical decision-making and did not influence patient management. All AI-generated outputs were evaluated solely for validation purposes in a shadow mode setting.

Statistical Analysis

All statistical analyses were performed using Jamovi software, version 2.6. Categorical variables were presented as frequencies and percentages. Continuous variables were summarized as mean ± standard deviation or median with interquartile range (IQR), depending on data distribution.

The primary endpoint of the study was the concordance between AI-generated recommendations and MDT decisions. Concordance rates were calculated as the proportion of cases in which AI recommendations and MDT decisions agreed.

All data were anonymized and handled in accordance with data protection regulations. The study followed ethical standards for retrospective observational research. The Acıbadem Mehmet Ali Aydınlar University Institutional Review Board approved the study (approval number: 2026/02, date: 22.02.2026).

Results

A total of 30 female patients with breast cancer were included in the study. The mean age of the cohort was 52±13.1 years, and the median interval from pathology reporting to MDT discussion was 7 days (IQR, 4-11). Nineteen patients (63.3%) were postmenopausal, while 11 (36.7%) were premenopausal (Table 1).

Most tumors were invasive carcinomas (n=27, 90.0%). Three patients (10.0%) had in situ disease. Among invasive tumors, cT2 was the largest subgroup (59.3%). Seventeen patients (63%) had node-negative disease, and 10 (37%) patients had nodal involvement. All tumors were ductal carcinomas. Thirteen patients (44.8%) had grade 2 tumors; 16 (55.2%) had grade 3 tumors. Estrogen receptor positivity was observed in 20 patients (66.7%), and progesterone receptor positivity was observed in 14 patients (46.7%). HER2 status was negative in 16 patients (57.1%), positive in 6 (21.4%), and inconclusive in 6 (21.4%). The median Ki-67 index was 17%.

Neoadjuvant therapy had been administered to 11 patients (36.7%), whereas the remaining 19 patients (63.3%) underwent upfront surgery. Among those who received neoadjuvant treatment, 4 patients (36.6%) achieved a pathological complete response. Breast-conserving surgery was performed on 22 patients (73.3%), while 8 patients (26.7%) underwent mastectomy. Axillary staging was performed by sentinel lymph node biopsy in 28 patients (93.3%), whereas axillary dissection was required in 2 patients (6.7%).

Concordance between AI-generated recommendations and MDT decisions varied according to treatment modality. The highest concordance was observed for radiotherapy, with 27 cases (90%) in agreement. Concordance was also relatively high for anti-HER2 therapy and endocrine therapy, with agreement rates of 73.3% and 70.0%, respectively. In contrast, concordance was lower for genomic testing and chemotherapy (53.3% and 33.3%, respectively) (Table 2).

Chemotherapy had the highest discordance rate: 20 cases (66.7%) showed disagreement. Among these discordant cases, the AI recommended chemotherapy in 6 cases (30%) while the MDT did not; conversely, the MDT recommended chemotherapy in 4 cases (20.0%) while the AI suggested genomic testing. In 5 additional cases (25%), the AI recommended genomic testing when the MDT did not, and in 5 cases (25%), the MDT recommended genomic testing when the AI did not.

For endocrine therapy, discordance was observed in 9 cases (30%). In 5 of these cases (55.6%), the MDT recommended endocrine therapy, whereas the AI did not; conversely, in 4 cases (44.4%), the AI recommended endocrine therapy, whereas the MDT did not. Discordance in anti-HER2 therapy was observed in 8 cases (26.7%), including 3 cases (37.5%) in which the MDT recommended anti-HER2 therapy but the AI did not, and 5 cases (62.5%) in which the AI recommended anti-HER2 therapy, but the MDT did not.

Radiotherapy had the lowest discordance rate, with 3 cases (10%) showing disagreement. Genomic testing showed discordance in 14 cases (46.7%). Overall, agreement was higher for radiotherapy, endocrine therapy, and anti-HER2 therapy, whereas it was lower for chemotherapy and genomic testing.

Discussion

In this study, we evaluated the concordance between AI-generated treatment recommendations and decisions made by an experienced MDT in patients with breast cancer. Our findings indicate that agreement between the AI system and the MDT varied across treatment modalities, highlighting several key factors.

Radiotherapy demonstrated the highest concordance rate (90%). This finding is not unexpected, as radiotherapy decisions in breast cancer are largely based on standardized treatment algorithms incorporating surgical procedure, tumor stage, and nodal status (8). In such relatively structured and guideline-driven settings, AI systems may be able to reproduce clinical decision pathways with a high degree of consistency.

Similarly, concordance was relatively high for endocrine therapy (70.0%) and anti-HER2 therapy (73.3%). These treatment decisions are primarily determined by receptor status and well-established therapeutic algorithms, which may explain the higher level of agreement between AI-generated recommendations and MDT decisions (9).

In contrast, chemotherapy demonstrated the lowest concordance rate (33.3%). Most discordant cases arose when the AI recommended chemotherapy, whereas the MDT favored genomic testing or clinical observation. This finding suggests that AI systems may rely more heavily on guideline-based risk stratification, while clinicians often incorporate additional contextual and patient-specific factors when determining the indication for chemotherapy (10). Variables such as patient age, tumor grade, proliferative activity, comorbidities, and clinical judgment may affect treatment selection beyond what algorithm-based systems can fully capture.

A similar pattern was observed for genomic testing, with moderate concordance between AI and MDT (53.3%). The AI system recommended genomic testing more frequently than the MDT. This may reflect the tendency of AI models to favor molecular testing in borderline cases where guidelines support the use of genomic assays as adjunctive decision-making tools, whereas clinicians may choose to omit such testing when the result is unlikely to alter the treatment plan (10).

Taken together, these findings stress an important distinction between treatment decisions that are predominantly guideline-driven and those that require more individualized clinical interpretation. AI-based systems appear to perform well in structured decision-making contexts, although they show lower agreement in more detailed scenarios that depend on physician experience and broader clinical judgment.

Study Limitations

This study has several limitations. First, the sample size was relatively small, which may limit the generalizability of the findings. Second, the analysis was conducted within a single-institution MDT setting, and treatment decisions may vary across institutions according to local practice patterns, clinical expertise, and patient characteristics. In addition, the AI system was evaluated retrospectively in shadow mode and, therefore, was not assessed within a real-time clinical workflow.

Regardless of these limitations, the present study provides preliminary insight into the interaction between AI-based decision support tools and multidisciplinary clinical decision-making in breast cancer care. Rather than replacing clinicians, such systems may serve as complementary tools that support clinical reasoning and facilitate guideline-informed treatment planning.

Conclusion

AI demonstrated high concordance with MDT decisions across treatment modalities that were largely guided by standardized criteria, including radiotherapy, endocrine therapy, and anti-HER2 therapy. By contrast, lower agreement was observed in modalities requiring more individualized judgment, particularly in chemotherapy and genomic testing. These findings suggest that AI may be most useful as a decision-support tool integrated into multidisciplinary clinical practice rather than as a substitute for expert clinical evaluation.