Recurrence detection in patients with triple-negative breast cancer following the current standard of care: a microsimulation model

Abstract

Objectives

Breast cancer is the most common cancer among women in the USA, accounting for approximately 31% of new cases and 15% of cancer-related deaths in females. Triple-negative breast cancer (TNBC) is the most lethal subtype with 29 months median overall survival. Despite curative-intent surgery and systemic adjuvant therapy, many patients remain at risk for recurrence. We developed a comprehensive model for patients with operable stage II and III TNBC, adhering to the current standard-of-care clinical follow-up recommendations, to evaluate the effectiveness of these recommendations in detecting recurrence among patients with newly diagnosed TNBC.

Design

An individual-level state transition (microsimulation) model was built with a lifetime horizon using monthly cycles. The model integrated treatment and surveillance recommendations per current guidelines.

Setting

The currently recommended treatment and surveillance modalities by the National Comprehensive Cancer Network guidelines for surveilling patients with TNBC after diagnosis.

Participants

Female patients aged 49 years who were newly diagnosed with operable stage II or III TNBC.

Outcome measures

Cumulative recurrences, detected recurrences, detection rate, overall survival and recurrence-free survival in a 5-year horizon, as well as average life expectancy, were the outcome measures used.

Results

Our model estimated a 5-year overall survival rate of 81.4%±0.1% (SE) for patients with stage II or III TNBC, with a recurrence-free survival rate of 78.3%±0.1% (SE). Over the same period, 7.4%±0.1% (SE) of patients with TNBC were projected to experience recurrences (68% distant recurrence, 32% locoregional recurrence), with a detection rate of only 59.3%±0.2% (SE) among these recurrences. Our model estimated that 399 mammograms would be required to detect one locoregional recurrence and 279 other imaging tests would be needed to identify one distant recurrence.

Conclusions

Our clinically validated model sheds light on the outcomes for patients with TNBC, revealing the limitations in recurrence detection rates associated with current guidelines. These findings underscore the need for improved strategies during the surveillance period. As novel testing methods become available, this model can be instrumental in exploring how these innovations may improve patient outcomes.

STRENGTHS AND LIMITATIONS OF THIS STUDY.

• Our study integrates natural triple-negative breast cancer disease progression with real-world surveillance practices and treatment guidelines, providing a holistic assessment of current management strategies.

• The model’s design enables detailed evaluation of surveillance performance, resource utilisation and modifiable factors such as adherence, offering a robust platform to inform future research, optimise follow-up protocols and support evidence-based updates to clinical guidelines.

• Extrapolation of survival data beyond 5 years and assumptions regarding outcomes for undetected recurrences introduce some uncertainty, reflecting limitations in available real-world evidence.

• Due to limited real-world data, the model assumed poorer survival among patients with undetected recurrences than among those with detected recurrences, operationalised by applying an increased mortality risk to patients, reflecting the impact of delayed diagnosis until recurrence detection.

Introduction

Breast cancer is the most common cancer among females, accounting for approximately 31% of new cases in the USA.¹ It is responsible for 15% of cancer-related deaths in females, making it the second deadliest cancer in women in the US.¹

Breast cancer is classified based on the presence of oestrogen receptors (ER), progesterone receptors (PR) and human epidermal growth factor receptor 2 (HER2),^{2 3} with treatment approach and effectiveness varying significantly based on these subtypes. Triple-negative breast cancer (TNBC) is characterised by the absence of ER, PR and HER2 receptors,^{2 3} and is the most lethal form of breast cancer;⁴ hence, managing TNBC presents a significant clinical challenge. Evidence also suggests that current practices may lack targeted therapeutic strategies, leading to suboptimal treatment and survival outcomes.⁵

Across all subtypes, patients who undergo treatment for primary breast cancer may still experience disease recurrence. Approximately 20.5% of patients who receive treatment in early-stage disease can experience recurrence within 10 years, and 35% of these recurrent cases belong to the TNBC subtype.⁶ Patients who complete primary treatment for breast cancer undergo clinical follow-up to monitor any signs of disease recurrence, which can occur locally, regionally (ie, locoregional) or in distant organs. During this surveillance period, routine follow-up can include mammograms and physical examinations, with additional imaging performed if suspicious findings suggest potential recurrence.⁷

Despite the established protocols for the postdefinitive treatment surveillance period, the implications of current standard-of-care (SoC) guidelines remain uncertain. Comprehensive data on the long-term clinical effectiveness of these strategies for patients with breast cancer are lacking, particularly for those with TNBC, who face a higher risk of recurrence and poorer prognosis.⁴ The absence of tailored clinical follow-up guidelines may result in either overtreatment, leading to unnecessary anxiety and increased healthcare costs, or undertreatment, potentially delaying recurrence detection and worsening patient outcomes. Therefore, further research is needed to evaluate and refine practices during the surveillance period for patients with TNBC to ensure timely recurrence detection and improved survival outcomes.

Our objective was to evaluate the implications of current SoC clinical follow-up recommendations for patients with newly diagnosed operable stage II and III TNBC by developing a comprehensive microsimulation model that simulates the natural disease progression, treatment and surveillance period of these patients.

Methods

Model overview

We built an individual-level state transition model, that is, microsimulation model, to follow the natural disease history of patients with newly diagnosed operable stage II and III TNBC. The model was developed in Python version 3.12. The model incorporated treatment and surveillance recommendations along with the natural disease progression in patients with TNBC. It encompassed three main health states, with substates included to capture different patient and disease characteristics: (1) recurrence-free (with two substates: ‘presurgery’ and ‘postsurgery’), (2) recurrence (with two substates: ‘locoregional recurrence’ and ‘distant recurrence’) and (3) death (figure 1).

Figure 1. Model schematic for the natural disease progression and surveillance period for triple-negative breast cancer (TNBC). *76% of patients receive radiation therapy in postadjuvant settings. BC, breast cancer.

Overall, the model considered time-dependent transitions based on best-fitted parametric survival distributions to available Kaplan-Meier curves (online supplemental figures A1–A3 in online supplemental appendix A1), with individual patient data reconstructed using Guyot’s method and transition probabilities derived from the fitted survival functions (online supplemental table A1 in online supplemental appendix A1). The model used a lifetime horizon with monthly cycle length. It was designed to have the flexibility to generate and compare different modalities in the surveillance period and treatment strategies.

Baseline population

The baseline population consisted of female patients aged 49 years who were newly diagnosed with operable stage II or III primary TNBC, reflecting the patient population in the KEYNOTE-522 study, the pivotal phase III perioperative pembrolizumab trial in Stage II-III TNBC.⁸ In the overall study population, 76.7% had stage II, while 23.3% had stage III.⁸ Online supplemental table A2 in online supplemental appendix A2 presents patients’ characteristics in the model.

Natural disease progression

All patients started in a recurrence-free state and received neoadjuvant therapy (NAT), as detailed in the primary treatment for TNBC section. During NAT, some patients could experience disease progression preventing them from undergoing definitive surgery and were considered to have progressed TNBC. Their mortality rates were assumed to be higher than those of patients with non-progressive TNBC, based on findings from the literature.^{8 9} The mortality rate of patients with non-progressive TNBC during NAT was assumed to be lower than that of the overall TNBC population, which included both patients with non-progressive TNBC (early stage) and patients with progressive TNBC (advanced stage). To account for this, an HR of 0.507 was applied to the treatment line for the overall population, capturing the lower mortality rate of patients with non-progressive TNBC compared with patients with progressive TNBC during NAT (online supplemental appendix A3).^{8 9}

At any point in the model, patients could die due to background mortality or breast cancer-specific mortality, that is, differentiated by the underlying disease stage. Stage-specific mortality was adjusted using an HR of 2.18 to inform increased mortality in stage III patients compared with stage II patients, calculated from the reported overall survival (OS) outcomes by stage from Luo et al.¹⁰ The background mortality probabilities were obtained from the US Life Tables.¹¹ Additionally, while currently there is no evidence on the impact of undetected recurrences on survival, in this study, we assumed that, compared with those with detected recurrences, patients with undetected recurrences would result in lower survival outcomes (ie, increased mortality) due to the potential impact of delayed treatment.^{12 13} At the same time, we conducted sensitivity analysis that included a lower bound of the HR to be near 1 to test the possibility that potential delays in treatment may not have any significant impact on survival.

The probabilities of disease progression before and after surgery were calculated using the event-free survival curve from the study by Schmid et al.⁸ When patients experienced recurrence, 31.8% of recurrences were assumed to be locoregional, while the remaining 68.2% were assumed to be distant recurrences.⁸ The sites of distant recurrences were assumed to be the lung, liver, bone and brain (online supplemental appendix A4).¹⁴

Additionally, to reflect the natural history of the disease, including molecular biology of metastasis, the model incorporated circulating tumour DNA (ctDNA) as a measure of patient-specific molecular residual disease indicator/biomarker. Because ctDNA has been associated with a significantly increased risk of recurrence,¹⁵ disease progression was stratified based on the presence of ctDNA, with separate survival curves for ctDNA+ and ctDNA− patients at each stage. This stratification was based on the prevalence of ctDNA+ patients and the HR of progression for ctDNA+ patients compared with ctDNA− patients at different time points, as detailed in online supplemental table A2.

Treatment

Primary treatment for TNBC

Regardless of the cancer stage, all patients with TNBC were assumed to receive a comprehensive primary treatment regimen, which included neoadjuvant chemotherapy with immunotherapy, surgery and adjuvant immunotherapy. The neoadjuvant treatment consisted of two phases: (1) phase I (3 months): patients received paclitaxel and carboplatin (PC) with pembrolizumab and (2) phase II (3 months): following the completion of the first phase, patients continued with pembrolizumab but switched to cyclophosphamide and doxorubicin instead of PC.

After completing NAT, all eligible patients underwent curative surgery, although specific surgery types (such as mastectomy or lumpectomy) were not explicitly modelled in the current model, rather, they are grouped into the same category of ‘surgery’.

For the adjuvant treatment phase, pembrolizumab was modelled as the primary agent. No additional chemotherapy drugs were included in the adjuvant setting, as the neoadjuvant treatment was assumed to encompass a comprehensive chemotherapy regimen.

Postadjuvant therapy, 76% of patients were assumed to receive radiation therapy, depending on the patient’s status and the doctor’s preference to mimic current clinical practice.⁸ However, disease progression was not stratified based on the uptake of radiation therapy by the patients. Figure 1 depicts the treatment approach in the model, representing the SoC.

Treatment for recurrent TNBC

In the model, when a recurrence was detected, patients would receive treatment to improve their survival. To estimate the survival of patients with locoregional recurrence, data from the study by Allen et al were used on the OS of patients with TNBC who received a variety of treatments, rather than a specific regimen.¹⁶ For distant recurrences, although there is no standard treatment, pembrolizumab combined with a chemotherapy drug has become a preferred first regimen for advanced disease in current clinical practice.⁷ To inform this estimate, we used the median OS of patients in the pembrolizumab-chemotherapy arm from Cortes et al to estimate the survival of patients with distant recurrence undergoing treatment.¹⁴

Surveillance

In our model, patients who underwent surgery were subsequently monitored using clinical protocols with predetermined adherence rates for each modality. The clinical follow-up algorithms employed in the model were informed by the National Comprehensive Cancer Network (NCCN) guidelines and validated by clinical experts.⁷

Surveillance for locoregional recurrence

The NCCN guidelines recommend that patients undergo a screening mammogram every 12 months as part of routine surveillance after surgery.⁷ If a screening mammogram revealed a suspicious finding, the patient would undergo a diagnostic mammogram, followed by a core needle biopsy as the confirmatory test in the same cycle (online supplemental figures A4 and A5 in online supplemental appendix A5).

Surveillance for distant recurrence

According to clinical guidelines, routine screening for distant recurrence, other than follow-up physical examinations, is not recommended unless there are suspicious signs or symptoms present.⁷ Schumacher et al reported that not all detected distant metastases are related to exhibited symptoms; approximately 25% of recurrences are incidentally detected or identified through non-cancer-specific screenings.¹⁷ To estimate the incidental detection rate for patients who did not show symptoms, that is, asymptomatic patients, two main assumptions were made:

• Distant recurrent patients had a 75% probability of presenting with symptomatic disease, meaning 25% were asymptomatic, in line with the findings reported by Schumacher et al.¹⁷

• All symptomatic patients were assumed to undergo screening for metastatic disease, consistent with standard clinical practice.

This led to a calibration estimate that 2.48% of all distant recurrences were detected incidentally in each month.

The screening methods for detecting recurrence varied depending on the site of metastasis. Online supplemental table A3 presents the distribution of metastasis sites and the performance characteristics of associated screening methods. In the model, we used a weighted average for sensitivity and specificity based on the distribution of metastasis sites. A biopsy with perfect sensitivity and specificity was included as the confirmatory test after an abnormal finding. All parameters during the surveillance period are presented in online supplemental table A2.

Model validation

The model was validated according to The Professional Society for Health Economics and Outcomes Research (ISPOR) model validation guidelines.¹⁸ The validation process was categorised into face validity, internal validity, external validity and cross validation. The details of the validation procedures can be found in online supplemental appendix A6.

To ensure the face validity of the model, each model component was reviewed in detail and feedback from clinical experts was incorporated to align the model with the most up-to-date scientific evidence and clinical guidelines.

For internal validity, stress testing was performed by inputting extreme values to evaluate the model’s behaviour under various conditions. Model outputs were compared with the input data to ensure that the model produced coherent and logical results. Mortality and recurrence data for patients with non-recurrent TNBC were sourced from Schmid et al⁸ and fitted using survival analysis techniques. Scenario analyses were conducted to test extreme cases as part of internal validation.

External validation was performed by comparing model outcomes with real-world data from Surveillance, Epidemiology, and End Results (SEER) database. Among the real-world studies reporting survival outcomes of patients with TNBC (online supplemental table A4), the study by He et al,¹⁹ with 31 941 patients with TNBC, was selected for external validation due to its stage-specific survival data over a 5-year period (online supplemental table A5). At the same time, He et al¹⁹ included cases from the 2010–2015 interval, before pembrolizumab was made available for patients with TNBC, as pembrolizumab was approved by the Food and Drug Administration in 2021 for the treatment of TNBC.²⁰ To align the baseline population of the model with the external validation target, the model’s baseline population treatment pathway was informed by the ‘placebo-chemotherapy’ arm from the study by Schmid et al (online supplemental appendix A6).⁸

To further assess model performance, cross-validation was conducted using a prior modelling study by Huang et al,²¹ which focused on the natural history of TNBC. Given the lack of published studies specifically addressing the TNBC surveillance period, the model’s OS and recurrence-free survival (RFS) outcomes were compared with those from Huang et al.²¹

Base case analysis

In the base case analysis, clinical outcomes were evaluated following the current SoC for surveillance period and treatment. Patient life-years, mortality rates and recurrence rates were estimated for each model cycle and aggregated over a lifetime horizon.

Sensitivity analysis

One-way sensitivity analyses were conducted to identify the key drivers of the model outcomes. This analysis involved varying the parameters within their available measures of uncertainty, such as SEs, or using reasonable ranges for each input.

Results

Model validation

As a result of internal validation efforts, the model’s OS and RFS curves closely matched the observed curves from the study by Schmid et al,⁸ confirming the model’s internal validity (online supplemental figures A6 and A7).

To conduct scenario analyses for internal validation, recurrence and death rates were set to 0% and 100%, the model’s behaviour aligned with expectations, demonstrating the robustness of the event-free survival outcomes (online supplemental figure A8). Similarly, when mortality rates were adjusted to 0% and 100%, the OS curves were aligned with the anticipated outcomes (online supplemental figure A9). Additional scenarios, such as perfect versus no clinical follow-up, were modelled, and the survival outcomes adhered to expectations, with perfect surveillance yielding better survival outcomes compared with baseline and no clinical follow-up (online supplemental figure A10).

External validation results showed that both stage-specific and OS curves matched reasonably well with those of SEER database, that is, 75.2% vs 75.0% for stage II patients, 51.3% vs 47% for stage III patients and 69.6% vs 69.0% for overall population for model-predicted versus SEER, respectively (figure 2). The small deviation between model-generated results versus SEER database for stage III patients may be due to the assumption that all patients get neoadjuvant chemotherapy in the model, which may not be the case for the entire population in the SEER database. The other possible factor is the time frame of SEER database (ie, 2010–2014) being earlier than Schmid et al, which could be a result of improved medical interventions.^{8 19}

Figure 2. External validation outcomes: model versus SEER database.

Cross-validation efforts showed that, when compared with Huang et al,²¹ OS and RFS outcomes demonstrated consistency across both studies (online supplemental figure A11). This comparison underscores the model’s reliability and supports its generalisability across different TNBC populations.

Base case analysis

As part of the base case analysis, the model was run over a lifetime horizon for patients with newly diagnosed TNBC to estimate OS, RFS, predicted recurrences (ie, all including both detected and undetected cases within 5 years), detected recurrences (ie, those identified through standard clinical practice), average life expectancy and the number of tests needed to detect one recurrence.

Survival outcomes for patients with TNBC

Following the standard of care, the model predicted OS for patients with TNBC to be 81.4%±0.1% (SE) at 5 years, 67.7%±0.1% (SE) at 10 years and 58.6%±0.1% (SE) at 15 years. RFS was projected to be 78.3%±0.1% (SE) at 5 years, 65.4%±0.1% (SE) at 10 years and 57.0%±0.1% (SE) at 15 years. The average life expectancy over a lifetime was estimated at 22.3 years.

For stage II patients, the model predicted a 5-year RFS of 82.2%, with 70.8% at 10 years and 63.0% at 15 years. The 5-year OS rate was 85.3%, with 73.3% at 10 years and 64.8% at 15 years. The average postdiagnosis life expectancy for stage II patients was 24.4 years.

For stage III patients, the 5-year RFS was 65.5%, with 47.8% at 10 years and 37.3% at 15 years. The 5-year OS rate was 68.4%, with 49.2% at 10 years and 38.1% at 15 years. The average postdiagnosis life expectancy for stage III patients was 15.5 years.

Risk of recurrence and detection rates in patients with TNBC

The model estimated that in 5 years, 7.4%±0.1% (SE) of patients with TNBC would have recurrences and only 4.4%±0.1% (SE) of the patients would be detected by the current modalities used in the surveillance period, corresponding to an overall detection rate of 59.3%±0.2% (SE) (figure 3).

Figure 3. Cumulative clinical and detected recurrences among all patients. Predicted recurrence refers to the proportion of individuals who are projected to experience recurrence among the entire baseline population (ie, patients with newly diagnosed stage II and III triple-negative breast cancer (TNBC)). Detected recurrences refer to the predicted recurrences identified via surveillance, while undetected recurrences are those missed by surveillance. Predicted recurrence and detected recurrences are shown cumulatively. The goal is to showcase the gap between detected and undetected recurrences among all patients with TNBC over a 5-year horizon. These rates do not reflect the TNBC recurrence rates outside of this context. The stepwise pattern in the cumulative detection rates is driven by annual mammography, which results in periodic accumulation and detection of locoregional recurrences occurring within each 12-month interval.

Our model estimated that over 5 years, 2.3% of patients with TNBC would experience locoregional recurrence, while 5.1% were projected to experience distant recurrence (online supplemental figure A12a,b in online supplemental appendix A7). In stage-specific subpopulations, the model predicted that 6.1% of patients with stage II TNBC would experience recurrence within 5 years, compared with 11.6% of patients with stage III TNBC (online supplemental figure A12c,d in online supplemental appendix A7).

Detection outcomes per 10 000 patients with TNBC

Because nearly all recurrences are reported to occur within the first 5 years following curative-intent surgery for patients with TNBC, we focused on reporting the 5-year recurrence and detection outcomes.^{22 23} Specifically, the model predicted 614 recurrences per 10 000 stage II patients (194 locoregional and 420 distant) and 1155 recurrences per 10 000 stage III patients (366 locoregional and 789 distant) within 5 years. However, only 394 (64.2%) and 586 (50.7%) recurrences in stage II and stage III, respectively, were detected by current clinical follow-up methods, leaving 35.8% recurrences in stage II and 49.3% in stage III undetected within the 5-year period (table 1).

Table 1. Recurrence detection outcomes per 10 000 patients with TNBC.

5-Year outcomes per 10 000 population	Locoregional*		Distant†		Total
	Predicted recurrences	Detected recurrences	Predicted recurrences	Detected recurrences	Predicted recurrences	Detected recurrences
Stage II	194	137 (70.6%)	420	257 (61.2%)	614	394 (64.2%)
Stage III	366	214 (58.5%)	789	372 (47.2%)	1155	586 (50.7%)
Overall	234	155 (66.2%)	506	284 (56.1%)	740	439 (59.3%)

^*Detected via mammogram

^†Detected asymptomatically or symptomatically following suspicious findings and confirmed via CT or MRI.

TNBC, triple-negative breast cancer.

For the overall population (76.7% stage II and 23.3% stage III), the model predicted 740 recurrences per 10 000 patients with TNBC, 234 of which were locoregional (66.2% detection rate) and 506 distant (56.1% detection rate). Consequently, 33.8% of locoregional recurrences and 43.9% of distant recurrences remained undetected within the 5-year period (table 1).

Number of tests in the surveillance period needed to detect one TNBC

To assess healthcare resource utilisation, we calculated the number of tests required to detect one case of TNBC recurrence. We found that detecting one locoregional recurrence required 399 mammograms and five subsequent core needle biopsies, while detecting one distant recurrence required 279 imaging tests and 39 biopsies over the same period.

Sensitivity analysis

To assess the uncertainty in the model, one-way sensitivity analysis was performed to examine the effect of each parameter on the 5-year recurrence detection rate. Using a triangular distribution approach, each parameter was varied between its minimum and maximum values (online supplemental table A6 in online supplemental appendix A8) to evaluate its impact on the detection rates. The 10 most sensitive parameters were identified and displayed in a tornado diagram (figure 4). The sensitivity analysis showed that the 5-year recurrence detection rate was most sensitive to the percentage of symptomatic cases among distant recurrence patients, followed by HR for increased risk of mortality for undetected patients compared with detected patients. Again, due to the fact it is not currently clear what potential impact undetected recurrences may have on survival, a wide range of HR values was explored, with the lower bound of HR 1.0864. The sensitivity of imaging tests for detecting symptomatic cases had the third largest effect. Furthermore, the ‘adherence to mammogram’ parameter, ranking fourth in the sensitivity analysis, further underscores the critical role of adherence to screening protocols in improving locoregional recurrence detection.

Figure 4. One-way sensitivity analysis results—percentage of detected recurrences among all recurrences. ctDNA, circulating tumour DNA; NAT, neoadjuvant therapy.

Discussion

Current guidelines are designed to monitor for recurrence; however, their effectiveness in improving outcomes, particularly for patients with TNBC, remains unclear. Our clinically validated model is the first to integrate the surveillance period and treatment guidelines with the natural disease progression implemented, providing a comprehensive assessment of TNBC management. By simulating both treatment outcomes and the effectiveness of the surveillance period in detecting recurrences, our findings highlighted the suboptimal detection of recurrences in patients with TNBC under the current SoC regimen, particularly for distant metastases.

A key insight from our study is the notable gap in the ability of current SoC modalities used in the surveillance period to detect recurrences in patients with TNBC. Despite predicting 740 recurrences per 10 000 patients with TNBC over 5 years, nearly 40% of the patients went undetected, especially with distant recurrences. This was particularly striking in patients with stage III TNBC, where almost 50% of the recurrences remained undetected. These findings underscore the inadequacies of current recommended practices in both stages, especially for distant recurrences, highlighting the need for more tailored or advanced methods.

Additionally, our model emphasised the considerable resource utilisation associated with the current surveillance period, with hundreds of mammograms, imaging tests including PET, CT and MRI and biopsies required to detect a single case of recurrence. This indicates high resource utilisation relative to the number of detected recurrences, suggesting that current protocols may be resource intensive without delivering proportional benefits in early recurrence detection.^24–26 Future research should explore more efficient strategies, such as incorporating more precise biomarkers that could tailor when and which imaging techniques might be most appropriate to employ during clinical follow-up, for the purpose of improving detection rates while optimising resource use. Sensitivity analysis further revealed that detection rates were mostly dependent on adherence to mammogram and other imaging protocols, particularly in cases of symptomatic distant recurrence. This finding underscores the need for improving patient adherence to schedules during the surveillance period, as doing so could potentially enhance the effectiveness of current practices.

The results of this study should be interpreted considering its limitations. First, while it was assumed that some patients would receive radiation therapy, as reported by Schmid et al,⁸ disease progression was not stratified based on radiation therapy uptake. This may limit the model’s ability to fully reflect the complexity of real-world patient outcomes. Second, the survival data used to inform the model had a 5-year time horizon, necessitating the extrapolation of transition probabilities, particularly for transitions from recurrence-free to recurrence stages and death. Third, due to limited real-world data, we assumed that the clinical outcomes for patients with undetected recurrence were equivalent to those of untreated patients, leading to increased mortality due to delayed treatment. It should be noted that all recurrences are eventually detected; however, to capture the impact of delayed detection, increased mortality risk was applied until recurrence detection. To reflect this, we applied an HR to estimate the mortality differences between detected and undetected recurrent cases at baseline and tested using the sensitivity analysis. While this assumption was necessary for modelling purposes due to lack of data, it is a hypothesis-generating opportunity for future clinical research to further investigate the impact of early detection on survival outcomes. Furthermore, patient heterogeneity in TNBC presented additional challenges that warrant further investigation.²⁷

In conclusion, our validated model highlights the clinical and operational challenges in the surveillance period of patients with TNBC. Current clinical follow-up guidelines demonstrate relatively modest detection rates, particularly for distant recurrences, while demanding substantial resource utilisation, underscoring the need for more efficient strategies. Improving adherence to existing protocols, enhancing the sensitivity of modalities used in the surveillance period and exploring novel detection approaches are essential steps towards better outcomes for patients with TNBC. Future research should focus on optimising clinical follow-up protocols to maximise early detection while reducing unnecessary resource use. Findings from this model could serve as a valuable foundation for future updates to guidelines, contributing to improved management and outcomes for patients with TNBC.

Conclusion

Our model-predicted clinical outcomes for patients with stage II and III TNBC, following current SoC guidelines in the surveillance period, revealed notably suboptimal detection rates, especially for distant recurrences. These findings highlight the limitations of existing practices and their implications for patient outcomes. As innovative technologies emerge, our model serves as a valuable tool to inform clinical decision-making in this evolving landscape and guide future advancements in the management of patients with TNBC.

The funder had no influence on the study outcomes, despite affiliations between some authors and the funder.

Data availability statement

The statistical analysis code and associated files are available upon reasonable request from the corresponding author for academic use.