AI pilot reconstructs breast cancer staging logic in low-resource settings

A pilot project in Nigeria has demonstrated that a lightweight, guideline-informed machine-learning tool can reproduce the American Joint Committee on Cancer’s eighth-edition breast cancer staging rules from structured clinical observations. The decision tree classifier achieved perfect classification on its held-out test set and on 39 de-identified retrospective cases, but the researchers emphasize that these results verify the correct operation of a deterministic mapping rather than demonstrate independent clinical validation or predictive performance in new populations. The system was designed for low-resource centres, where large annotated datasets, specialist staffing and robust digital infrastructure may be scarce.

The study addresses a central feature of breast cancer care in sub-Saharan Africa, where diagnostic and treatment services are constrained by shortages of trained personnel, uneven access to high-quality facilities and disparities in the timely delivery of care. Many artificial-intelligence systems are built around data-intensive models that require large proprietary datasets and substantial computing resources, making them difficult to introduce in settings where those resources are limited. The pilot instead began with the AJCC Breast Cancer Staging Manual, eighth edition, and converted its anatomical TNM framework into a computational model built with open-source Python libraries.

Unlike a conventional predictive model intended to discover previously unknown relationships in patient data, this programme was explicitly a functional verification exercise. The researchers extracted descriptors corresponding to tumour extent, regional nodal involvement and distant metastasis, then generated every valid feature-to-stage combination permitted within the selected AJCC criteria. A decision tree was trained to reconstruct the relationship between those observations and anatomical stage groups. Its branching structure and terminal leaves were intended to reproduce the guideline logic rather than estimate an uncertain clinical prognosis.

The input features reflected findings that can be recorded during clinical evaluation. Tumour size was represented by ordinal thresholds corresponding to T1, T2 and T3, while T4 was captured separately as evidence of local anatomical invasion such as skin or chest-wall involvement. Three binary variables described nodal status: discrete or mobile axillary nodes, fixed or matted axillary nodes, and supraclavicular or contralateral axillary involvement. A further binary feature recorded the presence or absence of distant metastasis. Together, these variables uniquely determined stage within the study’s defined anatomical framework.

The model was developed in seven stages, beginning with identification of the AJCC manual and proceeding through extraction and structuring of clinical descriptors, construction of valid feature-stage combinations, resampling, model training and verification. The initial dataset contained 39 cases and did not cover the complete staging space. Because some stage categories were under-represented, early decision-tree splits were unstable. An attempt to use the Synthetic Minority Over-sampling Technique generated clinically invalid combinations because several variables were categorical and constrained by the staging rules, so the researchers instead used structured resampling to represent every valid combination.

After resampling, the dataset contained 158 rows. Each valid combination was repeated five times in total except Stage IV combinations, which were repeated four times because distant metastasis alone determined that stage. The records were saved in comma-separated values format, loaded with Pandas into a Jupyter Notebook and converted into numerical arrays: a feature matrix with 158 rows and six input columns, and a target vector containing the corresponding stage labels. The data were shuffled before an 80–20 split for training and testing, with a fixed random state used for reproducibility.

During training, the decision tree’s Gini impurity fell from 0.807 to zero across its terminal nodes, indicating fully pure classification. On the 32-row test subset, the model correctly classified every sample. Precision, recall and F1 scores were each 1.00 for every included stage category, and both macro-average and weighted-average measures were also 1.00. Three-fold cross-validation produced accuracy scores of 1.0 in every fold, with a mean of 1.0 and a standard deviation of zero. The investigators said these zero-variance results reflected the deterministic nature of the selected features and the complete coverage of their staging combinations.

The researchers then applied the trained system to 39 fully de-identified clinical records from a previously published study that had not been used in model training or testing. The model again classified every case in agreement with its reference AJCC label, producing a confusion matrix with complete diagonal agreement. This retrospective exercise was described as instance-level functional verification rather than independent validation because both the reconstruction data and the clinical-case labels were based on the same AJCC ontology. In a sensitivity analysis, retraining on the original 39 cases without structured resampling reduced accuracy to 75%, reinforcing the finding that coverage of the valid feature space, not the raw number of rows, was central to reproducing the mapping.

Technically, the resulting tool remains functionally equivalent to a rule-based staging engine when supplied with fully specified, valid inputs. Its distinction lies in how that logic was constructed: the researchers reconstructed the guideline mapping through a standard machine-learning framework rather than manually writing nested conditional rules. They argue that this representation may be easier to integrate with contemporary machine-learning workflows and more readily extended to higher-dimensional inputs, including additional prognostic variables. Nevertheless, the pilot did not include Stage 0, which often requires pathological information, and it did not incorporate tumour biology, histological grade, hormone-receptor status, HER2 status or genomic-assay results needed for modern prognostic stage groups.

Several practical limitations remain before clinical use could be considered. The model was not tested against incomplete, ambiguous, mis-recorded or noisy observations, nor was its behaviour defined for missing inputs or internally inconsistent combinations. It also did not measure disagreement among clinicians estimating tumour size or assessing nodal, skin and chest-wall involvement. The authors propose prospective testing in consecutively enrolled patients, independent comparison with expert-assigned stages, explicit robustness analysis, and quantification of inter-observer variability. Data governance, ethical oversight, regulation, usability and effects on clinical decisions would also need evaluation. The pilot therefore offers an infrastructure-frugal proof of concept for consistent, auditable staging support, but it does not yet show that the system can safely classify the messier cases encountered in routine practice or improve outcomes for patients.

Subject of Research: Pilot use of guideline-informed machine learning to reconstruct AJCC breast cancer staging in a low-resource Nigerian setting.

Article Title: Leveraging Artificial Intelligence for effective breast cancer management in low-resource centres: a pilot project of a TNM staging machine

Article References: Olayide, A., Chijindu, C., Ganiyu, R., Kehinde, O., Jibril, A., Julius, O., Halima, A., & Samuel, O. (2026). Leveraging Artificial Intelligence for effective breast cancer management in low-resource centres: a pilot project of a TNM staging machine. Discover Artificial Intelligence, 6(1), Article 1098. https://doi.org/10.1007/s44163-026-01811-1

Image Credits: AI Generated

DOI: 10.1007/s44163-026-01811-1

Keywords: breast cancer, artificial intelligence, machine learning, decision trees, TNM staging, AJCC guidelines, low-resource healthcare, clinical decision support, Nigeria, Leveraging, Artificial, Intelligence