A machine learning clinicogenomic model was developed to identify patients with metastatic breast cancer at high risk of developing brain metastases. Findings from the development and validation of the CNSPredict model were presented at the 2026 ASCO Annual Meeting (Abstract 106).

Current guidelines do not recommend screening for brain metastases in patients with metastatic breast cancer unless they have neurologic symptoms. According to presenting author Luke Roy George Pike, MD, DPhil, Director of Brain Radiation Oncology and an assistant attending radiation oncologist at Memorial Sloan Kettering (MSK) Cancer Center, routine brain imaging for all patients with metastatic breast cancer would not be feasible at a population level. As a result, brain metastases are often detected at more advanced stages, when treatment options are more limited and neurologic complications are more common.  

A biomarker capable of identifying patients with metastatic breast cancer at the highest risk of developing brain metastases could enable earlier imaging and detection. Several clinical and genomic features have been associated with the development of brain metastases in this population, including HER2-positive disease and baseline alterations in NOTCH, PTEN, EGFR, and other genes. The researchers hypothesized that integrating these clinical and genomic features into a machine learning model could stratify patients by their risk of developing brain metastases and identify a high-risk subgroup that may benefit from earlier intervention.  

“To address this gap, we developed and validated CNSPredict, which is a machine-learning clinicogenomic model that identifies patients with metastatic breast cancer who are at high risk of developing brain metastases,” said Dr. Pike. 

Study and Model Methods 

The study included patients with metastatic breast cancer who were treated at MSK within 6 months of their diagnosis and underwent MSK-IMPACT sequencing of extracranial tissue within 1 year of diagnosis. MSK-IMPACT assesses alterations in 505 genes. Patients who did not develop brain metastases were required to have at least 6 months of follow-up.  

A total of 1,789 patients were included in the final cohort and followed longitudinally at MSK. Of these, 359 developed brain metastases. These patients tended to be younger and were more likely to have high-grade histology and HER2-positive or triple-negative disease. 

Clinicopathologic and genomic data from the full cohort were integrated and processed for model development. The cohort was then split into a training set (70%) and an internal validation set (30%).  

Using the training cohort (n = 1,255), the researchers developed CNSPredict through feature selection, least absolute shrinkage and selection operator (LASSO)–penalized regression, and internal cross-fold validation. The time-dependent DeepHit neural network was then used to model survival outcomes while accounting for competing risks.

The model was first validated in an internal holdout cohort of 534 patients. It was then reapplied to the full cohort to refine the analysis and identify key features associated with the risk of developing brain metastases.  

Dr. Pike noted that LASSO-penalized regression was selected to address concerns about correlated predictors and overfitting in high-dimensional clinicogenomic datasets. Risk scores were calculated using 100 iterations of LASSO-penalized regression, after which patients were stratified into low-, intermediate-, and high-risk groups based on tertiles.  

Additionally, the model was externally validated in a cohort of 257 patients from eight institutions and in data from a single-arm, phase II prospective screening clinical trial conducted at Moffitt Cancer Center. Because the trial enrolled a heavily pretreated population with limited sequencing data, only 19 patients had sequencing data of sufficient quality to be included in the analysis.  

Results 

Brain metastasis–free survival differed significantly across the three risk groups. At 2 years, brain metastasis–free survival rates were 95% in the low-risk group, 90% in the intermediate-risk group, and 75% in the high-risk group. Compared with the low-risk group, the intermediate-risk group had a hazard ratio (HR) of 2.63 (95% confidence interval [CI] = 1.70–4.05; P < .001), whereas the high-risk group had an HR of 7.62 (95% CI = 5.11–11.37; P < .001).  

Similar results were observed in the internal validation cohort. Compared with the low-risk group, the intermediate-risk group had an HR of 3.07 (95% CI = 1.54–6.12; P = .001), whereas the high-risk group had an HR of 8.97 (95% CI = 4.48–15.33; P < .001).

Model performance was characterized by a mean time-dependent area under the curve of 0.77 and a Harrell’s concordance index of 0.74.  

After model refitting, features selected in 70 of 100 bootstrap iterations of LASSO regression were identified as the most important predictors of brain metastasis risk. 

“These were reproducibly identified as being important predictors across repeated sampling of the data, suggesting that they're likely true and stable signals, rather than chance findings,” Dr. Pike explained. “Notably, the high-frequency features are biologically plausible.” 

The importance of these features was further supported by differences in their prevalence across the risk groups.  

Among the top predictors associated with a higher risk of brain metastases were triple-negative breast cancer, four or more metastatic sites, adrenal gland metastases, and TP53, PTEN, and ERBB2 mutations or amplifications. 

A confirmatory analysis was performed using DeepHit, which accounts for competing risks and nonlinear feature interactions when modeling event incidence. In the internal validation cohort, the model achieved an area under the curve of 0.72.  

Dr. Pike noted that patients with brain metastases were overrepresented in the external validation cohort because of selection bias. The model was applied to this cohort without retraining and demonstrated similar performance, with a mean area under the curve of 0.71 and a Harrell’s concordance index of 0.67. 

Compared with the low-risk group, the intermediate-risk group had an HR of 1.51 (95% CI = 0.92–2.47; P = .16), whereas the high-risk group had an HR of 3.53 (95% CI = 2.18–5.70; P < .001). Although these HRs were lower than those observed in the internal training and validation cohorts, Dr. Pike noted that the survival curves remained clearly separated over time. 

In the external Moffitt cohort, CNSPredict showed a similar trend toward distinguishing patients at high risk from those at intermediate or low risk (HR = 3.76; 95% CI = 0.41–33.95; P = .23).  

Limitations and Next Steps 

Dr. Pike noted that the study was limited to patients who underwent MSK-IMPACT sequencing, and the findings therefore may not be generalizable to the broader population of patients with metastatic breast cancer who lack access to specialized genomic testing and care. He also acknowledged that asymptomatic brain metastases were likely underdetected because routine screening is not recommended for patients without neurologic symptoms. In addition, treatment paradigms evolved during the study period to incorporate agents with central nervous system activity, such as trastuzumab deruxtecan, which may have influenced outcomes.  

The research team plans to prospectively validate its translational findings in the randomized phase II BRAINSTORM trial. That study is evaluating intensified MRI surveillance vs standard of care in patients with metastatic breast cancer who do not have neurologic symptoms but are identified by CNSPredict as being at high risk for brain metastases. The primary endpoint is neurologic event-free survival.  

DISCLOSURES: Dr. Pike disclosed a stock or ownership interest in Clovis Oncology, Novavax, and Schrodinger, as well as a consulting or advisory role with Aviko, Dxcover, Genece Health, Monograph Capital, and Turnstone Bio. He has also received research funding from Caris Life Sciences, Delfi Diagnostics, Genece Health, and Harbinger Health. For disclosures of the other study authors, visit asco.org.

ASCO AI in Oncology is published by Conexiant under a license arrangement with the American Society of Clinical Oncology, Inc. (ASCO^®). The ideas and opinions expressed in ASCO AI in Oncology do not necessarily reflect those of Conexiant or ASCO. For more information, see Policies.