Machine-Learning Models May Help Determine Brain Metastasis Risk in Patients With NSCLC

Machine learning–aided prediction models are capable of learning the relationship between clinical features and later development of brain metastases (BM) in patients with non-small cell lung cancer (NSCLC) who are at a high risk of BM and might benefit from surveillance brain MRI, according to a study published in Clinical Lung Cancer.

Lung cancer is one of the most diagnosed cancers around the globe, with an estimated 2.2 million cases just in 2020, and is the leading cause of cancer death. NSCLC makes up approximately 85% of lung cancers, and the incidence of BM is 20-40% of patients with NSCLC. The incidence of BM in NSCLC is one of the highest among cancer subtypes, and early detection is vital to improve clinical prospects.

In the study, researchers trained and validated classifier models to pinpoint patients at high risk of developing BM, as they could possibly benefit from surveillance brain MRI.

A total of 395 patients made up the clinical cohort, which included patients with an initial diagnosis of NSCLC between January 2011 and April 2019 and an in-clinic chest CT scan who were retrospectively recruited at a German lung cancer center. Brain imaging was conducted at initial diagnosis and in case of neurological symptoms during follow-up. Subjects lost to follow-up or still alive without BM at the data cut-off point of December 2020 were not included.

Clinical and/or 3D radiomics features of the primary tumor from staging chest-CT were included as covariates. A total of 4 machine-learning models for prediction (80/20 training) were compared. Gini Importance and SHAP (serum-derived HA-associated protein) were used as measures of importance, while sensitivity, specificity, area under the precision-recall curve, and Matthew’s Correlation Coefficient were used as evaluation metrics.

“Classification models that employed only clinical features as input showed the best performance when tuned to maximize recall to improve prospective studies, obtaining at best a sensitivity of ∼ 70% and a specificity just above 60%,” the authors wrote.

Radiomics features did not provide sufficient information, probably because of the heterogeneity of imaging data. Adenocarcinoma histology, lymph node invasion, and histological tumor grade at diagnosis were positively correlated with the prediction of BM, while age and squamous cell carcinoma histology at the time of diagnosis were negatively correlated. The reason why younger patients have a higher BM risk has not been fully explored. Subpopulation analyses found that 2 subgroups seemed to be at a higher risk of BM: female patients with adenocarcinoma histology, and patients with adenocarcinoma with no other distant metastases.

An evaluation of the importance of the input features displayed a high agreement with the existing literature, suggesting that the prediction models are learning the relevant relationships between the clinical features of NSCLC and later BM development.

Some known predictors, like the presence of EGFR mutations, did not show statistically significant results, perhaps because of the small dataset size, compounded with the high fraction of missing values, the authors noted. Additionally, the rate of EGFR mutations in the primarily Caucasian population is expected to be extremely low, even in complete datasets.

In the analysis, female patients seemed to be more vulnerable to BM, which is mirrored in the correlation of SHAP values with BM. This could possibly be an indirect effect, the authors noted, since women are more likely to develop adenocarcinomas, and female sex is a predictive factor for EGFR mutations, which alone are positively correlated with the development of BM.

In general, the analysis suggests that the highly ranked features identified in the study should be included in any machine-learning model that aims to strongly predict the development of BM.

An additional approach that has shown potential in the identification of high-risk subgroups is the combination of clinical parameters with gene expression signatures in primary tumors that are specifically linked to BM and not associated with the development of metastasis to other sites or simply to disease recurrence. The oxidative phosphorylation pathway especially seems to be strongly associated with BM risk, the authors noted.

An important limitation was the fact that systemic treatment and testing for driver mutations had completely changed over the recruitment period, and so the true number of patients possessing driver mutations and their possible impact on the population was not clear.

“Employed prospectively at initial diagnosis, [prediction] models can help identify high risk patients who might benefit from surveillance brain MRI for early detection and treatment of newly developed, asymptomatic BM,” concluded the researchers.

Reference

Visonà G, Spiller LM, Hahn S, et al. Machine-learning-aided prediction of brain metastases development in non-small cell lung cancers. Clin Lung Cancer. Published online August 6, 2023. doi:10.1016/j.cllc.2023.08.002