Can Predictive Models Determine 30-Day Mortality for Patients With Lung Cancer?

By: Joshua D. Madera, MS
Posted: Monday, February 13, 2023

Some decision support tools may play an efficacious role in limiting late systemic anticancer treatments for patients with advanced lung carcinoma, according to a study published in JCO Clinical Cancer Informatics. However, additional efforts are necessary to better establish an easy-to-use prototype for prospective clinical studies, suggested Charles Vesteghem, PhD, of Aalborg University, Denmark, and colleagues.

“Using predictive modeling may potentially help to limit late [systemic anticancer treatments], reducing the risk of causing unnecessary harm to patients in the late stage of life,” explained the study authors.

A total of 2,368 patients with advanced lung carcinoma who died between 2010 and 2019 were included in the study. Patients had either previously received systemic anticancer treatments or had metastatic lung cancer. Patients were stratified into three groups based on their date of death: patients who died between 2010 and 2017 (training cohort), patients who died in 2018 (validation cohort), and patients who died in 2019 (test cohort). All patient data were collected from the Aalborg University Hospital and included histopathologic results, biochemical data, diagnoses, and treatments. These data were compared using five different predictive models.

The most effective predictive models were the gradient tree boosting and random forest models, and the least effective predictive model was the artificial neural network–based model. The average precision for the gradient tree boosting and random forest models increased from 0.500 to 0.505 and from 0.498 to 0.509 when summative variables were added, respectively. A utility analysis of the predicted risk of 30-day mortality revealed that 40% of late systemic anticancer treatments administrations could have been avoided for alectinib, carboplatin, etoposide, and osimertinib.

Disclosure: For full disclosures of the study authors, visit ascopubs.org.

JCO Clinical Cancer Informatics