A machine learning model trained on real-world clinical data from patients with advanced nonsquamous non–small cell lung cancer (NSCLC) receiving first-line therapy accurately predicted overall survival (OS), closely matching Kaplan-Meier survival curves. Findings were presented in a poster during the 2025 ESMO AI & Digital Oncology Congress as part of a digital twin modeling analysis (Abstract 386P).

The model’s predictive accuracy enables the application of digital twin modeling, even amid a rapidly evolving standard of care for first-line treatment in nonsquamous NSCLC, according to Melissa Estevez, Director, Research Sciences, Flatiron Health, and colleagues.

Background

A digital twin is a virtual representation of a physical object, system, or process that can be analyzed and optimized.

Digital twin models powered by real-world data require reliable outcome predictions to support more personalized cancer care and faster drug development. However, this remains challenging in the advanced NSCLC population, where first-line treatment standards—and patient outcomes—are evolving rapidly.

To improve outcome predictions for digital twin modeling, the study authors incorporated factors influencing whether patients received chemotherapy alone or chemotherapy plus immune checkpoint inhibition in the first-line setting.

Study Methods and Materials

The researchers extracted electronic health record data from the Flatiron Health Research Database for patients with advanced nonsquamous NSCLC treated with first-line chemotherapy or chemotherapy plus immunotherapy between January 2018 and June 2024. Patients with ALK or EGFR alterations identified prior to initiating first-line therapy were excluded. All patients (n = 9,050) were followed through May 2025 or until death.

The predictive model used a random survival forest to analyze survival data and estimate the time from treatment initiation to death based on clinically relevant features derived from the health record, laboratory data, and variables extracted using large language models.

Predicted survival outcomes were compared with observed Kaplan-Meier curves to assess predictive accuracy, and integrative Brier scores were calculated to evaluate model performance over time.

Outcomes

The study cohort was randomly divided into training (60%), validation (20%), and test (20%) sets.

Only 17% of the cohort received first-line chemotherapy alone. The most common reasons for omitting immunotherapy were PD-L1 status considerations (25%), planned delays in immunotherapy (14%), and contraindications such as autoimmune comorbidities (10%). However, no documented reason was identified for 45% of these patients, and 6% had “other” reasons.

As a result, digital twin modeling incorporated large language model–extracted features, including PD-L1 status and history of autoimmune comorbidities, to address potential confounding. “Large language model–extracted features enhance clinical relevance and explainability as well as mitigate confounding by indication amid evolving standards of care,” the study authors wrote in their poster.

In the training cohort (n = 5,377), the model achieved an integrated Brier score of 0.17. Predicted and observed real-world OS rates matched at 1, 2, and 5 years (54%, 35%, and 17%, respectively).

Predicted and observed OS rates were similar, though not identical, in the validation and test cohorts at each time point. In the validation cohort (n = 1,817), predicted vs observed OS rates were 55% vs 54% at 1 year, 36% vs 37% at 2 years, and 19% for both at 5 years. In the test cohort (n = 1,856), predicted vs observed OS rates were 55% vs 53% at 1 year, 36% vs 35% at 2 years, and 19% vs 16% at 5 years.

The most important features in the model, based on permutation importance, were laboratory values, vitals, performance status, and stage.

Estevez et al reported that the model “produced well-calibrated and discriminative OS predictions, enabling digital twin modeling.”

The study authors also emphasized that external validation is needed to assess the model’s generalizability across target populations and additional data sources.

DISCLOSURES: The study was sponsored by Flatiron Health, Inc. All of the study authors are employed by Flatiron Health.

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.