(188) Improved Prediction of Overall Survival Using Machine Learning Survival Models Incorporating Time-Dependent Features

Background: There is little information about the benefits of machine learning (ML) survival models and time-dependent features in the prediction of mortality using real-world data. There is a need to develop good ML models to predict patient risk in support of research and development of novel cancer medications of interest.

Objectives: To develop and assess ML models using random survival forest (RSF) and gradient boosting (GDB) to predict overall survival (OS) of locally advanced and/or metastatic non-small cell lung cancer (aNSCLC) patients.

Methods: A retrospective cohort of 16,092 aNSCLC patients was created from the IQVIA Oncology EMR – US database from 2015 to 2021. The outcome was overall survival (OS) from initial diagnosis of aNSCLC. Eighty baseline features included patient demographics, vital signs, stage, TNM, histology, biomarkers, therapies, and functional labs for ML survival models. Time-dependent biomarkers and labs were measured at each quarter of the first year. Data were randomly split 70% for training with cross validation and 30% for test. Time-dependent AUC (area under the curve) and Brier score were used to evaluate performance of ML models, which were developed using Python packages on a cloud-based Azure Machine Learning Platform.

Results: The study cohort consisted of 53% male patients, 32% aged 75+ years and 86% in stage IV. Maximum follow-up was 36 months with a median OS of 18.9 months. By using baseline features, AUC of 0.71, 0.71, 0.68 and 0.68 from RSF, and 0.72, 0.71, 0.68 and 0.69 from GDB for the OS prediction in 18, 24, 30 and 33 months, were observed respectively. Top baseline features of importance included albumin, neutrophil, age, white blood cell and platelet. To evaluate the additional predictive value of quarterly longitudinal markers in the first year, the models were developed by using patients surviving >12 months (n = 6,957) and five datasets with increasing number of the longitudinal markers: (1) baseline; (2) baseline and markers from Q1; (3) baseline and markers from Q1 to Q2; (4) baseline and markers from Q1 to Q3; and (5) baseline and markers from Q1 to Q4. This resulted in increasing AUC of 0.61, 0.64, 0.65, 0.67, and 0.71 from RSF, and 0.64, 0.65, 0.67, 0.70, and 0.74 from GDB at 18 months, respectively. There were monotonically decreasing Brier scores also indicating performance improvement from 0.1658 to 0.1571 for RSF and from 0.1662 to 0.1545 at 18 months for GDB, respectively. Similar patterns of AUC and Brier scores were observed for the OS prediction in 24, 30 and 33 months.

Conclusions: The prediction substantially improved when time-dependent marker data were added to the survival model while the performance of RSF and GDB was moderate overall in this cohort.