Publication Date

Spring 2026

Degree Type

Master's Project

Degree Name

Master of Science in Computer Science (MSCS)

Department

Computer Science

First Advisor

Saptarshi Sengupta

Second Advisor

Faranak Abri

Third Advisor

Genya Ishigaki

Keywords

Survival modeling, Synthetic EHR Data, Synthea, Temporal features, Cox proportional hazards model, Concordance index

Abstract

Survival modeling is a critical component of oncology, enabling personalized care and improving clinical decision making. However, the development of these models is consistently bottlenecked by the scarcity of high quality longitudinal Electronic Healthcare Record (EHR) data, caused by strict legal restrictions. To bypass these restrictions and enable open research as well as faster iteration, synthetic data generators such as Synthea are being increasingly adopted. While previous literature has validated Synthea for various clinical analysis tasks, and survival modeling has been applied to real world EHRs, there remains a significant gap in evaluating whether the temporal event landmarks generated by Synthea are clinically valid for complex survival modeling. This study investigates the viability of utilizing temporal data extracted from Synthea generated cancer oncology cohorts to train predictive survival models. To evaluate their effectiveness, an ablation study was conducted using the Cox Proportional Hazards model. A baseline model was trained using static features, and the experimental model was trained using static and temporal data. The measured difference in the concordance index of the model, was +0.0224, from a baseline of 0.7811 to 0.8036. These findings establish that Synthea generated landmarks can be used effectively for survival modeling research in clinical oncology, as their inclusion in a CPH model increases its predictive ranking performance.

Download

Available for download on Saturday, May 22, 2027

Share

COinS