A Time Series Foundation Model for Cancer Management

Chenlian Fu, Ruian Shi, Michele Waters, Nikolaus Schultz, Justin Jee, Quaid Morris

Keywords: Cancer, time series, outcome prediction, foundation model

TL;DR: A foundation model for predicting cancer outcomes using routinely collected lab and vital signs time series data.

Abstract: In cancer, predicting critical events, such as death or complications of treatment, is crucial for personalized medicine. Prediction of such events frequently ignores dense, longitudinal data such as laboratory tests and vital signs collected as part of standard of care. We present a foundation model for adverse event prediction, ‘Surveillance of Patient Adverse events using Routine Clinical data’ (SPARC). SPARC outperforms models based on single-timepoint measurements as well as previous time series architectures for predicting adverse cancer outcomes, i.e. side effects of chemotherapy, immunotherapy and death. SPARC generalizes to external validation datasets, excels with limited training data availability, and incorporates non-obvious features to improve outcome prediction. Overall, SPARC provides a generalizable and efficient solution for optimizing cancer treatment decisions.

Submission Number: 24