Predictive models transform patient biomarker measurements into dynamic forecasts of disease and treatment response, enabling clinically actionable prediction rather than retrospective description. While invasive biomarkers—such as tissue biopsies, surgical pathology specimens, and cerebrospinal fluid sampling—provide rich biological detail, they are costly, risky, and sparsely collected, limiting their utility for longitudinal modeling. In contrast, minimally- and non-invasive biomarkers, including liquid biopsy measures and patient-reported outcomes (PROs), can be collected frequently with low patient burden, making them well-suited for dynamic prediction despite their lower biological resolution. Incorporating these data into predictive mathematical frameworks can present key modeling challenges, including noisy measurements, indirect observability of disease state, and inter-patient heterogeneity. Assessing a model’s predictive ability requires careful consideration of these potential issues. In this talk, we demonstrate how PRO data collected biweekly from non–small cell lung cancer patients receiving immunotherapy can be integrated into a calibrated and validated predictive model to forecast volumetric disease progression and generate patient-specific predictions. We conclude by highlighting practical modeling challenges, such as working with noisy and incomplete longitudinal data and balancing biological realism with mathematical tractability and discuss strategies for addressing these challenges in predictive oncology models.
Copyright. All Rights Reserved.