Patient-reported outcomes (PRO) are measures of symptoms (e.g., difficulty swallowing, insomnia, blood in stool) that are directly reported by the patient. They are measured using questionnaires that contain individual items that pertain to particular facets of a patient’s symptom (e.g., presence/absence, frequency, intensity, interference with everyday tasks). Individual items are typically reported on an ordinal n-point Likert scale from 0 (low symptom) to n-1 (high symptom). Individual symptoms can be an indication of disease progression, treatment toxicities, or comorbidities and thus can give a wholistic measure of patient wellbeing. Here, we develop a mathematical model to leverage PROs to predict hospitalization of cancer patients treated with radiotherapy. First, in order to model the effects of radiotherapy on PROs, we introduce a novel concept of radiation exposure analogous to drug exposure. Radiation exposure is modeled using a one-compartment PK model with a constant elimination rate. Additionally, we introduce another novel concept of tumor pressure to describe the effect of tumor burden on PROs. Tumor pressure is defined simply as proportional to the tumor burden. In turn, we model tumor burden dynamics using the U-shaped Claret tumor growth inhibition (TGI) model. Individual PRO items are factored into two groups: cancer-related symptoms and treatment-related symptoms. Individual PRO item response dynamics are then modeled using an n-state continuous-time inhomogeneous Markov chain model. Cancer-related symptoms include tumor pressure as a time-varying covariate in the transition rate matrix, whereas treatment-related symptoms include radiation exposure as a time-varying covariate in the transition rate matrix. We simulate PROs every 2 weeks for 104 weeks. We then model two latent variables describing overall cancer-related symptom burden and overall treatment-related symptom burden, respectively, using the 2-parameter (i.e., difficulty, discrimination) graded response model (GRM) from item response theory (IRT). Finally, we model hospitalization events using another 2-state continuous-time inhomogeneous Markov chain model with cancer-related and treatment-related symptom burden as time-varying covariates in the transition rate matrix. Future directions include validating the developed predictive model to real world data.