<h3>Purpose/Objective(s)</h3> Unplanned hospitalizations during cancer treatment can impact cancer outcomes, quality of life, and healthcare costs. Prior studies have suggested that early prediction may facilitate interventional strategies to reduce hospitalization rates. Consumer wearable devices and patient-generated health data (PGHD) have created much enthusiasm, but their clinical utility remains unclear. The objective of this study was to develop and internally validate machine learning (ML) approaches to daily step counts before and during chemoradiation (CRT). <h3>Materials/Methods</h3> Patients enrolled in three prospective, single-institution trials of activity monitoring during chemoradiation (NCT02649569, NCT03102229, NCT03115398) were included in this retrospective ML model development study. Study patients were asked to wear commercial fitness trackers continuously before and during curative-intent CRT for multiple cancer types. Patient (age, ECOG performance status, sex, and diagnosis), radiotherapy plan metrics (mean esophagus, heart, and lung dose), and step count data were integrated. Patients with missing demographics or a hospitalization within a week prior to CRT were excluded. Step counts were smoothed as a 3-day running average and normalized. Weekly features were developed, including mean, median, minimum, maximum, range, and standard deviation, and the absolute and relative differences in weekly features between two given weeks were determined. An elastic net-regularized logistic regression (EN) was trained in a training cohort, with and without step count-derived features, to predict a first hospitalization event within one week based on data from the preceding two weeks. The models were then evaluated on a separate, hold-out test population in terms of the area under the receiver operating characteristic curve (AUC). <h3>Results</h3> 214 patients, median age 61 (IQR 53-68), were included in this study. The most common diagnoses were head and neck cancer (30%) and lung cancer (29%). 151 patients (70%) were in the training cohort, and 63 (30%) were in the validation cohort. EN with step count features had strong predictive performance (AUC = 0.81 [0.62-0.91]) and significantly outperformed EN without step counts (AUC = 0.57 [0.40-0.74], p = 0.004). Top five contributing variables were step counts from each of the past two days, absolute difference in minimum step counts over the past two weeks, relative decrease in the maximum step count over the past two weeks, and relative decrease in the step count range over the past two weeks. <h3>Conclusion</h3> PGHD has the potential to improve predictive ML modeling and direct clinical care. The results highlight the ability to predict hospitalization events based on daily step counts. This model is planned for clinical validation in NRGF-001, which will randomize patients undergoing CRT for lung cancer to treatment with or without daily step count monitoring.