Enhancing Non-Small-Cell Lung Cancer Patients’ Urgent Care Visit Prediction after Combined Modality Therapy with Patient-Reported Outcome and Wearable Sensor Data
Abstract
Purpose
Previous studies in predicting the risk of urgent care visits for non-small-cell lung cancer (NSCLC) patients receiving combined modality therapy (CMT) have generally focused on pre-treatment risk factors. Our study aimed to develop a risk prediction model with patient-reported outcome (PRO) data and wearable sensor data.
Methods
Participants included 58 patients with NSCLC treated with CMT at Moffitt Cancer Center. We collected demographic, clinical, and PROs (PROMIS-57) data before starting CMT and abstracted data on usage of urgent care within 60 days after starting CMT. We also collected wearable sensor data from Fitbit activity trackers during this 60-day period. The follow-up period was split into two 30-day periods. For clinical variables, we included the most extreme value within each 30-day period but before the first urgent care visit. For PROs, changes since baseline but before the first urgent care visit were included. First, we generated a Bayesian network (BN) predicting risk of urgent care visit using only demographic and clinical data available before starting CMT and a second BN including these data during-CMT. Next, we generated an enhanced pre-treatment BN that added pre-CMT PRO data. Lastly, we created an enhanced during-treatment BN that added wearable sensor and clinical data assessed during-CMT. Cross-validation and area under the receiver operating characteristic (AU-ROC) curve were used to evaluate the predictive performance of each BN.
Results
Participants were mostly female (57%), White (88%), and non-Hispanic (95%) with average age 69 years (range = 35-89). The enhanced pre- and during-treatment BNs significantly outperformed the BNs with only on pre-CMT (p <0.001) and during-CMT (p = 0.002) clinical and demographic data respectively.
Conclusion
Wearable sensor and PRO data before and during CMT can significantly improve prediction of NSCLC patents’ urgent care visit. Future studies should aim to validate this model in independent and external datasets.