A Bayesian network to improve patient reported symptom monitoring in lung cancer

Abstract

Background: The SYMPRO-Lung study showed that patient-reported outcome (PRO) symptom monitoring significantly improved health-related quality of life (HRQOL) in lung cancer patients. Nonetheless, a substantial amount of alerts resulted in unnecessary con sults. To reduce the healthcare burden, we aimed to improve the alerting-algorithm through the use of a Bayesian network (BN) to compute the probability that PRO symptoms indi cate an alarming situation. Methods: A retrospective analysis was performed using data from the SYMPRO-Lung study, which included 248 patients who collectively completed 7.239 weekly questionnaires reporting their symptoms. The study focused on constructing a BN based on this data to model the alerting-algorithm, with the goal of reducing unnecessary alerts. The BN’s per formance was evaluated against the alerting-algorithm. The learned relationships between variables were used to infer the probability of the underlying reasons for symptoms and alerts, to better understand the different alerts. Results: In total 1326 alerts were sent by the alerting-algorithm, of which 468 (35%) were found unnecessary by participants or healthcare practitioners (HCPs). The constructed BN scored lower (accuracy 86% and precision 44%) on the performance metrics compared to the alerting-algorithm (accuracy 94% and precision 86%). Treatment necessity exhibited the highest likelihood of underlying symptoms and alert, followed by other reasons and impaired emotional functioning, indicating varying alert contexts. Conclusion: Although the BN can provide insight into the relationships between the vari ables in the alerting-algorithm and medical models, it did not perform better than the alerting-algorithm. The BN found differences within both the necessary and unnecessary alerts, showing alarming and non-alarming representations in both groups. Future work should improve the clinical validity of its captured relationships between variables and ex plore a dynamic approach to the BN.