Enhancing Explainability and Personalisation with Bayesian Networks for Patient Reported Symptom Monitoring
Keywords
Loading...
Authors
Issue Date
2024-10-31
Language
en
Document type
Journal Title
Journal ISSN
Volume Title
Publisher
Title
ISSN
Volume
Issue
Startpage
Endpage
DOI
Abstract
Background: The SYMPRO-Lung study showed that patient-reported outcome (PRO) symptom moni toring for lung cancer patients significantly improved the health-related quality of life (HRQoL) of patients
in comparison to care as usual. However, to implement PRO-symptom monitoring on a larger scale, im provements to the alert algorithm of the SYMPRO-Lung study were required to decrease the number of
unnecessary alerts that contributed to a higher workload for health care providers (HCPs). Additionally,
with the growing use of technological tools, such as clinical decision support systems, it was explored how
explainable models like Bayesian networks (BNs) could be used to provide additional information about
the patients’ condition to the HCP.
Methods: A total of 7232 PRO-symptom monitoring reports from 244 participants with in the SYMPRO Lung study were used to develop a Noisy-OR network aimed at improving the alert algorithm. The
performance of the Noisy-OR network was evaluated by comparing the total number of predicted alerts
against the outcomes of the original alert algorithm. Furthermore, the PRO-symptom reports and the
patient characteristics, were used to construct a BN to explore potential underlying reason(s) for the
reported symptoms. The BN’s predictions for the most probable underlying reasons were compared to
the actual patient characteristics. To give more insight about the influence of variables on the BN’s pre dictions, MAP-independence was performed to find which symptoms contributed to the most probable
underlying reasons. Finally, all components of the study were integrated into the PROM-pipeline tool,
which not only functioned as an alert algorithm for PRO-symptom monitoring but also provided patient
insights to HCPs.
Results: The alerts from the SYMPRO-Lung study were divided into 799 (60.3%) necessary alerts and
526 (39.7%) unnecessary alerts depending on the follow-up of the alert. The Noisy-OR network showed a
reduction in the number of unnecessary alerts, while effectively retaining alerts for severe symptoms. The
outcomes of the BN indicated that undergoing treatment was the most probable underlying reason for the
reported symptoms, followed by impaired emotional functioning. Patients who reported necessary alerts
were more likely to have comorbidities, disease progression, or impaired emotional functioning. Moreover,
the MAP-independence revealed that different symptoms contributed to the underlying causes depending
on the stage of lung cancer. Lastly, it was concluded that the design of the PROM-pipeline tool had
to give patient-specific insights that were easy to interpret for HCPs. The HCPs had the option to also
access more detailed patient insights, such as the most probable underlying reason for the symptoms and
the MAP-independence results for certain variables. However, to avoid overreliance on the BN outcomes,
it was necessary to provide adequate caution about the uncertainty of the model’s predictions.
Conclusion: Several BN techniques were combined in the PROM-pipeline tool which was used to pre dict alerts for reported symptoms and to optimize the personalization of PRO-symptom monitoring care.
The Noisy-OR network for alert prediction was able to reduce the overall number of alerts and the BN
and MAP-independence were able to give insights about the relation between symptoms and underlying
reasons. With additional data where certain patient characteristics are better represented it is possible
to optimize the performance of both the Noisy-OR network as well as the BN. Finally, it is suggested
that expert knowledge should be incorporated into the Noisy-OR network and the BN to improve the
credibility of the outcomes
Description
Citation
Faculty
Faculteit der Sociale Wetenschappen