Can the use of artificial intelligence and wearable technology enable the optimisation of fluid status in haemodialysis patients?
Volume overload is a modifiable risk factor for cardiovascular and all-cause mortality in haemodialysis patients. Despite the risks associated with volume overload, maintaining, and assessing volume is difficult to perform adequately. This multi-step research examines the use of digital health solutions and artificial intelligence in assessing fluid status in haemodialysis patients.
Digital health solutions that target interdialytic weight gain were examined first. A post-hoc analysis of a trial involving a digital health app demonstrated that there was insufficient justification to use home blood pressure as a metric for volume assessment. No association between home blood pressure and ultrafiltration volume was demonstrated using a linear mixed effects model. Next, a prospective validation study involving a wearable hydration monitor was conducted in 20 haemodialysis patients, termed ‘HOPE-02’. The diffuse reflectance spectroscopy device was neither accurate nor developed enough to detect changes inpatients’ fluid status.
The second phase of this research involved assessing dry weight automatically. Several single-centre, proof-of-concept algorithms were developed for the prediction of a reference dry weight and a bioimpedance-determined normohydration weight or overhydration index, respectively. A Long-short term memory network algorithm was able to predict a reference dry weight within a root mean squared error (RMSE)of 1.25kg.A linear regression algorithm predicted bioimpedance normohydration weight within an RMSE of 1.33 kg, using a dataset of 20patients.A prospective trial, termed ‘HOPE-03’,was conducted in 24 haemodialysis patients to assess the temporal validity of the linear regression algorithms in real-time using a prototype platform. Although performance degradation was evident, the insights acquired are applicable to future iterations. Lastly, a classification model was created using the HOPE-03 dataset. A logistic regression model hadan overall accuracy of 85% for the prediction of fluid categories.
The findings of this research led to a collaboration with Fresenius. Research involving the redevelopment and external validation of these algorithmsis ongoing.
History
First Supervisor
Prof. Conall O’SeaghdhaSecond Supervisor
Prof. Donal SextonThird Supervisor
Prof. Peter ConlonComments
Submitted for the Award of Doctor of Philosophy to RCSI University of Medicine and Health Sciences, 2024Published Citation
Sandys V,. Can the use of artificial intelligence and wearable technology enable the optimisation of fluid status in haemodialysis patients?. [PhD Thesis] Dublin: RCSI University of Medicine and Health Sciences; 2024Degree Name
- Doctor of Philosophy (PhD)
Date of award
2024-05-31Programme
- Doctor of Philosophy (PhD)