Towards a data fusion model for predicting deterioration in dialysis patients

The accumulation and relatively rapid removal of fluid in haemodialysis patients is often accompanied by intradialytic hypotension (IDH). Current patient monitoring during haemodialysis includes intermittent measurements of tympanic temperature, blood pressure and haematocrit. However, this information is mostly used retrospectively rather than as a means for preventing adverse events. We suggest the use of a probabilistic data fusion model based on dialysis vital sign data to predict IDH. We continuously monitored the vital signs of 40 haemodialysis patients during 8 sessions over a 6-month period in the Oxford Renal Unit. The study involved non-invasively monitoring the heart rate, blood oxygen saturation, systolic and diastolic blood pressures as well as the tympanic temperature throughout each dialysis session. The 4-dimensional vital sign data was initially visualised on 2D projections using the Neuroscale algorithm. The projections show a distinction between data from unstable and stable patients, with data from hypotensive events appearing outside the region of the 2D projection corresponding to "normal" physiology. A data fusion model based on an estimate of the probability density function of data from stable patients was then created. With this model, instabilities in patient physiology can be identified, and the adverse event can be predicted ahead of time in some cases.