Abstract Background and Aims Adequate small solute clearance and ultrafiltration (UF) are the main requirements for fluid and salt homeostasis on peritoneal dialysis (PD). New automated peritoneal dialysis (APD) systems that allow remote patient monitoring in terms of PD treatment analysis have the potential to improve clinical outcomes. UF on APD although dependent on PD prescription, i.e. fill volumes and glucose concentration, exhibits substantial intra- and interindividual variation, even day-by-day. As conventional statistical methods are not able to assess these complex non-linear relationships adequately, we aimed to build a deep learning algorithm to predict ultrafiltration trajectories of individual patients 12 weeks ahead. Method Data on daily UF and PD prescriptions (i.e., glucose concentration, fill volumes) was extracted from the APD cycler management software (PD Link, Baxter, IL, USA) and patient charts at the Medical University of Vienna, for a secondary analysis of this cohort. The Keras high-level API for TensorFlow in R software (R Core Team 2023) was used to create a Long Short-Term Memory (LSTM) recurrent neural network based on historic prescription and gcUF data of 26 weeks (after the first 90 days of APD) to predict individual UF trajectories for the following 12 weeks. Results APD machine readouts from a total of 123 patients, with a mean time on PD of 117 (±73) days, were retrieved. The final recurrent neural network consisted of four LSTM layers with 64 cells each, sigmoid activation. The final model was fit using 30 epochs, 49 steps per epoch, and recurrent dropout to avoid overfitting. A 40-40-20% data split was used to train, validate, and test the model adequately. This deep learning algorithm was able to predict individual UF values, for each day of the upcoming 12 weeks, with an accuracy of at mean 201ml. Upon visual inspection, predicted individual UF trajectories smoothly resemble the overall day-by-day trajectory of the patient but fail to account for extreme spikes in either direction. Conclusion In conclusion, historic daily APD cycler and prescription data in coalescence with respective daily measurements of membrane function (i.e., gcUF) can be used to predict individual daily ultrafiltration values and trajectories of APD patients 12 weeks ahead using a deep learning algorithm. Therefore, continuous collection of PD prescription data and UF measurements may be used to monitor peritoneal membrane function of patients on APD and alert clinical teams upon detection of relevant changes of UF trajectories already 12 weeks ahead to facilitate early intervention. In contrast to current “screening” methods on PD, i.e., Peritoneal Equilibrium Test, these methods could reduce treatment burden.
Read full abstract