*Purpose: Alike other clinical biomarkers, eGFR trajectories are characterized by intra-individual variability, related to biological fluctuations and analytical error. These fluctuations hamper the distinction between alarming graft function deterioration or harmless fluctuation within the patient-specific expected normal range of eGFR. We postulated that a deep learning approach could forecast future eGFR sequences and inform clinicians on deviations of measured values from this patient-specific predicted range.

*Methods: The data consisted of 159 912 eGFR measurement from 2 155 transplantations, divided into a derivation cohort of 985 kidney transplant patients with 105 244 eGFR measurements, and two independent validation cohorts with 39 999 eGFR measurements from 1 171 patients. We used deep learning with innovative sequence-to-sequence modelling to predict patient-specific eGFR trajectories in the first 3 months after transplantation.

*Results: Both in the training and in the independent validation sets, the sequence-to-sequence models accurately predicted future patient-specific eGFR trajectories in the first 3 months after transplantation, based on the graft’s previous eGFR values. Accuracy increased with more eGFR values as input and more adjacent timeframe predictions as output (Figure 1). The sequence-to-sequence model predictions performed better than more conventional prediction models like autoregressive integrated moving average.

*Conclusions: We developed and validated a sequence-to-sequence deep learning model for individual forecasting of kidney transplant function. The patient-specific sequence predictions could be used in clinical practice to guide physicians on deviations from the expected intra-individual variability.

« Back to 2021 American Transplant Congress