P1374DEVELOPMENT OF HEMOGLOBIN PREDICTION AND ERYTHROCYTE STIMULATING AGENT RECOMMENDATION ALGORITHM (HPERA) USING RECURRENT NEURAL NETWORK IN END-STAGE KIDNEY DISEASE PATIENTS

Abstract

Abstract Background and Aims The optimization of anemia management is a challenging task due to the complexities of underlying diseases and heterogeneous responses to erythropoiesis-stimulating agents (ESA) in patients with end-stage kidney disease (ESKD). Recent studies have shown that machine learning (ML) algorithms can be an effective tool to predict hemoglobin (Hb) levels and determine the ESA doses in these patients. However, most of the proposed ML approaches are not designed to handle multivariate longitudinal patient data. Thus, we developed Hb prediction and ESA doses recommendation algorithm (HPERA) using recurrent neural networks (RNN). Method A total of 466 participants, who underwent hemodialysis in 7 hospitals in the Republic of Korea, were included in the present study. We selected 15 variables from an extreme gradient boosting (XGBoost) algorithm. The outcome of the prediction algorithm was Hb levels in next month. In the recommendation algorithm, the outcome was the ESA dose for target Hb next month. Among various types of RNN families, gated recurrent units (GRU) were used to build both the prediction and recommendation algorithms. In addition to holding out a separate validation dataset, we used a Gaussian noise layer following each input layer to avoid overfitting. We also performed linear regression, multilayer perceptrons, and extreme gradient boosting with an extensive hyperparameter search to validate our GRU-based prediction algorithm. The performances of each model were evaluated in terms of the mean absolute error (MAE). Results The mean age of the study population was 57.8 years, 248 (53.2%) participants are male, and the mean observation period is 30.0 months. The best result of our prediction algorithm in terms of MAE was 0.59 g/dL and was obtained by two stacked GRU layers followed by a single hidden feedforward network with 6-month follow-up patient data. The best recommendation algorithm had 43.2 μg in MAE and this was obtained by one GRU layer followed by two layers of a feedforward network. The HPERA had a lower overall ESA dose (μg/months) [155 (80–240) vs. 140 (70–210), P<0.001], decreased Hb difference (g/dL) [0.8 (0.4–1.4) vs. 0.6 (0.3–1.0), P<0.001)], and had a higher success and a lower failure rates of reaching target Hb compared to those in real practice. Conclusion The GRU-based prediction model outperformed previous ML methodologies, though hyperparameter turning was much simpler. Using the HPERA showed the possibility of a reduced amount of ESA, decreased Hb difference, and increased the reaching rate of target Hb levels. Our study revealed a great potential direction of anemia management using ML in ESKD patients.