Although hemodynamic instability after starting continuous renal replacement therapy (CRRT) due to acute kidney injury is related with worse outcomes, its prediction model has not been established. The present study addresses whether machine learning algorithms can predict the hypotension after starting CRRT. We randomly divided a total of 1,191 adult patients who started CRRT for acute kidney injury into training (70%) and testing (30%) sets. The primary outcomes were a decrease in mean arterial blood pressure (MAP) ≥ 10 mmHg (MAP-1) or ≥ 20 mmHg (MAP-2) within 1 hour after starting CRRT. We used random forest (RF), extreme gradient boost (XGB) and artificial neural network (ANN) models for prediction. For evaluating the model performance, the area under the receiver operating characteristic curve (AUROC), accuracy, and F1 score were used. Furthermore, a multiclass classification model was applied to score the severity of hypotension, as follows: 0, no decrease; 1, < 10 mmHg; 2, 10–19 mmHg; 3, 20–29 mmHg; and 4, ≥ 30 mmHg (MAP-3).  In predicting MAP-1, the RF, XGB and ANN had AUROCs of 0.73 (0.675–0.781), 0.74 (0.691–0.793), and 0.76 (0.715–0.812), respectively. In predicting MAP-2, the RF, XGB and ANN had AUROCs of 0.75 (0.697–0.811), 0.74 (0.714–0.825), and 0.76 (0.715–0.822), respectively. All of these predictive performance were not different between models. Average of AUROCs for predicting MAP-3 were 0.652, 0.654, and 0.638 in RF, XGB, and ANN, respectively. Accuracy in MAP-3-prediction RF, XGB, and ANN models were 0.590, 0.621, and 0.612, respectively. In the XGB model, MAP itself at the time of starting CRRT was determined as the most important variable, followed by pH, serum sodium, and body temperature. Machine learning can be applied to predict the risk of hypotension after CRRT.