The shale barrier (SB) tends to impede the steam chamber development because of the low permeability of shale during the steam-assisted gravity drainage (SAGD) process. The interpretation of 4-D seismic data in geostatistical reservoir modelling is traditionally used to understand better the uncertainty of distribution of SB. On the other hand, long-term 4-D seismic survey to monitor the steam chamber development incurs additional cost.In this study, machine learning (ML) approaches, such as multiple linear regression (MLR), Random Forest, and artificial neural network (ANN), were performed to predict the horizontal location, width, and length of the SB using the inflection points of the SAGD production data; oil rate, cumulative oil, steam-oil ratio, and cumulative steam-oil ratio and the reservoir and operating parameters; reservoir thickness, horizontal permeability, vertical-horizontal permeability ratio, oil saturation, and steam injection pressure without the application of 4-D seismic data.Data pre-processing was carried out to examine and select 38 parameters (5 reservoir and operating parameters, the vertical location and thickness of the SB, and 31 inflection points), and the dataset (2131 cases) was divided into 75% of the train set and 25% of the test set. The k-fold cross validation and grid search process for hyperparameter tuning were performed to optimize the ML models. A comparison of the ML models revealed the ANN models to have larger coefficient of determination (R2) values than those of the MLR and Random Forest models. In the optimal ANN model, the R2 values of the test set for predicting the width (0.869) and length (0.918), but the R2 value for predicting the horizontal location was as low as 0.569.For verification, the optimal ANN model was tested using the production data from Husky's Sunrise and Suncor's Firebag SAGD projects in Canada. The field production data and simulation results using the reconstructed SB showed a similar trend, and the similarity between the field data and simulation results was improved compared to the results reported by Kim and Shin (2018).