The prediction of speech recognition is an important tool for the optimization of speech enhancement algorithms. The first Clarity Prediction Challenge was organized to find the most accurate prediction models for listeners with hearing-impairment and stimuli processed by different speech enhancement algorithms. The modified binaural short-time objective intelligibility (MBSTOI) represents the baseline. Our challenge contribution is based on a model for predicting listening effort. Predictions are obtained non-intrusively using only the output signals from the hearing aid processors. The challenge is split into a closed data set where all listeners and enhancement algorithms are included in training and testing, and an open data set where some listeners and one algorithm are missing in the training set. For the closed set, an individual mapping from the model output to speech intelligibility scores is used whereas for the open set the same mapping is applied for all data points. The model achieves a prediction accuracy of 25.88% root mean squared error (RMSE) (MBSTOI: 28.52%) and a correlation of 0.70 for the closed set. The open set results in an RMSE of 32.07% (MBSTOI: 36.52%) and a correlation of 0.54. The proposed non-intrusive model outperforms the intrusive MBSTOI for both data sets.