A cognitive and physical stress co-classification effort started with acquisition of a training dataset and generation of machine learning models from 17 heart rate variability parameters. Accuracy was improved with multilayer perceptron models and tested on 85 firefighters in a cage maze. A specific platform acquired a dataset with better label accuracy providing a second model. Feature importance and model performance were assessed using the cage maze data. A SHAP analysis provided the basis for the model comparison and feature important assessment. Conclusions were drawn on best time windows, feature selection, and model hyperparameters.