A computer-implemented method is provided for taking one or more actions relating to therapy, the method comprising: obtaining data comprising audio data relating to a therapy session between a therapist and one or more patients; extracting text data from the audio data to form a transcript; dividing the transcript into a plurality of utterances; using at least a first part of a deep learning model to assign a semantic representation to each of the plurality of utterances to produce a plurality of assigned utterances; compiling the plurality of assigned utterances to form a representation of the therapy session; using at least a second part of a deep learning model, and an input comprising the representation of the therapy session, to obtain an output predicting a characteristic of the therapist, and/or the therapy, and/or the one or more patient; and causing the system to take one or more actions relating to the therapy, wherein the one or more actions are selected based on the output meeting one or more predetermined criterion.