While increasingly appreciated for its importance, the interaction between health care professionals (HCP) and patients is notoriously difficult to study, with both methodological and practical challenges. The former has been addressed by the so-called Verona coding definitions of emotional sequences (VR-CoDES) - a system for identifying and coding patient emotions and the corresponding HCP responses - shown to be reliable and informative in a number of independent studies in different health care delivery contexts. In the preset work we focus on the practical challenge of the scalability of this coding system, namely on making it easily usable more widely and on applying it on larger patient cohorts. In particular, VR-CoDES is inherently complex and training is required to ensure consistent annotation of audio recordings or textual transcripts of consultations. Following up on our previous pilot investigation, in the the present paper we describe the first automatic, computer based algorithm capable of providing coarse level coding of textual transcripts. We investigate different representations of patient utterances and classification methodologies, and label each utterance as either containing an explicit expression of emotional distress (a `concern'), an implicit one (a `cue'), or neither. Using a data corpus comprising 200 consultations between radiotherapists and adult female breast cancer patients we demonstrate excellent labelling performance.