Self-care is essential in managing chronic diseases when patients could not always be monitored by medical staff. It therefore fills in the gap to provide patients with advice in improving their conditions in day-to-day practices. However, effectiveness of self-interventions in encouraging healthy behaviour is limited, as they are often delivered in the same manner for patients regardless of their demographics, personality and individual preferences. In this paper, we propose strategies to generate personalized health intervention messages departing from assumptions made by theories of social cognition and learning, planned behaviour and information processing. The main task is then defined personalised argument generation task. Specifically, an existing well-performing Natural Language Generation (NLG) pipeline model is extended to modulate linguistic features by ranking texts generated based on individuals’ predicted preferences for persuasive messages. Results show that the model is capable of generating diverse intervention messages while preserving the original intended meaning. The modulated interventions were approved by human evaluators as being more understandable and maintaining the same level of convincingness as human-written texts. However, the generated personalised interventions did not show significant improvements in the power to change health-related attitudes and/or behaviour compared to their non-personalised counterparts. This is attributed to the fact that human data collected for the model’s training was rather limited in size and variation.