Limited access to textual data, particularly de-identified mental health records, hinders NLP development in the clinical domain. This study presents a method for generating artificial clinical documents using discharge summaries from a large mental healthcare provider and an intensive care unit. The generated text undergoes intrinsic evaluation (text preservation, memorization of training data, clinical validity) and extrinsic evaluation (impact on NLP text classification). Results show that using artificial data for training yields comparable classification results to using original data, and that a small amount of original data is sufficient to condition generation, minimizing the risk of retaining sensitive information. This approach is promising for creating shareable artificial clinical data to advance computational methods using healthcare data.

Julia Ive, Natalia Viani, Joyce Kam, Lucia Yin, Somain Verma, Stephen Puntis, Rudolf N. Cardinal, Angus Roberts, Robert Stewart, Sumithra Velupillai