The value of computational language analyses as biomarkers in schizophrenia-spectrum disorders is increasing as a result of rapidly advancing linguistic techniques. Language technology evolves quickly and analytic techniques such as machine learning allow for the application of complex features to a clinically relevant goal. Language analyses show potential for a range of applications in schizophrenia; for example in identifying at risk groups on social media,[82,83] monitoring psychosis relapse through smartphone applications or predicting treatment response. Recent work using computational semantic tools such as semantic space and graph analysis, as well as phonetic acoustic markers, have proved successful in both diagnosis and prognosis of schizophrenia-spectrum disorders. Accuracy scores in differentiating patients from healthy controls, family members or at risk groups range from 80 to 90%, often outperforming clinical raters. Even the clinically difficult differentiation between psychosis and mania showed high specificity and sensitivity with language analysis (both 94%).
Further longitudinal studies across a broader range of ages, disease severity and illness durations will be needed to understand the trajectory of language disturbances in schizophrenia-spectrum disorders. Future research is needed to fully appraise the potential of language as a diagnostic or prognostic tool. For example, a variety of language characteristics could be targeted by combining disparate computational tools. This may improve the predictive power substantially; since the most often used tools (semantic space and acoustic measures) are thought to be a reflection of a different set of symptoms. Semantic incoherence is often associated with FTD or disorganized language,[24,60,85] while acoustic measures are often used to objectify negative symptoms.[29,75,86,87] Bringing these methods together acknowledges the heterogeneity of symptoms associated with schizophrenia-spectrum disorders. Combining several quantifiable aspects of language may also pave the road towards cross-diagnostic analyses. Finally, researchers in this field should aim to do cross-linguistic analyses, to examine whether these models hold for the great diversity of languages in the world.
We acknowledge the valuable contribution of authors we have been unable to cite due to space constraints.
Financial support and sponsorship
I.S. received a TOP grant from The Netherlands Organization for Health Research and Development (ZonMW, project: 91213009).
Curr Opin Psychiatry. 2020;33(3):212-218. © 2020 Lippincott Williams & Wilkins