Automated analysis of free speech predicts psychosis onset in high-risk youths

Abstract

Background/objectives: Psychiatry lacks the objective clinical tests routinely used in other specializations. Novel computerized methods to characterize complex behaviors such as speech could be used to identify and predict psychiatric illness in individuals.

Aims: In this proof-of-principle study, our aim was to test automated speech analyses combined with Machine Learning to predict later psychosis onset in youths at clinical high-risk (CHR) for psychosis.

Methods: Thirty-four CHR youths (11 females) had baseline interviews and were assessed quarterly for up to 2.5 years; five transitioned to psychosis. Using automated analysis, transcripts of interviews were evaluated for semantic and syntactic features predicting later psychosis onset. Speech features were fed into a convex hull classification algorithm with leave-one-subject-out cross-validation to assess their predictive value for psychosis outcome. The canonical correlation between the speech features and prodromal symptom ratings was computed.

Results: Derived speech features included a Latent Semantic Analysis measure of semantic coherence and two syntactic markers of speech complexity: maximum phrase length and use of determiners (e.g., which). These speech features predicted later psychosis development with 100% accuracy, outperforming classification from clinical interviews. Speech features were significantly correlated with prodromal symptoms.

Conclusions: Findings support the utility of automated speech analysis to measure subtle, clinically relevant mental state changes in emergent psychosis. Recent developments in computer science, including natural language processing, could provide the foundation for future development of objective clinical tests for psychiatry.

Figure 1

Pipeline for automated extraction of the semantic coherence features. Texts were initially split into sentences/phrases. Each word was represented as a vector in high-dimensional semantic space using Latent Semantic Analysis (LSA). Summary vectors were calculated as the mean of each vector in each phrase. Coherence was determined based on the semantic similarity between adjacent phrases, calculated as the cosine of their respective vectors. The semantic coherence feature that best discriminated those who transitioned to psychosis from those who did not was the minimum semantic coherence value (i.e., the coherence at the point of maximal discontinuity) within each transcribed text.

Figure 2

Pipeline for cross-validation of the Machine Learning classifier. A vector of features for each participant is extracted and fed into the classifier that was trained on the other participants’ data. The classifier is used to predict outcome for the left-out, or test, participant. Each participant is sequentially left out of the training data set to serve as the test subject once, resulting in accuracy of prediction data for all participants.

Figure 3

Discrimination between individuals who transitioned to psychosis (clinical high risk+ (CHR+); in red) and those who did not (CHR−; in blue) presented as the convex hull of CHR− individuals. Color shading within the convex hull is used only to illustrate volume. Discrimination was based on three features extracted from free speech using automated methods. The frequency of use of determiners (‘that’, ‘what’, ‘whatever’, ‘which’, and ‘whichever’) normalized by phrase length; the minimum semantic coherence between two consecutive phrases within the interview; and the maximum phrase length.

Figure 4

Effect of randomly shuffling a proportion of classic literary texts (degree of ‘disorder’) on the measure of semantic coherence developed. Data points represent the minimum semantic distance between two adjacent sentences within a text. Increasing levels of ‘disorder’ were associated with a decrease in the coherence measure employed.

Figure 5

Correlations of text features and clinical ratings (top panel) and between positive (sips-A) and negative (sips-B) symptoms (bottom panel). Upper panel: canonical correlation between the text features, and the Structured Interview for Prodromal Syndromes (SIPS) features. Lower panel: scatter-plot of A_total and B_total shows no association between subthreshold psychotic symptoms and negative symptoms. In both panels, clinical high risk− (CHR−) and CHR+ are labeled with blue and red dots, respectively. For the analysis, all features were centered and normalized.