Abstract
Background and Hypothesis: Mapping a patient’s speech as a network has proved to be a useful way of understanding formal thought disorder in psychosis. However, to date, graph theory tools have not incorporated the semantic content of speech, which is altered in psychosis. Study Design: We developed an algorithm, “netts”, to map the semantic content of speech as a network, then applied netts to construct semantic speech networks for a general population sample, and a clinical sample comprising patients with first episode psychosis (FEP), people at clinical high risk of psychosis (CHR-P), and healthy controls. Study Results: Semantic speech networks from the general population were more connected than size-matched randomised networks, with fewer and larger connected components, reflecting the non-random nature of speech. Networks from FEP patients were smaller than from healthy participants for a picture description task but not a story recall task. For the former task, FEP networks were also more fragmented than those from controls, showing more, smaller connected components. CHR-P networks showed fragmentation values in-between FEP patients and controls. A clustering analysis suggested that semantic speech networks captured novel signal not already described by existing NLP measures. Conclusions: Overall, these data suggest that semantic networks can enable deeper phenotyping of formal thought disorder in psychosis. We are releasing Netts as an open Python package alongside this manuscript.
Introduction: The Challenge of Formal Thought Disorder
Formal Thought Disorder (FTD) is a hallmark of psychosis, characterized by disorganized speech, "loosening of associations," and incoherence. While clinicians use tools like the Thought and Language Index (TLI) to measure this, manual assessment is time-consuming and subjective.
Existing Natural Language Processing (NLP) approaches often focus on either syntax (the order of words) or semantic coherence (a single summary score of how related words are). However, they often miss the structural "map" of meaning. This research addresses the gap by applying graph theory to the semantic content of speech—treating entities as nodes and their relationships as edges—to see if the structural topology of a patient's ideas can predict or describe their clinical state.
Methodology: The Netts Framework
The authors developed an algorithm called netts (Networks of Transcribed Speech) to transform transcribed speech into quantifiable semantic networks. The pipeline consists of four primary stages:
- Preprocessing: Utilizing Stanford CoreNLP for lemmatization and co-reference resolution (e.g., recognizing that "the man" and "he" refer to the same entity).
- Extraction: Employing OpenIE5 (Open Information Extraction) to identify (subject, relation, object) triples within sentences.
- Graph Construction: Entities become nodes, and the relations between them become directed edges. Additional edges are extracted from adjectives and prepositions.
- Refinement: Nodes referring to the same entity are merged, and labels are cleaned of superfluous words.
To normalize for speech volume, the resulting networks are analyzed using graph metrics z-scored against 1,000 random Erdős-Rényi graphs.
Results: The Fragmentation of Thought
The study utilized a general population cohort (N=436) to establish baseline properties and a clinical dataset (N=53) including patients with First Episode Psychosis (FEP), individuals at Clinical High Risk (CHR-P), and healthy controls. Participants performed both a spontaneous picture description task and a structured story recall task.
Speech is Non-Random
Semantic networks from the general population were found to be significantly more connected and possessed fewer, larger components than size-matched random networks, confirming that semantic speech structures are inherently organized.
Quantifying Disorganization
In psychosis, the clinical observation of "loosening of associations" manifests physically in the graph as fragmentation. FEP patients do not simply say less; their ideas form more "islands" that are smaller and disconnected from one another. This was particularly evident in the picture description task, where FEP patients produced networks with significantly fewer nodes ($z = 2.33, p = 0.02$) and edges ($z = 1.97, p = 0.048$) than healthy controls.
The CHR-P Gradient and Novel Signal
Individuals at clinical high risk (CHR-P) showed fragmentation values that sat squarely between healthy controls and FEP patients, suggesting a potential biomarker for disease progression. Furthermore, a clustering analysis showed that semantic network features (like component size) were distinct from traditional NLP markers like tangentiality or semantic coherence, suggesting netts captures a novel clinical signal.
Discussion and Future Directions
While promising, the clinical sample size (N=53) is relatively modest. Larger longitudinal studies are required to determine if these network features can accurately predict transition from high risk to full psychosis. Additionally, the tool is currently optimized for English; adapting it to other languages would require different underlying NLP models.
Future work could incorporate word embeddings to weight edges based on semantic similarity or analyze the temporal order of edge formation to measure how well a speaker "stays on topic" over time. By releasing netts as an open-source Python package, the authors provide a foundation for deeper automated phenotyping of formal thought disorder in clinical settings.
Summary updated