Using Machine Learning to Predict Schizophrenia Diagnosis via Speech Connectedness | Schemes and Mind Maps Psychiatry

Quantifying ‘word salad’: The structural randomness of verbal reports predicts

negative symptoms and Schizophrenia diagnosis 6 months later

- Thought disorder measured as random speech structure -

Natália B. Mota (1), Mauro Copelli (2), Sidarta Ribeiro (1)

((1) Brain Institute, Federal University of Rio Grande do Norte – UFRN, (2) Department of Physics,

Federal University of Pernambuco – UFPE)

Background: The precise quantification of negative symptoms is necessary to improve differential

diagnosis and prognosis prediction in Schizophrenia. In chronic psychotic patients, the representation

of verbal reports as word graphs provides automated sorting of schizophrenia, bipolar disorder and

control groups based on the degree of speech connectedness. Here we aim to use machine learning to

verify whether speech connectedness during first clinical contact can predict negative symptoms and

Schizophrenia diagnosis six months later.

Methods: PANSS scores and memory reports were collected from 21 patients undergoing first clinical

contact for recent-onset psychosis and followed for 6 months to establish DSM-IV diagnosis, and 21

healthy controls. Each report was represented as a graph in which words corresponded to nodes, and

node temporal succession corresponded to edges. Three connectedness attributes were extracted from

each graph, z-scores to random graph distributions were measured, correlated with the PANSS

negative subscale, combined into a single Fragmentation Index, and used for predictions.

Findings: Random-like speech was prevalent among Schizophrenia patients (64% x 5% in Control

group, p=0.0002). Connectedness explained 92% of the PANSS negative subscale variance

(p=0.0001). The Fragmentation Index classified low versus high scores of PANSS negative subscale

with 93% accuracy (AUC=1), predicted Schizophrenia diagnosis with 89% accuracy (AUC=0.89), and

was validated in an independent cohort of chronic psychotic patients.

Interpretation: The structural randomness of speech graph connectedness is increased in

Schizophrenia. It provides a quantitative measurement of “word salad” as a Fragmentation Index that

tightly correlates with negative symptoms and predicts Schizophrenia diagnosis during first clinical

contact of recent-onset psychosis.

Address correspondence to Natália Bezerra Mota,

Brain Institute - Federal University of Rio Grande do Norte (UFRN): Av Nascimento de Castro 2155,

CEP: 59056-450, Natal, RN, Brazil; Phone: +55 84 32152708

e-mail: nataliamota@neuro.ufrn.br;

Using Machine Learning to Predict Schizophrenia Diagnosis via Speech Connectedness, Schemes and Mind Maps of Psychiatry

Related documents

Partial preview of the text

Download Using Machine Learning to Predict Schizophrenia Diagnosis via Speech Connectedness and more Schemes and Mind Maps Psychiatry in PDF only on Docsity!

Quantifying ‘word salad’: The structural randomness of verbal reports predicts

negative symptoms and Schizophrenia diagnosis 6 months later

- Thought disorder measured as random speech structure -

Natália B. Mota (1), Mauro Copelli (2), Sidarta Ribeiro (1)

((1) Brain Institute, Federal University of Rio Grande do Norte – UFRN, (2) Department of Physics,

Federal University of Pernambuco – UFPE)

Background: The precise quantification of negative symptoms is necessary to improve differential

diagnosis and prognosis prediction in Schizophrenia. In chronic psychotic patients, the representation

of verbal reports as word graphs provides automated sorting of schizophrenia, bipolar disorder and

control groups based on the degree of speech connectedness. Here we aim to use machine learning to

verify whether speech connectedness during first clinical contact can predict negative symptoms and

Schizophrenia diagnosis six months later.

Methods: PANSS scores and memory reports were collected from 21 patients undergoing first clinical

contact for recent-onset psychosis and followed for 6 months to establish DSM-IV diagnosis, and 21

healthy controls. Each report was represented as a graph in which words corresponded to nodes, and

node temporal succession corresponded to edges. Three connectedness attributes were extracted from

each graph, z-scores to random graph distributions were measured, correlated with the PANSS

negative subscale, combined into a single Fragmentation Index, and used for predictions.

Findings: Random-like speech was prevalent among Schizophrenia patients (64% x 5% in Control

group, p=0.0002). Connectedness explained 92% of the PANSS negative subscale variance

(p=0.0001). The Fragmentation Index classified low versus high scores of PANSS negative subscale

with 93% accuracy (AUC=1), predicted Schizophrenia diagnosis with 89% accuracy (AUC=0.89), and

was validated in an independent cohort of chronic psychotic patients.

Interpretation: The structural randomness of speech graph connectedness is increased in

Schizophrenia. It provides a quantitative measurement of “word salad” as a Fragmentation Index that

tightly correlates with negative symptoms and predicts Schizophrenia diagnosis during first clinical

contact of recent-onset psychosis.

Address correspondence to Natália Bezerra Mota,

Brain Institute - Federal University of Rio Grande do Norte (UFRN): Av Nascimento de Castro 2155,

CEP: 59056-450, Natal, RN, Brazil; Phone: +55 84 32152708

e-mail: nataliamota@neuro.ufrn.br;

Background

Schizophrenia is associated with negative symptoms and poor prognosis 1. In particular,

elevated negative symptoms are associated with lower rates of recovery 1,2. Formal thought disorder -

which comprises poverty of speech, derailment, and incoherence - constitutes an important set of

psychotic symptoms, and negative formal thought disorder is associated with the Schizophrenia

diagnosis, even during first episode psychosis 2,3. The early stages of the disease constitute a critical

opportunity for prevention of major cognitive damage 4.

Improved behavioral measures subjected to novel mathematical analyses are emerging as part of

a new field that uses computational tools to better characterize psychiatric phenomena (computational

phenotyping) 5-13. In particular, verbal report quantification by graph tools predicts diagnosis 9,10^ and

makes precise and automated quantification of speech features related with negative symptoms 9. By

representing each word as a node and the temporal sequence of consecutive words as directed edges, it

is possible to calculate attributes that characterize graph structure. The assessment of dream reports

from chronic psychotic patients has shown that word connectedness (number of edges between words,

or amount of nodes in connected components) is negatively correlated with negative symptoms 9. The

same graph attributes calculated from short-term memory reports from healthy children are positively

correlated with Intelligence Quotient and Theory of Mind scores, and can predict academic

performance independently of other cognitive measures 14. Altogether, the data suggest that word

connectedness measured by graph analysis rises during healthy development, but not during the course

of Schizophrenia. It is therefore possible that early markers of speech disorganization during recent-

onset psychosis, such as decreased connectedness, may be able to predict the severity of negative

symptoms as well as the Schizophrenia diagnosis.

Our hypotheses were: 1. Speech connectedness from dream reports 9 and short-term memory

reports 14 can predict the Schizophrenia diagnosis ; 2. Patients in the Schizophrenia group will produce

verbal reports less connected and more similar to random connectedness then those from other groups;

3. Connectedness attributes negatively correlate with negative symptoms measured by PANSS 9 ; 4. A

single index combining connectedness attributes correlated with negative symptoms should improve

Schizophrenia diagnosis and prediction of negative symptom severity.

Graph measures:

The search for a predictive index of connectedness was exploratory, and for that we tested 6 different

kinds of memory reports, searching for the best combination of connectedness attributes. Memory

reports were transcribed and represented as graphs in which each word was represented as a node, and

the temporal sequence between consecutive words was represented by directed edges (Figure 1A)

using the software SpeechGraphs (http://www.neuro.ufrn.br/softwares/speechgraphs) 9. Three

connectedness attributes were calculated: Edges (E), which measures the amount of links between

words; largest connected component (LCC), which measures the amount of nodes in the largest

component in which each pair of nodes has a path between them; and largest strongly connected

component (LSC), which counts the amount of nodes in the largest component in which each pair of

nodes has a mutually reachable path, i.e., node “a” reaches node “b” and node “b” reaches node “a”

(Figure 1A).

Analyses:

Since formal thought disorder symptoms are frequently described as „word salad‟, i.e. as random-like

speech, we compared each memory report graph to 1,000 random graphs built with the same nodes and

number of edges, but with a random shuffling of the edges that amounts to shuffling word (Figure 1B).

Next we estimated the LCC and LSC z-scores between each original graph and the corresponding

+ LSCz_negative(−2.4) + E_dream (−0.23) + LCC_dream * 0.