User login
In the first study, an algorithm was created to analyze speech patterns and semantic content to create novel “speech networks.” Compared with their healthy peers, patients with FEP had smaller and more fragmented networks. At-risk individuals had fragmented values that were in between those of the FEP and healthy control groups.
“This suggests that semantic speech networks can enable deeper phenotyping of formal thought disorder and psychosis,” said lead author Caroline Nettekoven, PhD, department of psychiatry, University of Cambridge, England.
In the second study, Janna N. de Boer, MD, University of Groningen, the Netherlands, and colleagues examined patients with FEP who did and did not experience relapse after 24 months of follow-up.
An algorithm based on natural language processing (NLP) of speech recordings predicted the relapses with an accuracy of more than 80%.
NLP “is a powerful tool with high potential for clinical application and diagnosis and differentiation, given its ease in acquirement, low cost, and naturally low patient burden,” said de Boer.
The findings for both studies were presented at the annual congress of the Schizophrenia International Research Society.
Fragmented networks
Dr. Nettekoven noted that previous research has shown “mapping the speech of a psychosis patient as a network and analyzing the network using graph theory is useful for understanding formal thought disorder.”
However, these tools ignore the semantic content of speech, which is a “key feature” that is altered in psychotic language, she added.
The researchers therefore proposed a “novel type of network to map the content of speech.”
For example, if someone said, “I see a man,” a semantic speech network developed from this sentence would have the first and last words connected by “the edge” to the word “see,” Dr. Nettekoven explained.
To explore further, the investigators developed an algorithm known as “netts” that automatically creates semantic speech networks from transcribed speech.
They first applied the algorithm to transcribed speech from a general population sample of 436 individuals and then to a clinical sample (n = 53) comprising patients with FEP, those at clinical high risk for psychosis, and a healthy control group.
Comparing the general population sample with randomly generated semantic speech networks, the investigators found that networks from the general population had fewer but larger connected components, which “reflects the nonrandom nature of speech,” said Dr. Nettekoven.
In the clinical sample, networks from the FEP group had a significantly higher number of connected components compared with the healthy control group (P = .05) and a significantly smaller median connected-component size (P < .01).
“So patients’ mental speech networks are more fragmented than those from controls,” said Dr. Nettekoven. She added that the networks from clinically high-risk individuals “showed fragmentation values in between [those of] patients and controls.”
A further clustering analysis suggested the semantic speech networks “capture a novel signal that is not already described” by other NLP measures, Dr. Nettekoven said. In addition, the network features were related to negative symptom scores and scores on the Thought and Language Index.
However, Dr. Nettekoven noted that these relationships “did not survive correcting for multiple comparisons.”
Relapse predictor
During her presentation of the second study, Dr. de Boer said that “predicting relapse remains challenging” in FEP.
However, she noted that recent developments in NLP have proved to be effective in a “range of applications,” including early symptom recognition and differential diagnosis in psychosis.
To determine whether NLP could help predict relapse, the study included 104 patients aged 16-55 years with FEP whose conditions had been in remission for 3-6 months. Speech recordings were made at baseline and after 3 and 6 months and were analyzed via OpenSMILE software.
After a follow-up of 24 months, 24 of the patients remaining in the study had not experienced relapse, while 21 patients had experienced relapse. There were no significant age, education, or gender differences between those who did and those who did not experience relapse.
On the basis of speech analysis, the investigators identified a machine learning classifier, which showed an accuracy of 80.8% in predicting relapse 3 months in advance of the occurrence.
‘Valid and informative’
Commenting on the studies, Eric J. Tan, PhD, Centre for Mental Health, Swinburne University of Technology, Melbourne, said they are “but two of a variety of ways in which speech can be analyzed and are both equally valid and informative.”
The key takeaway “is that both studies are examples of the ways in which speech can be used clinically, such as for predicting relapse and for the potential proxy measure for the assessment of symptom severity,” said Dr. Tan, who was not involved with the research.
The studies also show that “speech is sensitive to different stages of the disorder, as well as its individual symptoms,” he added.
However, Dr. Tan noted that although “speech may be more of a sign of an underlying pathology or dysfunction, given that it waxes and wanes with illness severity, more analyses are needed before drawing definitive conclusions.” This is especially needed “given the relative infancy of quantitative speech analysis,” he said.
“It would also be useful to conduct these analyses across a variety of different languages to look for commonalities and differences that will help shed light on the variables most closely linked to the disorder,” Dr. Tan concluded.
The investigators have reported no relevant financial relationships. Dr. Tan has received an Early Career Research Fellowship from the National Health and Medical Research Council of Australia.
A version of this article first appeared on Medscape.com.
In the first study, an algorithm was created to analyze speech patterns and semantic content to create novel “speech networks.” Compared with their healthy peers, patients with FEP had smaller and more fragmented networks. At-risk individuals had fragmented values that were in between those of the FEP and healthy control groups.
“This suggests that semantic speech networks can enable deeper phenotyping of formal thought disorder and psychosis,” said lead author Caroline Nettekoven, PhD, department of psychiatry, University of Cambridge, England.
In the second study, Janna N. de Boer, MD, University of Groningen, the Netherlands, and colleagues examined patients with FEP who did and did not experience relapse after 24 months of follow-up.
An algorithm based on natural language processing (NLP) of speech recordings predicted the relapses with an accuracy of more than 80%.
NLP “is a powerful tool with high potential for clinical application and diagnosis and differentiation, given its ease in acquirement, low cost, and naturally low patient burden,” said de Boer.
The findings for both studies were presented at the annual congress of the Schizophrenia International Research Society.
Fragmented networks
Dr. Nettekoven noted that previous research has shown “mapping the speech of a psychosis patient as a network and analyzing the network using graph theory is useful for understanding formal thought disorder.”
However, these tools ignore the semantic content of speech, which is a “key feature” that is altered in psychotic language, she added.
The researchers therefore proposed a “novel type of network to map the content of speech.”
For example, if someone said, “I see a man,” a semantic speech network developed from this sentence would have the first and last words connected by “the edge” to the word “see,” Dr. Nettekoven explained.
To explore further, the investigators developed an algorithm known as “netts” that automatically creates semantic speech networks from transcribed speech.
They first applied the algorithm to transcribed speech from a general population sample of 436 individuals and then to a clinical sample (n = 53) comprising patients with FEP, those at clinical high risk for psychosis, and a healthy control group.
Comparing the general population sample with randomly generated semantic speech networks, the investigators found that networks from the general population had fewer but larger connected components, which “reflects the nonrandom nature of speech,” said Dr. Nettekoven.
In the clinical sample, networks from the FEP group had a significantly higher number of connected components compared with the healthy control group (P = .05) and a significantly smaller median connected-component size (P < .01).
“So patients’ mental speech networks are more fragmented than those from controls,” said Dr. Nettekoven. She added that the networks from clinically high-risk individuals “showed fragmentation values in between [those of] patients and controls.”
A further clustering analysis suggested the semantic speech networks “capture a novel signal that is not already described” by other NLP measures, Dr. Nettekoven said. In addition, the network features were related to negative symptom scores and scores on the Thought and Language Index.
However, Dr. Nettekoven noted that these relationships “did not survive correcting for multiple comparisons.”
Relapse predictor
During her presentation of the second study, Dr. de Boer said that “predicting relapse remains challenging” in FEP.
However, she noted that recent developments in NLP have proved to be effective in a “range of applications,” including early symptom recognition and differential diagnosis in psychosis.
To determine whether NLP could help predict relapse, the study included 104 patients aged 16-55 years with FEP whose conditions had been in remission for 3-6 months. Speech recordings were made at baseline and after 3 and 6 months and were analyzed via OpenSMILE software.
After a follow-up of 24 months, 24 of the patients remaining in the study had not experienced relapse, while 21 patients had experienced relapse. There were no significant age, education, or gender differences between those who did and those who did not experience relapse.
On the basis of speech analysis, the investigators identified a machine learning classifier, which showed an accuracy of 80.8% in predicting relapse 3 months in advance of the occurrence.
‘Valid and informative’
Commenting on the studies, Eric J. Tan, PhD, Centre for Mental Health, Swinburne University of Technology, Melbourne, said they are “but two of a variety of ways in which speech can be analyzed and are both equally valid and informative.”
The key takeaway “is that both studies are examples of the ways in which speech can be used clinically, such as for predicting relapse and for the potential proxy measure for the assessment of symptom severity,” said Dr. Tan, who was not involved with the research.
The studies also show that “speech is sensitive to different stages of the disorder, as well as its individual symptoms,” he added.
However, Dr. Tan noted that although “speech may be more of a sign of an underlying pathology or dysfunction, given that it waxes and wanes with illness severity, more analyses are needed before drawing definitive conclusions.” This is especially needed “given the relative infancy of quantitative speech analysis,” he said.
“It would also be useful to conduct these analyses across a variety of different languages to look for commonalities and differences that will help shed light on the variables most closely linked to the disorder,” Dr. Tan concluded.
The investigators have reported no relevant financial relationships. Dr. Tan has received an Early Career Research Fellowship from the National Health and Medical Research Council of Australia.
A version of this article first appeared on Medscape.com.
In the first study, an algorithm was created to analyze speech patterns and semantic content to create novel “speech networks.” Compared with their healthy peers, patients with FEP had smaller and more fragmented networks. At-risk individuals had fragmented values that were in between those of the FEP and healthy control groups.
“This suggests that semantic speech networks can enable deeper phenotyping of formal thought disorder and psychosis,” said lead author Caroline Nettekoven, PhD, department of psychiatry, University of Cambridge, England.
In the second study, Janna N. de Boer, MD, University of Groningen, the Netherlands, and colleagues examined patients with FEP who did and did not experience relapse after 24 months of follow-up.
An algorithm based on natural language processing (NLP) of speech recordings predicted the relapses with an accuracy of more than 80%.
NLP “is a powerful tool with high potential for clinical application and diagnosis and differentiation, given its ease in acquirement, low cost, and naturally low patient burden,” said de Boer.
The findings for both studies were presented at the annual congress of the Schizophrenia International Research Society.
Fragmented networks
Dr. Nettekoven noted that previous research has shown “mapping the speech of a psychosis patient as a network and analyzing the network using graph theory is useful for understanding formal thought disorder.”
However, these tools ignore the semantic content of speech, which is a “key feature” that is altered in psychotic language, she added.
The researchers therefore proposed a “novel type of network to map the content of speech.”
For example, if someone said, “I see a man,” a semantic speech network developed from this sentence would have the first and last words connected by “the edge” to the word “see,” Dr. Nettekoven explained.
To explore further, the investigators developed an algorithm known as “netts” that automatically creates semantic speech networks from transcribed speech.
They first applied the algorithm to transcribed speech from a general population sample of 436 individuals and then to a clinical sample (n = 53) comprising patients with FEP, those at clinical high risk for psychosis, and a healthy control group.
Comparing the general population sample with randomly generated semantic speech networks, the investigators found that networks from the general population had fewer but larger connected components, which “reflects the nonrandom nature of speech,” said Dr. Nettekoven.
In the clinical sample, networks from the FEP group had a significantly higher number of connected components compared with the healthy control group (P = .05) and a significantly smaller median connected-component size (P < .01).
“So patients’ mental speech networks are more fragmented than those from controls,” said Dr. Nettekoven. She added that the networks from clinically high-risk individuals “showed fragmentation values in between [those of] patients and controls.”
A further clustering analysis suggested the semantic speech networks “capture a novel signal that is not already described” by other NLP measures, Dr. Nettekoven said. In addition, the network features were related to negative symptom scores and scores on the Thought and Language Index.
However, Dr. Nettekoven noted that these relationships “did not survive correcting for multiple comparisons.”
Relapse predictor
During her presentation of the second study, Dr. de Boer said that “predicting relapse remains challenging” in FEP.
However, she noted that recent developments in NLP have proved to be effective in a “range of applications,” including early symptom recognition and differential diagnosis in psychosis.
To determine whether NLP could help predict relapse, the study included 104 patients aged 16-55 years with FEP whose conditions had been in remission for 3-6 months. Speech recordings were made at baseline and after 3 and 6 months and were analyzed via OpenSMILE software.
After a follow-up of 24 months, 24 of the patients remaining in the study had not experienced relapse, while 21 patients had experienced relapse. There were no significant age, education, or gender differences between those who did and those who did not experience relapse.
On the basis of speech analysis, the investigators identified a machine learning classifier, which showed an accuracy of 80.8% in predicting relapse 3 months in advance of the occurrence.
‘Valid and informative’
Commenting on the studies, Eric J. Tan, PhD, Centre for Mental Health, Swinburne University of Technology, Melbourne, said they are “but two of a variety of ways in which speech can be analyzed and are both equally valid and informative.”
The key takeaway “is that both studies are examples of the ways in which speech can be used clinically, such as for predicting relapse and for the potential proxy measure for the assessment of symptom severity,” said Dr. Tan, who was not involved with the research.
The studies also show that “speech is sensitive to different stages of the disorder, as well as its individual symptoms,” he added.
However, Dr. Tan noted that although “speech may be more of a sign of an underlying pathology or dysfunction, given that it waxes and wanes with illness severity, more analyses are needed before drawing definitive conclusions.” This is especially needed “given the relative infancy of quantitative speech analysis,” he said.
“It would also be useful to conduct these analyses across a variety of different languages to look for commonalities and differences that will help shed light on the variables most closely linked to the disorder,” Dr. Tan concluded.
The investigators have reported no relevant financial relationships. Dr. Tan has received an Early Career Research Fellowship from the National Health and Medical Research Council of Australia.
A version of this article first appeared on Medscape.com.
FROM SIRS 2022