Psycholinguistic Screening for Cognitive Decline in Cancer Survivors: A Feasibility Study

Abstract

Objectives.

Cancer survivors frequently experience treatment-related cognitive decline. No standard of care for cognitive assessment currently is recommended. Traditional neurocognitive testing is time consuming, expensive, and anxiety provoking. Because language requires complex cognitive processes and reflects cognition, we tested feasibility of using psycholinguistic speech analysis as a proxy for cognitive function in men undergoing prostate cancer treatment.

Sample and Setting.

Audio-recorded speech samples were collected from thirteen men enrolled in a parent study.

Methods and Variables.

Relationships between neurocognitive and psycholinguistic measures were evaluated in audio-recorded speech during the parent study clinical interviews and in response to an elicitation (prompt) question at two time points.

Results.

Significant correlations between neurocognitive and psycholinguistic metrics were identified for prompted speech although these relationships varied between time points. No relationships were identified in clinical interview speech samples.

Implications for Nursing.

Feasibility for recording, transcribing, and analyzing speech from cancer research interviews was demonstrated. Findings suggest relationships between neurocognitive and psycholinguistic measures in prompted speech. If validated, psycholinguistic assessment of prompted speech samples may be used to assess cognitive function. Advances in natural language processing may provide opportunities for future automated speech analyses for inexpensive, unobtrusive screening for cancer treatment-related cognitive decline.

Introduction

Prostate cancer primarily occurs in older men (mean age is 66 at diagnosis); thus, age-related comorbidities are common (Nguyen et al., 2015). Androgen deprivation therapy (ADT) is used to treat prostate cancer throughout the disease trajectory by reducing testosterone to castration levels. Forty-seven to 69% of men undergoing ADT experience treatment-related cognitive decline that may significantly impact their functioning, ability to live independently, and quality of life (Gonzalez et al., 2015; Nelson et al., 2008). With new interventions demonstrating potential for overcoming negative cognitive and other effects of ADT, early identification of treatment effects on cognition is increasingly important (Manson et al., 2019).

Currently, no standard of care exists for the assessment of cognitive change in cancer survivors. Existing methods include the use of self-report measures that may trigger follow up assessment using a complex battery of abstract neurocognitive tests. This testing is conducted by trained psychometricians over 2 or more hours and is typically disliked by older adults, raising their anxiety about cognitive abilities (Williams et al., 2014).

Because language production depends on intact function in a number of cognitive domains, spoken language reflects cognition that can be ascertained in everyday speech by measuring complexity of grammar and vocabulary and idea density in spoken language (Chen et al., 2009; Fernandes & Cairns, 2017). Psycholinguistics and neurolinguistics, scientific disciplines studying the psychological and neurobiological factors linking cognition with language production and the effects of neurological disorders on language, provide the conceptual basis for this study and for strategies designed to evaluate cognition as reflected in language (Fernandes & Cairns, 2017). Specific patterns of language in everyday conversations can differentiate conditions ranging from normal aging and mild cognitive impairment (Aramaki et al., 2016), to Alzheimer’s disease (Kemper et al., 2001), vascular dementia (Williams et al., 2003), and other neurodegenerative disorders. Recent studies have suggested that changes in speech may be one of the earliest signs of cognitive decline, offering insight about cognitive deficits years before mild cognitive impairment (Eyigoz et al., 2020) or dementia (Beltrami et al., 2018). Advances in natural language processing provide future potential for automated analysis of spoken language that may be used to assess cognition (Voleti et al., 2019).

Analysis of grammatical complexity and idea density of everyday conversations as well as narrative responses to a prompt have been utilized to characterize cognitive function and identify changes confirmed with neurocognitive testing (Kemper et al., 2001). However psycholinguistic measures have not yet been studied in the cancer survivor population. The goal of this study was to evaluate the feasibility of using psycholinguistic analyses to measure language complexity in men receiving ADT and to examine relationships between language complexity and traditional neurocognitive measures. A secondary aim was to compare use of psycholinguistic analysis between samples collected during clinical interviewing and those collected with prompted elicitation questions.

Approach

The parent study, “Staying Strong and Healthy for Androgen Deprivation Therapy for Men” (NR014518, S. Maliski, PI), is a randomized controlled trial of an intervention to minimize cardiovascular and metabolic risks of ADT through education, activity, and nutrition (Manson et al., 2019). We expanded data collection with the addition of self-reported and objective measures of neurocognitive function.

Recruitment

Following University IRB approvals, we recruited thirteen participants from the parent study. Inclusion criteria were: male; over the age of 21, diagnosed with prostate cancer; starting or had started ADT for prostate cancer within the last 3 months; re-initiating ADT after being on holiday for longer than their ADT dosage; able to speak and read English; reachable by telephone; and able to travel to the university for data collection.

Procedures

During 2018 and 2019 two speech samples were audio-recorded during two regularly scheduled parent study visits over a 6-month interval. One 5-minute recording was captured during the clinical interview for the study. Additionally, an elicitation (prompt) question designed to require thoughtful reflection were recorded . The elicitation questions, taken from established protocols, was randomly selected from the following: “What person, living or dead, famous or not, do you admire the most and why?” and “What is the most significant event you’ve experienced? It could be about the best thing that ever happened to your or the worst thing that ever happened.”

Four research assistants (RAs) were trained in transcription using archived practice audio recordings until a minimum of 90% agreement was achieved for five practice recordings for each psycholinguistic metric. Recordings were transcribed verbatim and segmented into utterances (sentences or sentence fragments) using standards for psycholinguistic analysis (Kemper et al, 2001). Each utterance was coded for the number and type of noun-verb clauses it contained. A simple sentence includes one main noun-verb clause. For example: “John (noun) called (verb) Mary.” Adding left and right-branching clauses reflects greater complexity of cognitive processes used to compose speech. For example: “Before he went (left-branching noun-verb clause) away, John called (main noun-verb clause) Mary to let her know (right-branching noun-verb clause) the plan.” Left-branching clauses are more cognitively demanding because the idea conveyed in the left-branching clause must be held in working memory while the remaining main clause is interpreted.

Coding was tabulated using the Systematic Analysis of Language Transcripts (SALT software LLC., Research Version 2012. Two measures of sentence length were also computed: (1) mean length of utterance in number of words (MLU) and (2) mean number of clauses per utterance (MCU) for main, left-branching (LCU), right-branching (RCU). Measures collected in the parent study included a battery of validated neurocognitive tests designed to measure the domains most likely affected by prostate cancer and cancer therapy. A summary of the psycholinguistic and neurocognitive assessments is listed in Table 1 (Dugbartey et al., 2000; Fieo et al., 2016; Strauss et al., 2006).

Table 1.

Psycholinguistic and Neurocognitive Assessments

Assessment	Description
Psycholinguistic grammatical complexity
Mean length of utterance	Average number of words per utterance
Mean clauses per utterance	Average number of clauses per utterance
Right-branching clauses	Clauses following the main clause
Left-branching clauses	Clauses preceding the main clause
Main clauses	Primary noun-verb utterance
Psycholinguistic semantic complexity
Type token ratio	Number of words compared to number of word roots (diversity of vocabulary)
Neurocognitive measures
Verbal memory	The World Health Organization/UCLA Auditory Verbal Learning Test involves a verbal presentation of a list of 15 words. Participants are asked to repeat back as many words as they can remember. The test includes 5 trials followed by introduction of an interference list of different words, immediate recall of the original list, delayed recall of the original list, and visual recognition for both lists.
Verbal fluency	In the Verbal Category Fluency Test participants are asked to produce as many words as possible within a prescribed category over one minute. Total sum of words produced for 3 categories (animals, foods, clothing) is calculated.
Executive function and processing speed	The timed Color Trail Making Tests (TMT) A and B assess participants’ ability to connect numbered circles in numerical order (A) and then alternating between numbered circles in two alternating colors (B). Higher scores on the TMT A and B indicate poorer performance.
Visuospatial ability	The Wechsler Adult Intelligence Scale IV Block Design Test is a timed test during which participants use 3-dimensional multicolored blocks to recreate geometric patterns.
Self-report of cognitive function	The Patient Reported Outcomes Measurement System (PROMIS) Applied Cognition, 8a, version 1, two short forms for 1) General Concerns and 2) Abilities related to cognitive function. Each instrument involves 8-items on a five-point scale. Participants are asked to rate both cognitive concerns and cognitive abilities during the past 7 days (Fieo et al., 2016).

Statistical analyses

Because of the small sample size for this feasibility study, nonparametric tests (Wilcoxon rank sum and Spearman correlations) were conducted with significance set at p = 0.05 level and reported at each time point. These results should be interpreted with caution due to the small sample size and potential violation of assumptions for the statistical tests.

Results

Sample Demographics.

Our feasibility study included data from 13 men ranging in age from 44 to 79 years (mean age = 66). Participants primarily were white (92%), in a relationship with a spouse or significant other (92%), and college educated (77%).

Correlations between psycholinguistic, neurocognitive, and self-reported cognition measures.

See Table 2 for the correlation matrix providing results for correlations between the neurocognitive and psycholinguistic measures. Only significant correlations in the prompted speech sample were identified and these varied between time 1 and time 2. No significant relationships between self-report measures of cognition and psycholinguistic measures were demonstrated.

Table 2.

Correlations between Grammatic Complexity and Neurocognitive Measures

RCU=Right-branching clauses per utterance; LCU=Left branching clauses per utterance; Color TMT A=Color Trail

	MLU		MCU		Main		LCU		RCU
	rs	p	rs	p	rs	p	rs	p	rs	p
Time 1
Prompt Question
Verbal Memory	.23	.45	.63	.02	.09	.77	.78	<.01	.61	.03
Verbal Fluency	.29	.34	.5	.09	.04	.89	.72	.01	.53	.06
Color TMT A	−.43	.14	−.32	.29	−.35	.24	−.2	.51	−.41	.17
Color TMT B	−.30	.34	−.62	.03	−.22	.48	−.56	.06	−.69	.01
Block Design	.12	.70	.42	.15	.22	.47	.32	.29	.38	.20
Clinical Interview
Verbal Memory	−.03	.87	−.06	.63	−.25	.92	.09	1.00	.07	.53
Verbal Fluency	.07	.29	.06	.30	−.03	.27	−.08	.95	−.02	.60
Color TMT A	−.36	.65	−.37	.19	−.27	.25	.18	.91	−.45	.25
Color TMT B	−.41	.08	−.37	.09	−.25	.14	.33	.98	−.42	.08
Block Design	−.11	.96	−.24	.88	−.29	.58	−.09	.32	−.2	.53
Time 2
Prompt Question
Verbal Memory	.10	.80	.19	.62	.26	.50	.30	.43	.13	.73
Verbal Fluency	.12	.77	.20	.61	.28	.46	.19	.62	.20	.61
Color TMT A	−.20	.61	−.33	.38	−.38	.31	−.63	.07	−.15	.70
Color TMT B	−.08	.83	−.15	.70	−.25	.52	−.28	.46	−.22	.58
Block Design	.80	.01	.83	.01	.85	<.01	.73	.03	.72	.03
Clinical Interview
Verbal Memory	.10	.80	.03	.95	.18	.65	−.26	.51	.13	.75
Verbal Fluency	.55	.12	.42	.26	.65	.06	−.01	.98	.25	.51
Color TMT A	.10	.80	−.20	.61	−.28	.46	−.10	.80	−.08	.83
Color TMT B	−.32	.41	−.52	.15	−.57	.11	−.35	.35	−.33	.38
Block Design	.10	.80	.22	.58	.03	.93	.52	.15	.15	.70

Note. MLU=Mean length of utterance; MCU=Mean clauses per utterance; Main=Main clauses per utterance; Color Trail Making Test Part A; Color TMT A=Color Trail Making Test Part B; Block Design =Weschler Adult Intelligence (WAIS) IV Block Design.

Comparison of clinical interview and elicitation question (prompt).

We compared findings from recordings collected as part of the study clinical interview and those responding an elicitation (prompt) question using the Wilcoxon signed rank test. Prompt and interview speech differed for MLU (W = 62, p = 0.04), LCU (W = 60.5, p = 0.03), and for complete utterances (W = 26.5, p = 0.02). However, no significant correlations between neurocognitive and psycholinguistic measures were demonstrated in the study interview speech sample (Table 2).

Discussion

This study established feasibility for enrolling men receiving ADT and collecting language samples during routine cancer research study visits. Transcription and psycholinguistic coding were successful, reaching reliability of 90% or greater agreement for utterance segmentation and coding of clauses by graduate research assistants. We identified a number of significant correlations but acknowledge that findings from this small feasibility study need to be validated in future research with larger and more diverse samples.

A number of significant correlations were identified between the neurocognitive and psycholinguistic measures in the prompted speech samples although these varied somewhat between time 1 and time 2. No relationships between semantic and grammatical complexity with self-reported measures of cognition were identified. This may be due to our small sample of 13 participants assessed at two time points.

At time 1, verbal memory was correlated with grammatical complexity measures mean clauses per utterance (MCU; r_s = .63, p = .02), and left and right-branching clauses per utterance (r_s = <.78, p = <.01 for LCU and r_s = .61, p = .03 for RCU). Verbal fluency also correlated with left-branching clauses per utterance (r_s = .72, p = .01). We interpret these results to show that increasing grammatical complexity reflected in more complex utterances mirrored higher scores for memory functions as assessed in the neurocognitive tests. Timed color trail-making test (TMT) B was negatively correlated with MCU, and RCU, two measures of grammatical complexity (r_s = <−.62, p = <.03, r_s = −56, p = .03). The TMT B is a measure of executive function and processing speed with higher scores indicating increased time needed to complete the test tasks. Higher language complexity correlated with poorer performance in executive function and processing speed domains in our sample.

At time 2, block design, a measure of visuospatial ability correlated with all of the psycholinguistic measures including mean length of utterances in words (MLU), mean clauses per utterance (MCU) and main, left, and right-branching clauses per utterance (see results in table). These positive correlations of scores on neurocognitive and psycholinguistic measures are all in the anticipated direction (higher performance on neurocognitive test correlates with more complex language). The reason for different measure correlations at two time points is not clear and should be explored further.

A number of psycholinguistic and neurocognitive measures were significantly correlated, but in the prompt (elicitation question) data only. The lack of correlation with the interview speech suggests that these two types of speech vary, and that prompted speech provides greater sensitivity to cognitive function. Prompted speech requires the participant to think about and compose a response in an answer to an open-ended question. Thus, prompted speech narratives rely more on cognitive and language processes. In comparison, the responses to questions in the research study interview are likely brief and frequently include yes and no responses to specific, more finite questions. Because brief responses in the research study interview are less dependent on cognitive processes, they may be less reflective of cognitive processes.

This feasibility study provided preliminary results for psycholinguistic analysis of conversation to measure cognition during cancer-related healthcare visits. If validated as a reliable and sensitive measure in cancer survivors, psycholinguistic analyses of speech samples could be compared over time to detect changes due to treatment effects, signaling a need for neurocognitive evaluation and potential interventions. If psycholinguistic analysis proves sensitive for identifying changes in cognition, an application may be developed at a low cost and readily available across care settings.

Implications for Nursing

No standard of care currently is accepted to assess cognitive changes in cancer survivors. Issues exist for both self-report measures and neurocognitive testing. Awareness of available assessment and screening tools and those under investigation is critical to nurses caring for cancer survivors. If relationships can be established between neurocognitive and psycholinguistic measures, psycholinguistic analyses may be used to assess cognitive function. Although the process of recording, transcribing and coding speech samples is laborious and not feasible at the point of care, advances in automated transcription and natural language processing may provide potential tools for cognitive assessment in the near future (Beltrami et al., 2018). A speech analysis app on a cell phone or other mobile device would be readily scalable in rural and other underserved areas where access to advanced neurocognitive testing is limited, improving care and could potentially be provided in telehealth visits. Using speech analyses to measure cognition, grounded in theories of psycholinguistics and neurolinguistics (Fernandes and Cairns, 2017) has potential to fill a gap to improve care by providing a new tool for assessment of cognitive function. Identification of cognitive decline is the first step in initiating evidence-based interventions to prevent or minimize cognitive effects of cancer therapies and to initiate decision making about therapy.

Conclusion

If relationships suggested in this feasibility study are validated in ongoing research, psycholinguistic analyses may be applied to assess cognitive function and screen for cognitive changes, avoiding the need for time-consuming and stressful neurocognitive testing in persons receiving cancer treatment.

Availability of data and material.

Interested parties may contact the corresponding author concerning sharing of deidentified study data that may be made available pending IRB and data sharing approvals.