Vocal Acoustic Biomarkers of Depression Severity and Treatment Response

Abstract

Background

Valid, reliable biomarkers of depression severity and treatment response would provide new targets for clinical research. Noticeable differences in speech production between depressed and nondepressed patients have been suggested as a potential biomarker.

Methods

One hundred and five adults with Major Depression were recruited into a four-week, randomized, double-blind, placebo-controlled research methodology study. An exploratory objective of the study was to evaluate the generalizability and repeatability of prior study results indicating vocal acoustic properties in speech may serve as biomarkers for depression severity and response to treatment. Speech samples, collected at baseline and study end-point using an automated telephone system, were analyzed as a function of clinician-rated and patient-reported measures of depression severity and treatment response.

Results

Regression models of speech pattern changes associated with clinical outcomes in the prior study were found to be reliable and significant predictors of outcome in the current study despite differences in the methodological design and implementation of the two studies. Results of the current study replicate and support findings from the prior study. Clinical changes in depressive symptoms among patients responding to the treatments provided also reflected significant differences in speech production patterns. Depressed patients who did not improve clinically showed smaller vocal acoustic changes and/or changes that were directionally opposite to treatment responders.

Conclusions

This study supports the feasibility and validity of obtaining clinically important, biologically-based vocal acoustic measures of depression severity and treatment response using an automated telephone system. National Institutes of Health Clinical Trials Registry: http://clinicaltrials.gov Identifier: NCT00406952.

Introduction

Depression and anxiety disorders are critical public health and safety concerns throughout the developed and developing worlds (1). Depression interferes with the lives of roughly 1 of every 6 adults in the United States and is associated with more than half of all suicides (2, 3). Development of evidenced-based treatments requires reliable and valid measures of clinical severity and patient response. Objective measures of illness severity in many medical fields reflect direct measurements of physical characteristics, such as tumor volume or blood chemistry tests. Treatment outcome measures of mental disorders typically rely upon subjective judgments provided by the patient and/or a trained clinician, such as the Hamilton Depression Rating Scale (HAM-D) (4, 5), Montgomery-Asberg Depression Rating Scale (6), Inventory of Depressive Symptomatology (7), or Quick Inventory of Depressive Symptomatology (QIDS) (8). Results of mood and anxiety clinical trials have been inconsistent in recent years (9). Methodological and study design features, such as patient enrollment pressures, high severity criteria for study inclusion, rater expectancies, baseline score inflation, and unblinding of raters may contribute to inaccurate clinical assessments and unreliable study findings (10-15).

Development of an objective, non-invasive, physiologically-based biomarker of depression severity that is sensitive to clinical change associated with treatment could provide new avenues for clinical research and development of treatments. Biomarkers for diagnosis and assessment of depression have been sought for years (16), including investigations of neuroendocrine response to dexamethasone and corticotrophin releasing hormone (17), electroencephalographic (EEG) measures during sleep (18), cytokine plasma concentrations (19, 20), monoamine levels in cerebrospinal fluids (21), and brain-derived neurotrophic factor (22). Validation and use of a practical, easy-to-use biomarker for diagnostic testing and treatment response for depression remains elusive however (23).

Qualitative differences in speech produced by depressed and nondepressed patients have been recognized for many years (24, 25), but quantitative validation of speech parameters as indicators of depression severity and treatment response were not demonstrated until the 1970’s and ‘80’s (26, 27). Change in speech pause times and other acoustic measures were identified and offered as non-invasive biomarkers for “reducing clinical subjectivity in monitoring treatment progress” (28, 29). Replication and generalizability of these findings to non-English speaking depressed patients has been demonstrated (30), and the effects of clinical depression on measures of speech pitch have also been established (31).

Clinical improvement associated with antidepressant treatments was modeled using multivariate equations of voice acoustic parameters in the 1990’s (32-34), and efforts to further refine and validate vocal acoustic measures as biomarkers of central nervous system functioning continue, particularly with respect to depression (35-39). Use of vocal acoustic measures of depression severity and treatment response in clinical trials has been limited, due to perceived needs for specialized recording equipment and environments, software, and the technical skills needed to analyze the speech samples (40). Automated collection and analysis of speech samples using inexpensive, widely available devices, public-domain software, and high capacity data servers now addresses these perceived barriers. Speech samples obtained over telephones are adequate for extracting useful clinical information (41), and analysis of these speech samples can be streamlined using automated batch processing (42, 43).

In 2007, an interactive voice response (IVR) system was used to collect speech samples and computer-automated HAM-D measures of depression severity (44-47) in a naturalistic observational study (48). Significant correlations were found between depression severity and several vocal acoustic measures; patients who responded clinically to their treatments showed greater change in several speech measures than the patients who did not respond to treatment. Additionally, patients responding to their treatments showed significant within-subject changes of vocal acoustic measures between the beginning and end of treatment, while treatment nonresponders did not. This suggests that the changes in vocal acoustic measures during speech production reflect underlying neurophysiologic changes associated with clinical response evident by changes in symptoms of depression over time. Speech data from the prior study were subsequently analyzed using different, independently developed methods; results were consistent with previous findings (49).

The present study was a Phase 4 randomized, double-blind, placebo-controlled methodology study. The main objective was to evaluate new methodology for detecting antidepressant treatment response. Another study objective was to replicate findings from the 2007 vocal acoustic study (48) and assess the generalizability of vocal acoustic measures as potential biomarkers of depression severity and treatment response in a randomized clinical trial. Important methodological similarities and differences between the previous and current study are identified in Table 1

Table 1.

Methodological design comparison between previous and current study to evaluate repeatability and generalizability of vocal acoustic speech measures as biomarkers of depression severity and treatment response.

Study Design	Prior Study (Mundt et al 2007) (48)	Current Study (NCT00406952)
Treatment Indication	Major Depressive Disorder	Major Depressive Disorder
Study Design	Open-label treatment at physician’s discretion, naturalistic observation of treatment outcomes	Placebo-controlled, double-blind, randomized clinical trial. Sertraline vs. placebo, forced-titration dosing regimen.
Inclusion/Exclusion	Any patient beginning treatment for new major depressive episode.	Screening/baseline clinician-rated HAMD ≥ 22, extensive list of co-morbid exclusion criteria.
Treatment Duration	Six weeks	Four weeks
Primary Outcome Measure	Computer-administered HAM-D	Clinician-rated QIDS-C
Patient Recruitment (N)	Single US site (N = 33)	Eleven US sites (N = 105)

Methods and Materials

Eleven investigational sites across the United States screened 183 participants and randomized 165 to study treatments between November 2006 and August 2007. The randomized sample included 61 males and 104 females. Mean participant age was 37.8 years (SD = 12.5); 125 were White, 26 were Black, 4 were Asian, 10 reported their race as ‘Other.’ Inclusion criteria were: 18 to 65 years old; not currently taking psychotropic medications; a 17-item clinician-administered HAM-D score of 22 or greater at baseline; primary diagnosis of major depressive disorder based on Diagnostic and Statistical Manual of Mental Disorders, Fourth Edition (DSM-IV) criteria, with symptoms present for at least one month and willingness to provide informed consent as required by federal and institutional guidelines. Exclusion criteria were: Prior failure to respond to sertraline treatment or two or more antidepressant treatments of adequate dose and duration in the past 5 years; current diagnosis of generalized anxiety, social anxiety, obsessive compulsive, panic, post-traumatic stress, anorexia, bulimia, or substance abuse disorder within the past 6 months; current or past diagnosis of delirium, schizophrenia, bipolar, psychotic, dementia, or amnestic cognitive disorder.

Participants attended four office visits during the 4-week study period. Screening procedures confirmed study eligibility and baseline data were collected at the same visit (Week 0). Clinical assessments of depression severity and treatment response were obtained at the end of Weeks 1, 2, and 4. Double-blind, randomized assignment to a forced-titration dosing regimen started all participants on 50 mg/day of sertraline or placebo at baseline. At the end of Weeks 1 and 2 dosages were either (a) maintained if treatment efficacy was evident with minimal side effects, (b) increased by 50 mg/day to improve inadequate therapeutic response, or (c) decreased to reduce adverse events.

Clinical assessments of depression severity and treatment response at weeks 1, 2, and 4 included clinician-rated QIDS (QIDS-C) and HAM-D, and a paper self-report form of the QIDS (QIDS-SR). The QIDS is a 16-item scale; total scores ranging from 0 – 27, based on severity scores of nine symptom domains used as DSM-IV (50) diagnosis criteria. 17-item HAM-D scores range from 0 – 52. Baseline HAM-D scores were used to determine study eligibility, but were not used as the primary clinical outcome measure. Use of different clinical measures for study eligibility and treatment response reduces confounding of results by inflated baseline scores (14). Treatment response, for evaluation of the vocal acoustic measures, was defined as a 50% or greater reduction in the total symptom scores between baseline and study end-point. The a priori statistical analysis plan for evaluating the vocal acoustic data identified the QIDS-C as the primary clinical outcome measure of depression severity and treatment response.

At baseline, and after weeks 1 and 4, speech samples were collected using a standardized, automated telephone interface developed previously (48, 51). Speech samples consisted of (a) “Free Speech” descriptions of recent physical and mental experiences and their effects on functioning over the prior week; (b) “Automated Speech” produced by reciting the alphabet and counting from 1 to 20; (c) “Reading” the phonetically-balanced, 169-syllable, 129-word Grandfather Passage commonly to assess communication disorders; and (d) “Sustained Vowels” of /a/, /i/, /u/, and /ae/ held for approximately 5 seconds. Participants were instructed to speak in a natural manner at their typical speaking rate. Their speech was recorded over standard telephone lines as single-channel, 8-bit .wav files sampled at 8 kHz.

Analysis of the speech samples focused on motor timing measures, such as the duration and proportion of silences and vocalizations within the samples, and frequency domain measures, such as the mean and variability of the fundamental frequency (F₀) and first and second formants (F₁ and F₂) produced by acoustic resonances within the human vocal tract. More complete description of the data processing and analytic methods have been published elsewhere (52) and are provided in Supplement 1. The acoustic measures extracted from each speech sample type are listed in Table 2.

Table 2.

Vocal acoustic characteristics measured for each type of speech sample obtained.

VA Measures	Types of Speech Samples Collected Via IVR telephone system
	Free Speech	Automated Speech	Reading Passage	Held Vowels
Total recording time	√	√	√
Total vocalization time	√	√	√
Total pause time	√	√	√
Number of pauses	√	√	√
Mean pause length	√	√	√
Pause variability (SD)	√	√	√
Percent pause time	√	√	√
Speech/pause ratio	√	√	√
Speaking rate (syllables/sec)		√	√
Mean, SD, COV of F0	√	√	√	√
Mean, SD, COV of F1				√
Mean, SD, COV of F2				√

SD = Standard Deviation; COV – Coefficient of variation (ratio of SD of frequencies divided by the mean frequency)

F₀ = Fundamental frequency; F₁ = First Formant frequency; F₂ = Second Formant frequency;

Reported results were specified in a statistical analysis plan created prior to study completion and data lock. Statistical Package for the Social Sciences (SPSS Version 16) was used to conduct the analyses and an alpha value of p <.05 was specified for the planned comparisons. Between-group differences (e.g., responders versus nonresponders; sertraline versus placebo) were tested using two-group t-tests of means (53), and change over time was tested using paired mean t-tests (54). Tests of proportional equivalence in cross-tabulated frequencies used Fisher’s exact test (55), and classification agreement was tested using Cohen’s kappa (56), a measure of inter-rater agreement for categorical items that takes into account the degree of agreement that would occur due to chance alone.

Prior to the current study, logistic regression analyses of data obtained from the previous study (48) were conducted. Additional details regarding these analyses, and the calculation of response probabilities and likelihoods using data from the current study, are provided in Supplement 1. In the first analysis of the data from the prior study, baseline to Week 6 change (Δ) in vocal acoustic measures were computed. These change measures were entered as independent variables to model the differences between patients who responded to treatment (Responders; N = 13) and those who did not (Nonresponders; N = 19). Equation 1 provides the multivariate equation that maximized classification accuracy of the patients (91%) in the prior study data, χ² (7) = 25.4, p. <.001.

$$\begin{array}{l}\mathit{Log\; Odds\; Response}=­4.71-5.68(\mathrm{\Delta }\mathrm{Speaking\; rate})-214.1(\mathrm{\Delta }{\mathrm{F}}_0\mathrm{COV})+0.28\\ (\mathrm{\Delta }\mathrm{Speech}/\mathrm{pause\; ratio})-0.30(\mathrm{\Delta }\mathrm{Total\; recording\; time})+76.3(\mathrm{\Delta }\mathrm{Percent\; pause}\\ \mathrm{time})+0.42(\mathrm{\Delta }\mathrm{Total\; vocalization\; time})-18.2(\mathrm{\Delta }\mathrm{Pause\; variability})\end{array}$$ EQ. 1

A second analysis entered the vocal acoustic measures observed at baseline (only) as potential predictors of response available prior to receiving treatment. Equation 2 shows the best logistic regression model that significantly ‘predicted’ treatment response status at Week 6 in the prior study, χ² (2) = 6.9, p. =.032.

$$\mathit{Log\; Odds\; Response}=­2.45+0.06\left(\mathrm{Baseline\; Total\; recording\; length}\right)-62.93\left(\mathrm{Baseline}{\mathrm{F}}_0\mathrm{COV}\right)$$ EQ. 2

Logistic regression equations can be used to estimate the probability that a particular case belongs to one binary class or another (e.g., treatment responder or nonresponder) by applying Equation 3 (see Supplement 1).

$$\mathrm{Prob}(\mathrm{R})=({\mathrm{e}}^{\mathrm{Log\; Odds\; Response}}/(1+{\mathrm{e}}^{\mathrm{Log\; Odds\; Response}})).$$ EQ. 3

Generalizability of Equation 1 was evaluated by entering baseline to Week 4 vocal acoustic measure change – from the present study – and classifying participants with probability estimates of .5 or greater as ‘Vocal Acoustic Responders’. Equation 2 generalizability was evaluated by entering the baseline speech data – from the present study – and classifying participants with probability estimates of .5 or greater as ‘Likely Vocal Acoustic Responders’. These a priori classifications were compared with the observed clinical measures of treatment response or nonresponse from the present study (see Tables 5 and 6).

Table 5.

Repeatability and generalizability of treatment response^a probability in current study using baseline to end-point changes in vocal acoustic measures based on logistic regression model derived from prior study.

Clinical Treatment Outcome Measure (QIDS-C)b	Predicted by Logistic Regression Model (EQ. 1)
	Responder	Nonresponder
Responder	8	43
Nonresponder	1	52
(HAM-D-17)c
Responder	8	45
Nonresponder	1	50
(QIDS-SR)d
Responder	8	47
Nonresponder	1	49

^aResponder = 50% or greater reduction of depression severity from baseline to Week 4

^bTwo-sided Fisher exact test, p. = .015

^cTwo-sided Fisher exact test, p. = .031

^dTwo-sided Fisher exact test, p. = .033

Table 6.

Repeatability and generalizability of treatment response^a propensity in current study using baseline (only) vocal acoustic measures based on logistic regression model derived from prior study.

Clinical Treatment Outcome Measure (QIDS-C)b	Predicted by Logistic Regression Model (EQ. 2)
	Responder	Nonresponder
Responder	34	17
Nonresponder	25	29
(HAM-D-17)c
Responder	34	19
Nonresponder	25	27
(QIDS-SR)d
Responder	36	19
Nonresponder	23	28

^aResponder = 50% or greater reduction of depression severity from baseline to Week 4

^bTwo-sided Fisher exact test, p. = .049

^cTwo-sided Fisher exact test, p. = .117

^dTwo-sided Fisher exact test, p. = .050

Results

Of the 165 participants randomized to treatment, 39 discontinued prior to study completion, 20 were missing speech samples at baseline or end-point prohibiting meaningful analysis, and 1 subject was missing baseline QIDS-C data. Data from 105 evaluable participants were analyzed. The QIDS-C total depression scores were significantly correlated with both the HAM-D and QIDS-SR scores at baseline (r = .43 and .53; r² = .18 and .28) and at Week 4 (r = .87 and .81; r² = .76 and .66, respectively).

Fifty-five participants were randomized to sertraline treatment and fifty received placebo treatment. Means and standard deviations of the clinical depression measures for each treatment group at baseline and Week 4 are shown in Table 3. Significant clinical improvement from baseline to end-point is evident in both treatment groups across all three clinical outcome measures. Neither the QIDS-C nor the QIDS-SR found statistically significant mean score differences between the treatment groups, t(103) = 0.31 and 0.14 (ns), respectively. The mean improvement of 13.4 HAM-D points in the sertraline group is significantly greater than the 10.4 point improvement seen in the placebo-treated participants, t(103) = -2.21, p.<.05.

Table 3.

Means (± Standard deviations) of primary and secondary clinical measures of depression severity by treatment assignment.

Treatment	N	Depression Measure	Mean Score (± Standard Deviation)			Number of Participants
			Baseline	Week 4	Change	Respondersa	Nonresponders
Sertraline	55
		QIDS-C	16.4 (± 3.3)	8.5 (± 4.3)	-7.9 (± 4.9) ***	27	28
		HDRS-17	24.9 (± 2.8)	11.5 (± 5.8)	-13.4 (± 5.7) ***	33	22
		QIDS-SR	15.8 (± 3.7)	8.1 (± 4.7)	-7.7 (± 5.5) ***	29	26
Placebo	50
		QIDS-C	17.4 (± 2.8)	9.2 (± 4.6)	-8.2 (± 4.7) ***	24	26
		HDRS-17	24.6 (± 2.5)	13.9 (± 6.4)	-10.7 (± 6.6) ***	20	30
		QIDS-SR	16.6 (± 3.6)	8.8 (± 4.6)	-7.9 (± 4.3) ***	25	25

^***p. < .001

^aResponder = 50% or greater reduction of depression severity from baseline to Week 4

The proportion of treatment responders and nonresponders within each treatment group was not statistically different by either the QIDS-C or QIDS-SR measure, χ² = 0.12 and 0.08 (ns), respectively. With the HAM-D measure, the proportion of treatment responders in the sertraline group (60%) is greater than the proportion of placebo responders (40%), but was not statistically significant, χ² = 4.19, p. = .051.

Statistical inferences regarding sertraline/placebo differences are consistent across the clinical outcome measures. All three measures are consistent in classifying individual study participants as treatment Responders or Nonresponders based on a 50% or greater improvement from baseline. The QIDS-C and QIDS-SR agreed in 84 of the 105 cases (kappa = .60; p. < .001), the QIDS-C and HAM-D agreed in 87 of the 105 cases (kappa = .66; p. < .001), and the QIDS-SR and the HAM-D agreed in 84 of the 107 cases (kappa = .60; p. < .001).

Seven vocal acoustic measures were significantly correlated with depression severity in the prior study: F₂ Coefficient of Variation (COV) (r = -.17; r² = .03); total recording time (r = .20; r² = .04); total pause time (r = .29; r² = .08); pause variability (r = .38; r² = .14); percent pause time (r = .18; r² = .03); speech/pause ratio (r = -.22; r² = .05); and speaking rate (r = -.23; r² = .05). Six of these seven measures, obtained from the automated/reading speech samples in this study, correlated significantly with QIDS-C total scores. These were total recording time (r = .25; r² = .06), total pause time (r = .25; r² = .06), pause variability (r = .27; r² = .07), percent pause time (r = .20; r² = .04), speech/pause ratio (r = -.14; r² = .02), and speaking rate (r = -.18; r² = .03). Two of the seven measures obtained from the Free speech samples were significantly correlated with the QIDS-C: pause variability (r = .26; r² = .07) and percent pause time (r = .11; r² = .01). Two speech measures that were not significantly correlated with depression severity in the prior study were significantly correlated in the present study. These were total vocalization time (r =.13; r² = .02) and the number of pauses (r = .16; r² = .03) during the automated/reading tasks.

Table 3 shows 51 treatment Responders (27 sertraline, 24 placebo) and 54 Nonresponders (28 sertraline, 26 placebo) based on the primary outcome (QIDS-C). Table 4 compares ten vocal acoustic measures that found significantly different baseline to end-point changes separating the treatment Responder and Nonresponder groups. The statistical significance of the within-group changes over time are also shown. In all eight timing measures (recording lengths, vocalization durations, pause durations and variability, speaking rate), the within-subject change was significant in the treatment Responders. Among Nonresponders, only four showed significant change over time; for the measures that did show significant change, the changes over time are of smaller magnitudes and/or in opposite directions than those of the treatment Responders.

Table 4.

Means (± Standard Deviations) for significant vocal acoustic measure differences between treatment responders ^a and nonresponders defined by primary clinical outcome measure (QIDS-C).

Response Category	N	VA Measure	Mean (± Standard Deviation)			Responder - Nonresponder
			Baseline	Week 4	Change	Group t-test	df	p.
Respondersa	51
		Free Speech total recording time	120.4 (± 67.0)	94.4 (± 37.5)	-26.0 (±50.0)***
		Free Speech total vocalization time	68.5 (± 39.6)	58.0 (± 23.5)	-10.5 (±31.3) *
		Free Speech number of pauses	118.6 (± 73.5)	98.5 (± 53.3)	-20.1 (± 59.4) *
		Free Speech total pause time	51.9 (± 31.5)	36.4 (± 19.4)	-15.5 (± 25.9)***
		Free Speech pause variability	0.69 (± 0.25)	0.51 (± 0.15)	-0.18 (± 0.20)***
		Automatic Speech total recording time	30.2 (± 9.28)	23.2 (± 7.03)	-7.00 (± 9.45)***
		Automatic Speech total vocalization time	15.6 (± 3.87)	14.3 (± 3.28)	-1.26 (± 3.88) *
		Automatic Speech speaking rate	2.27 (± 0.70)	2.89 (± 0.85)	0.62 (± 0.82) *
		Mean F0 frequency	151.8 (± 36.5)	153.3 (± 35.7)	1.45 (± 7.25) NS
		Mean F1 frequency	546.8 (± 67.1)	558.2 (± 51.8)	11.4 (± 48.9) NS
Nonresponders	54
		Free Speech total recording time	98.0 (± 57.9)	107.0 (± 67.7)	8.93 (± 49.9) NS	3.58	103	.001
		Free Speech total vocalization time	51.3 (± 27.5)	61.6 (±36.8)	10.2 (± 29.3) *	3.50	103	.001
		Free Speech number of pauses	97.1 (± 58.7)	108.2 (±79.9)	10.3 (± 54.6) NS	2.72	102	.008
		Free Speech total pause time	46.7 (± 34.3)	46.3 (± 34.6)	-1.00 (± 27.7) NS	2.75	102	.007
		Free Speech pause variability	0.70 (± 0.26)	0.61 (± 0.20)	-0.09 (± 0.20) **	2.26	102	.026
		Automatic Speech total recording time	28.9 (± 7.87)	25.2(± 7.00)	-3.68 (± 6.90) ***	2.06	103	.042
		Automatic Speech total vocalization time	14.6 (± 3.30)	15.9 (± 3.56)	1.24 (± 4.38) *	3.10	103	.003
		Automatic Speech speaking rate	2.46 (± 1.02)	2.70 (± 0.88)	0.24 (± 0.95) NS	-2.19	103	.031
		Mean F0 frequency	157.3 (± 33.6)	155.7 (± 33.5)	-1.65 (± 7.95) NS	-2.09	103	.040
		Mean F1 frequency	570.5 (± 62.1)	556.6 (± 59.9)	-14.3 (± 46.3) *	-2.76	103	.007

^***p. ≤ .001

^**p. ≤ .01

^*p. ≤ .05

^NSp. > .05

^aResponder = 50% or greater reduction of depression severity on primary outcome (QIDS-C) from Baseline to Week 4

Equations 1 and 3, applied to individual patient’s baseline to Week 4 change in the vocal acoustic measures, were used to compute an expected probability that each participant would or would not respond to treatment. Patients with probability estimates of 0.5 or greater were classified as ‘Predicted Responders’ and the other patients were classified as ‘Predicted Nonresponders’. Similarly, Equations 2 and 3 were applied to the baseline speech measures for each participant to compute an a priori propensity for response. The probability estimates above and below 0.5 were used to classify each participant as a ‘Likely Responder’ or ‘Likely Nonresponder.’ While only nine participants had a predicted response probability of .5 or greater (See Table 5), the 51 treatment Responders did have significantly higher mean response probability estimates (Mean = 0.137; SD = 0.326) than the Nonresponders (Mean = .025; SD = 0.111), t(102) = -2.36, p. = .020. Similarly, response propensity estimates of the true Responders (Mean = 0.639; SD = 0.356) were significantly higher than the mean propensity estimate of the Nonresponders (Mean = 0.467; SD = 0.374), t(103) = -2.41, p. = .018.

Chi-square analysis of the cross-tabulated “Predicted Responder/Nonresponder” and “Likely Responder/Nonresponder” classifications with the “True Responder/Nonresponder” categories are shown in Tables 5 and 6. Both a priori logistic regression predictions of (a) treatment response probability based vocal acoustic changes using Equation 1, and (b) treatment response propensity using baseline speech measures using Equation 2, were statistically significant when QIDS-C criterion to define treatment response was examined. The statistical inferences that would have been drawn had the secondary outcomes measures been used to define treatment response are also presented.

Equation 4 provides the results of logistic regression modeling using baseline to Week 4 acoustic changes (Δ) as comparing the treatment Responders and Nonresponders observed in the present study. The model produces a test sensitivity estimate of 70.6% and specificity estimate of 79.2%, χ² (5) = 41.3, p. <.001. Equation 5 models Responder/Nonresponder differences of the baseline acoustic measures, and suggests a potential test sensitivity and specificity of 62.7% and 66.7% respectively, χ² (5) = 17.0, p. =.004. Since these models are maximized to “fit” the speech data obtained in this study, the repeatability/generalizability of these models requires independent evaluation in future research (like the tests of Equations 1 and 2 described above).

$$\begin{array}{l}\mathit{Log\; Odds\; Response}=­0.385-0.014(\mathrm{\Delta }\mathrm{Free\; speech\; total\; recording\; time})+\\ 0.832(\mathrm{\Delta }\mathrm{Free\; speech\; speech}/\mathrm{pause\; ratio})­0.277(\mathrm{\Delta }\mathrm{Auto\; speech\; vocalization}\\ \mathrm{time})+0.079(\mathrm{\Delta }\mathrm{Mean}{\mathrm{F}}_0\mathrm{frequency})+0.013(\mathrm{\Delta }\mathrm{Mean}{\mathrm{F}}_1\mathrm{frequency})\end{array}$$ EQ. 4

$$\begin{array}{l}\mathit{Log\; Odds\; response}=­3.854+0.014(\mathrm{Baseline\; Free\; speech\; number\; of\; pauses})-\\ 0.048(\mathrm{Baseline\; Free\; speech\; total\; pause\; length})+0.033(\mathrm{Baseline\; Free\; speech}\\ \mathrm{total\; vocalization\; time})+9.761(\mathrm{Baseline}{\mathrm{F}}_2\mathrm{COV})\end{array}$$ EQ. 5

Discussion

This is the first study to investigate vocal acoustic speech measures as biomarkers of depression severity and treatment response in a multisite double-blind, randomized clinical trial. The results are generally consistent with prior research and describe reliable methods for collecting and analyzing clinically meaningful speech data that are readily available to clinicians and researchers. This study replicates prior studies that have shown depressed patients produce longer speech pause times, which shorten with clinical improvement following treatment (26, 27, 57, 58). Other studies have also found significant reductions in total recording and vocalization times due to increased speaking rates following antidepressant treatment of depressed patients with psychomotor retardation (35, 48).

An important aspect of this study is the replication and support of findings from a much smaller pilot study (48), strengthening confidence in the results of both studies. This replication was found despite important methodological differences that could have, but did not, influence study outcomes. The two studies used different depression severity measures and different methods for administering the assessments. The prior study was an open-label, six-week observational design, whereas the current study was a double-blind, four-week, placebo-controlled clinical trial. Despite these important differences in study design and implementation, the results obtained were very consistent across both studies. Replication of results relating vocal acoustic measures with depression severity and response to treatment using different clinical scales, different modes of administering the assessments, different treatment durations, and different methods of blinding patients and providers provides strong evidence for generalizability of the results from both studies.

Six of seven correlations between depression severity and specific vocal acoustic measures found statistically significant in the prior study were also significant in the current study. More severe depression produces longer recordings with more pause time, more variable pause lengths, a greater percentage of pause time, smaller speech/pause ratios and slower speaking rates. This is consistent with clinical observations that depressed patients often present with psychomotor retardation, report difficulties with thinking and concentration, and have problems making decisions – such as choosing words (50). These symptoms manifest behaviorally during speech production (particularly during automated speech tasks) as objectively measurable biomarkers of depression. Treatment responders decrease total recording times during automated tasks primarily by increasing speaking rate and producing less overall vocalization. They also reduce the overall length of recordings and total vocalization, as well as the number, total time, and variability of pausing during free speech generation. The extent to which the behavioral manifestations of speech production evident in these vocal acoustic measures primarily reflects changes in patients’ interests or motivations to engage externally or reflect internal changes associated with concentration or information processing capacity, or both, remains to be determined.

This study found the mean fundamental and first formant frequencies increased or remained unchanged among treatment responders, and decreased or remained unchanged in nonresponders. This pattern was not found in the prior study suggesting replication is needed to assure that the findings are not spurious. Prior research in healthy adults, however, using the same measurement extraction methods have found these frequencies to remain stable over 4-week periods (59).

It is important that the change defining an individual’s clinical response to treatment is reflected in significant within-patient changes of the speech measures over time. The consistency of this relationship with treatment effectiveness, regardless of the nature of the treatment received (i.e., placebo or active drug), is also important. It suggests common neurophysiologic mechanisms influence symptoms of depression and the production of speech. All eight motor-based vocal acoustic measures distinguishing observed Responders from Nonresponders showed significant within-group change from baseline to end-point among treatment Responders, while only half did so in the Nonresponders. Of the speech measures that did show significant within-group change in the Nonresponders, the changes were of smaller magnitudes and/or directionally opposite to those seen in the Responders. The same pattern was seen in the prior study.

The statistical significance of the categorical predictions based on logistic models from just 32 participants of the prior study (13 Responders, 19 Nonresponders) applied to the present data from 105 randomized clinical trial participants is quite remarkable. The predictive validity of the models and consistent pattern of significant relationships between the clinical and acoustic measures suggests that the bio-behavioral linkage between speech production and depressive symptoms is robust. Although application of Equation 1 substantially underestimated the number of actual responders (see Table 5), the positive predictive value of the model was quite high (89%). Suboptimal coefficient estimations in the applied regression models could reflect the limited sample size of the prior study, modeling of 6-week (rather than 4-week) change in speech measures, or both. Ongoing studies continue to collect clinical and vocal acoustic data that will permit continued evaluation of the repeatability, validity, and extensions of the models applied to, and derived from, the present study. The present results are encouraging. Whether or not vocal acoustic measures of speech production might someday serve as objective biomarkers of depression severity for patient screening or monitoring of clinical progress during treatment, as suggested by others (60), or could be used to identify patients likely to be treatment responsive (or resistant), as suggested by the present study, remains to be demonstrated by further research.