The effect of growth hormone-releasing hormone on cognition and brain connectivity in adults with cognition ranging from normal to mild cognitive impairment

Abstract

Cognitive decline with age and other clinical conditions are linked with reduced hypothalamic-pituitary-adrenal (HPA) axis function. Stimulating the HPA axis with supplemental growth hormone (GH) treatment can improve cognition, however, potential direct effects of stimulation with growth hormone releasing hormone (GHRH) are not established. In a double-blind, placebo-controlled pilot trial, we assessed 22 subjects with baseline cognition ranging from normal cognition to mild cognitive impairment before and after 10 weeks of treatment with low-dose tesamorelin (1 mg; GHRH analog) or placebo. We compared groupwise changes in body composition, fatigue, sleep, physical performance, glucose tolerance, cognitive function, and brain morphometry and functional connectivity. Low-dose GHRH treatment was not directly linked with significant changes in study measures. Using advanced machine learning (ML) models to further examine the data we identified potential treatment-related differences in areas of the brain related to cognitive function including the right anterior cingulate and left superior frontal occipital fasciculus. This pilot study highlights the potential benefits of pairing cognitive tests and neuroimaging with ML tools to achieve greater sensitivity for treatment-related effects. The clinical trial registration number is: NCT02553603

Graphical abstract

Highlights

• •GHRH may improve cognitive performance in healthy and cognitively impaired subjects.
• •GHRH may alter brain connectivity and morphometry in areas related to cognition.
• •Delayed Word Recall could serve as a sensitive indicator for early stages of cognitive deficits.
• •Machine learning analysis could potentially be utilized to identify biomarkers for cognitive deficits.

1. Introduction

The hypothalamic-pituitary-adrenal (HPA) axis plays an important role in aging, including impacts on frailty and cognition [1], [2]. Somatopause, or age-related decline in circulating GH and IGF-1 levels, is associated with detrimental physiologic effects of aging including decreased lean mass and bone mineral density, and increased adiposity [1], [2], [3]. Somatopause also has significant impacts on cognition [1]. Specifically, levels of growth hormone releasing hormone (GHRH), growth hormone (GH), and insulin-like growth factor 1 (IGF-1) decline with aging, and diminished levels of these somatotropic hormones are thought to contribute to reduced executive function, impaired memory formation, and Alzheimer's Disease (AD)-like brain pathology.

A large body of research focuses on the important roles of GH and IGF-1 treatments in cognition. One study in particular revealed that growth hormone and IGF-1 treatments increased both peripheral and central neuronal protein synthesis and survival, highlighting the important roles of downstream hepatic GH receptor signaling and IGF-1 on target tissues in neuroprotection [4]. Despite the rationale for GH therapy in individuals with cognitive impairment, responses to GH treatment are not always positive. Moreover, long-term use of GH therapy has been linked with adverse side effects such as hypertension, diabetes, fluid retention, cardiovascular disease, and cancer [2]. Because of these significant concerns, GHRH has become an increasingly explored safer alternative therapy, likely due to its more physiologic and pulsatile release of GH [2].

Hypothalamic regulation in production and release of GHRH stimulates GH secretion via adenylyl cyclase activation and cAMP production in somatotroph cells in the anterior pituitary gland [5]. The pulsatile actions of GHRH treatment may revitalize the aging endocrine axis and promote beneficial cognitive improvements in those with mild cognitive impairment (MCI), AD, and even non-impaired healthy patients [6], [7]. In addition to the direct GH secretagogue effects in neuroprotection, GHRH receptors are expressed in extrapituitary regions, such as the cortex and hippocampus, suggesting other signaling pathways [8]. One proposed pathway is that GHRH administration increases levels of γ-aminobutyric acid (GABA) in the cortex [9]. This is relevant given a relationship between declining GABA levels and worsening executive function has previously been established [10]. GH secretagogues such as GHRH may have additional independent effects on cognitive function that are not recapitulated when GH or IGF-1 levels are augmented directly [11]. This claim is supported by a recent study demonstrating that injection of GHRH agonist into mice reduced Aβ deposition, tau phosphorylation, and inflammatory cytokine expression, without altering systemic GH and IGF-1 levels [12]. Another study found that 20 weeks of 1 mg GHRH administration improved cognitive function in healthy older individuals with MCI and normal cognitive function [6].

In addition to cellular level changes, impaired cerebral perfusion and altered activation patterns of specific brain regions are associated with cognitive impairment in aging population with MCI and AD [13]. Notably, impaired perfusion is relevant to GHRH because GH pulses rely on blood circulation to modulate: 1) delivery of regulatory mediators (e.g. GHRH) to pituitary cells via the afferent microcirculation for GH release, 2) efflux of GH from pituitary cells into efferent capillaries, and 3) eventual delivery of IGF-1 back to specific brain regions where it can stimulate neurogenesis and protein synthesis necessary for normal morphology, memory formation, and cognitive function [6], [14].

The primary objective of this double-blind placebo-controlled study was to investigate the effects of the growth hormone releasing hormone (GHRH) analog tesamorelin (Egrifta®) on cognitive function in 55- to 85-year-old subjects with cognition ranging from MCI to normal. We hypothesized that GHRH may have direct cognitive effects via mechanisms independent of stimulated increases in GH and IGF-1. To study potential direct effects of GHRH while minimizing downstream signaling effects, study treatment was administered at half the standard daily dose (1 mg daily vs. 2 mg daily). Notably, the cognitive effects of GHRH administration over less than a 20-week timeframe have not been directly studied before. The present study employed a 10-week treatment duration to determine whether cognitive effects may arise within a shorter, more clinically practical timeframe. Over the course of 10 weeks of GHRH treatment or placebo, we measured changes in body composition, plasma IGF-1, glucose tolerance, fatigue, sleep, physical performance, cognition, and magnetic resonance neuroimaging including brain morphometry and resting-state functional brain connectivity (rsfMRI). Notably, the effects of GHRH on brain morphometry and connectivity have not been previously determined. Although neurocognitive testing and brain imaging alone often lack the sensitivity to detect modest cognitive changes, their combined effects may increase that sensitivity.

2. Materials and methods

2.1. Enrollment

Both men and women 55–85 years of age (inclusive) were recruited from UTMB's Claude D. Pepper Older Americans Independence Center, Internal Medicine Geriatric clinics, community flyers, and listings on clinical research websites. Interested subjects were screened to determine eligibility at the UTMB Institute for Translational Science Clinical Research Center (ITS-CRC). Subjects were screened for cognition using the mini-mental state examination (MMSE). Subjects with MMSE score > 23 were included in the study, including those with mild signs of cognitive deficit (MMSE score of 23 to 26) and no cognitive impairment (MMSE score of 27 to 30). Subjects were instructed to continue all regular activities of daily living and maintain their usual diet during the 10-week study.

2.2. Study design

A total of 27 men and women were initially enrolled in the current double-blind placebo controlled clinical trial (Fig. 1). Sex was defined based on whether patient was identified as male or female at birth. Subjects were block-randomized based on age and sex to receive 10 weeks of treatment with either GHRH analog or placebo. Baseline data was collected on 24 subjects and final analysis was conducted on 22 subjects which included 12 subjects in the placebo group (5 male, 7 female) and 10 subjects in the GHRH group (5 male, 5 female). At baseline screening, the MMSE was used to classify subjects based on cognitive level. Notably, of the completed subjects, 17 were classified as having normal cognition based on MMSE screening while only 5 met the criterion for MCI. As a result, clinical measures, MRI data, and longitudinal changes were correlated to baseline cognition.

Fig. 1.

Study flow diagram depicts subject enrollment, treatment, completion, and analysis throughout the clinical trial.

2.3. GHRH administration

The GHRH analog, tesamorelin (Egrifta®), was supplied in vials containing 1 mg of tesamorelin (peptide content) and 50 mg of mannitol. Placebo was provided in identical vials containing 50 mg of mannitol. Subjects were given instructions on how to reconstitute and self-inject Egrifta® or placebo by a member of the research team. In this study, single injections containing 1 mg Egrifta® or placebo were self-administered subcutaneously (typically abdominal) daily for 10 weeks. Subjects were given study medication diaries to record the time of injection. Subjects established the time of injection and were advised to keep it consistent every day. If a missed dose was identified within 6 h of the normal dosing time, subjects were instructed to inject that daily dose and note the time in their injection diary. For missed doses identified more than 6 h after the typical dosing time, subjects were instructed to skip the dose and resume the normal dosing schedule on the following day.

2.4. Dual energy X-ray absorptiometry (DEXA)

Body composition was determined at baseline and week 10 using dual-energy x-ray absorptiometry (DEXA; GE Lunar iDXA) to assess changes in total tissue mass, lean body mass, fat mass, and bone mineral content.

2.5. Fatigue and sleep questionnaires

Perceptual fatigue was assessed using the MD Anderson Brief Fatigue Inventory (BFI) [15], chosen for its ease of administration to impaired populations and sensitivity to detect change over time [16]. BFI was measured at baseline, and weeks 2, 4, 6, & 10. Sleep quality was assessed at baseline and weeks 6 & 10 using the Pittsburgh Sleep Quality Index [17].

2.6. Modified 6-min walk

A modified 6-min walk test was conducted to assess physical function. The standard 6-min walk test was modified as previously reported [18], [19], [20], [21], with subjects asked to walk at 25% perceived effort from minute 0–2, at 50% perceived effort from minute 2–4, and at 100% effort (as quickly as they can safely walk without running) from minute 4–6. Distance traveled for each 2-min category was recorded and combined for total distance for analysis. Performance on the 6-min walk test is related to cognitive function [18], [20] and brain volume [20], [21].

2.7. Oral glucose tolerance test (OGTT)

Glucose tolerance was measured at screening and week 10 to identify changes over the treatment period. An IV was placed in an antecubital vein on the subject's forearm. A baseline blood sample was drawn before the subject consumed a bolus of Glucola containing 0.75 g/kg body weight of glucose. Repeat samples were then taken 15, 30, 60, and 120 min after consuming the bolus to measure plasma glucose. Subjects were fasted (no food or drink except water) for approximately 12 h before testing.

2.8. ADAS cognitive behavior subscale analysis

A modified version of the Alzheimer's Disease Assessment Scale–Cognitive Subscale (ADAS-cog) was administered by a single trained and qualified neuropsychologist at baseline and again at the completion of the study (10 weeks) as a detailed assessment of cognition [22], [23]. The ADAS-cog13 has a maximum Total Score of 90 with a higher score indicative of greater impairment. To assess individual ADAS-cog subtests for correlation with MMSE scores, the raw subtest results were analyzed. Individual subtest results allowed for a greater data range without requiring weighting adjustments, ultimately allowing for more precision than scored results.

2.9. Brain composition and morphometry

MRI was used at baseline and after the 10-week treatment period to assess changes in resting state functional brain connectivity (rsfMRI) and regional white-matter changes (fractional anisotropy and mean diffusivity).

2.10. Machine learning

The results of all tests were combined, and the dataset was filtered to remove highly correlated values (cutoff = 0.80). The remaining values were scaled to create a set of 159 predictor variables, which were used to train Machine Learning (ML) models. The R package “caret” was used for machine learning analysis of the morphometric and ADAS results dataset. Five of the available ML models within caret were initially selected for training on the dataset. These models were: Random Forest (rf), Neural Network (nnet), Generalized Linear Model (glm), Partial Least Squares (pls) and Support Vector Machines with Radial Basis Function Kernel (svm). More than one ML technique was chosen to account for the inherent differences of these models and their strengths towards linear and non-linear relationships in the data. Including different ML models best accounted for the mixture of the expected nature of these relationships as their output was used in the reduction to multiple predictor variables.

The performance of ML models was evaluated using the area under the ROC curve (AUC) value, obtained after five-fold cross-validation with 100 repeats. For each run, the top 10 most important predictor variables were recorded from each model if the AUC output of the model was greater than 0.50. To determine the stability of the discovered relationships, a perturbation analysis was performed with repeated seeds (10 times). After 100 runs, the 20 most selected predictor variables were gathered into a separate dataset for the second step.

This dataset was split into training and testing groups to evaluate the performance of the five models. The chosen models included output predictor variables that had the greatest effect in discriminating between placebo and treatment groups. This best performance was observed with the random forest model, which is better suited to explore non-linear relationships as they might expected to exist in human structural connectivity expressed by the diffusion imaging dataset.

2.11. Statistics

Statistical analysis was conducted using GraphPad Prism (GraphPad Software, LLC; Version 9.2.0). Unpaired t-tests were used to compare change from baseline (Δ) between treatment groups. Unequal variance (Welch correction) t-tests were used to assess changes in ADAS-cog subtests due to skewed data distribution (floor effect). Comparison between treatment groups for change in repeated measures over time was done using repeated measures two-way ANOVA (BFI, PSQI, IGF-1). Correlations were assessed for the ADAS-cog subsets against baseline MMSE using Pearson correlation coefficients. Statistical significance was determined using a threshold of p < 0.05.

3. Results

3.1. IGF-1

Plasma IGF-1 values were measured at baseline and again at weeks 2, 4, 6, and 10. Two-way repeated measures ANOVA indicated that there was no effect of treatment (p = 0.499) or combined effects of time and treatment (p = 0.576) between the two groups (Fig. 2).

Fig. 2.

Plasma insulin-like growth factor 1 (IGF-1) levels for subjects receiving 10 weeks of treatment with either growth hormone releasing hormone (GHRH) or placebo.

3.2. Body mass and composition

Dual-energy x-ray absorptiometry (DEXA; GE Lunar iDXA) was used at baseline and week 10 to monitor changes in total tissue mass, lean mass, fat mass, and bone mineral content. Over 10 weeks of treatment, there was an average increase in tissue mass for the GHRH group due to increased lean mass (Table 1). However, the change over time was not significantly different between treatment groups.

Table 1.

Body mass and composition (Avg. ± SD), physical performance (Avg. ± SD), glucose tolerance (Avg. ± SD), for subjects receiving 10 weeks treatment with either placebo or growth hormone releasing hormone (GHRH). Two-way repeated measures ANOVA was used to assess changes over time between groups and within groups (Sidak's correction).

	Baseline	Placebo		Baseline	GHRH		p
		10 Weeks	Δ		10 Weeks	Δ
Body Composition
Tissue Mass (kg)	77.9 ± 18.8	77.3 ± 18.8	−0.6 ± 2.4	75.9 ± 20.1	77.3 ± 19.8	1.4 ± 2.4	0.069
Lean Mass (kg)	46.8 ± 9.0	47.1 ± 8.3	0.4 ± 2.0	46.9 ± 9.8	48.4 ± 10.6	1.5 ± 1.9	0.183
Fat Mass (kg)	31.1 ± 14.3	30.2 ± 15.6	−1.0 ± 2.8	29.0 ± 16.9	28.9 ± 15.7	−0.1 ± 1.7	0.437
Bone Mineral Content (kg)	2.6 ± 0.6	2.6 ± 0.6	0.0 ± 0.0	2.8 ± 0.5	2.8 ± 0.5	0.0 ± 0.0	0.931
Physical Performance
6 -Minute Walk (m)	421.2 ± 74.2	436.4 ± 78.5	15.2 ± 53.1	444.2 ± 86.1	451.1 ± 91.5	6.9 ± 39.1	0.686
Glucose Tolerance
Peak Glucose (mg/dl)	158.1 ± 29.7	166.0 ± 44.7	7.8 ± 37.5	135.1 ± 18.7	139.4 ± 25.3	4.2 ± 21.2	0.791

3.3. Physical performance

A modified 6-min walk test was used to assess treatment-related changes in physical performance. Over the duration of the study, the average 6-min walk distance increased slightly in both groups, though changes were not significantly different between groups (Table 1).

3.4. Glucose tolerance

Oral glucose tolerance was assessed at baseline and week 10. Average peak plasma glucose was minimally higher at 10 weeks for each group; however, no statistically significant differences were observed between groups (Table 1).

3.4.1. Fatigue and Sleep

Perceptual fatigue was assessed using the MD Anderson Brief Fatigue Inventory (BFI) [14] at baseline and again at weeks 2, 4, 6, and 10. Two-way repeated measures ANOVA indicated no effect of treatment (p = 0.102) or combined effects of time and treatment (p = 0.620) between the two groups (Fig. 3). Sleep quality was assessed at baseline and repeated at weeks 6 and 10 using the Pittsburgh Sleep Quality Index [17] Similar analysis indicated there was no effect of treatment (p = 0.186) or combined effects of time and treatment (p = 0.555) between the two groups (Fig. 3).

Fig. 3.

Throughout 10 weeks of treatment with either growth hormone releasing hormone (GHRH) or placebo, subjects were given questionnaires to self-assess fatigue (A) using the Brief Fatigue Inventory (BFI) and sleep quality (B) using the Pittsburgh Sleep Quality Index (PSQI).

3.4.2. ADAS cognitive assessment

Overall, ADAS-cog testing did not detect changes in cognition over the duration of the 10-week study, however, some trends were noted. The average Delayed Word Recall had a non-significant trend towards improvement with 10-weeks of GHRH (p = 0.051). The average orientation scores demonstrated minor improvement in placebo and declined in GHRH, resulting in a statistically significant difference between the two groups (p = 0.050) (Table 2). Of note, one GHRH group subject declined to participate in post-treatment ADAS-Cog testing and was therefore not included in analysis.

Table 2.

Average scores for ADAS-cog subtests before and after 10 weeks of GHRH treatment or placebo. Significant difference from baseline to 10 weeks within groups was assessed with paired t-test. Significant difference in the change from baseline to 10 weeks between groups was assessed with unpaired t-test with Welch correction. One subject declined to participate in post-treatment ADAS-Cog testing and was therefore not included in analysis.

	Placebo			GHRH			P (Δ)
	Pre treatment	Post treatment	Placebo Δ	Pre treatment	Post treatment	GHRH Δ
Total Score	9.2 ± 6.3	7.8 ± 5.6	−1.4	11.0 ± 5.7	9.6 ± 6.2	−1.5	0.985
Word recall	2.6 ± 1	2.7 ± 1.5	0.1	2.9 ± 1.2	2.5 ± 1.6	−0.3	0.347
Commands	0.2 ± 0.4	0 ± 0	−0.2	0.2 ± 0.4	0 ± 0	−0.2	0.768
Constructional praxis	0.5 ± 0.7	0.7 ± 0.7	0.2	0.7 ± 0.5	0.9 ± 0.6	0.2	0.867
Delayed word recall	2.9 ± 2.1	2.3 ± 2.1	−0.7	3.3 ± 2	2.2 ± 2.4	−1.1	0.495
Naming objects	0 ± 0	0 ± 0	0.0	0 ± 0	0.1 ± 0.3	0.1	0.347
Ideation praxis	0 ± 0	0 ± 0	0.0	0.5 ± 0.8	0.1 ± 0.3	−0.4	0.104
Orientation	0.3 ± 0.6	0.1 ± 0.3	−0.2	0.2 ± 0.4	0.4 ± 0.7	0.2	0.050
Word recognition	2 ± 2	1.3 ± 1.4	−0.8	1.7 ± 1.8	2.1 ± 1.6	0.3	0.082
Remembering test instructions	0 ± 0	0 ± 0	0.0	0 ± 0	0 ± 0	0.0
Comprehension of spoken language	0 ± 0	0 ± 0	0.0	0 ± 0	0 ± 0	0.0
Word finding	0.1 ± 0.3	0 ± 0	−0.1	0.1 ± 0.3	0.1 ± 0.3	0.0	0.339
Spoken language ability	0 ± 0	0 ± 0	0.0	0 ± 0	0 ± 0	0.0
Number cancellation	0.9 ± 1	0.8 ± 1	−0.1	1.4 ± 1.1	1.1 ± 0.9	−0.3	0.385
Maze task	0 ± 0	0 ± 0	0.0	0 ± 0	0 ± 0	0.0

	Placebo				GHRH
	Pre treatment	Post treatment	Δ	P (within)	Pre treatment	Post treatment	Δ	P (within)	P (Δ)
Total Score	9.2 ± 6.3	7.8 ± 5.6	−1.4		11.0 ± 5.7	9.6 ± 6.2	−1.5		0.985
Word recall	2.6 ± 1	2.7 ± 1.5	0.1	0.564	2.9 ± 1.2	2.5 ± 1.6	−0.3	0.439	0.347
Commands	0.2 ± 0.4	0 ± 0	−0.2	0.166	0.2 ± 0.4	0 ± 0	−0.2	0.169	0.768
Constructional praxis	0.5 ± 0.7	0.7 ± 0.7	0.2	0.504	0.7 ± 0.5	0.9 ± 0.6	0.2	0.347	0.867
Delayed word recall	2.9 ± 2.1	2.3 ± 2.1	−0.7	0.136	3.3 ± 2	2.2 ± 2.4	−1.1	0.051	0.495
Naming objects	0 ± 0	0 ± 0	0.0		0 ± 0	0.1 ± 0.3	0.1	0.347	0.347
Ideation praxis	0 ± 0	0 ± 0	0.0		0.5 ± 0.8	0.1 ± 0.3	−0.4	0.104	0.104
Orientation	0.3 ± 0.6	0.1 ± 0.3	−0.2	0.167	0.2 ± 0.4	0.4 ± 0.7	0.2	0.169	0.050
Word recognition	2 ± 2	1.3 ± 1.4	−0.8	0.095	1.7 ± 1.8	2.1 ± 1.6	0.3	0.471	0.082
Remembering test instructions	0 ± 0	0 ± 0	0.0		0 ± 0	0 ± 0	0.0
Comprehension of spoken language	0 ± 0	0 ± 0	0.0		0 ± 0	0 ± 0	0.0
Word finding	0.1 ± 0.3	0 ± 0	−0.1	0.339	0.1 ± 0.3	0.1 ± 0.3	0.0		0.339
Spoken language ability	0 ± 0	0 ± 0	0.0		0 ± 0	0 ± 0	0.0
Number cancellation	0.9 ± 1	0.8 ± 1	−0.1	0.586	1.4 ± 1.1	1.1 ± 0.9	−0.3	0.195	0.385
Maze task	0 ± 0	0 ± 0	0.0		0 ± 0	0 ± 0	0.0

3.5. Correlation of screening MMSE and baseline ADAS-cog

MMSE scores from pre-treatment screening data were significantly correlated with baseline performance on ADAS-cog subtests for Word Recall, Commands, Delayed Word Recall, Naming Objects/Fingers, Orientation, Word Recognition, Word Finding, Maze Errors, and Maze Time (Table 3). Correlation was not calculated for Target Errors as no errors were reported at baseline.

Table 3.

Correlation of screening MMSE score to baseline measure of ADAS Cog. subtest scores.

ADAS Cog. Subtest	Pearson Correlation (r)	(p)
Word Recall	−0.585	0.004 *
Commands	−0.583	0.004 *
Construct. Praxis	0.043	0.848
Delayed Word Recall	−0.544	0.009 *
Naming Objects/Fingers	−0.666	0.001 *
Ideation Praxis	0.016	0.943
Orientation	−0.610	0.003 *
Word Recognition	−0.519	0.016 *
Word Finding	−0.575	0.005 *
Target Hits	0.292	0.187
Target Errors†
Maze Errors	−0.397	0.068 *
Maze Time	−0.586	0.004 *

^†No Target Errors were reported.

3.6. Brain connectivity and diffusion tensor imaging

Using resting state fMRI (rsfMRI), whole brain intrinsic connectivity (IC) was assessed for 140 different brain regions. Our initial rsfMRI analysis did not detect significant differences in response to GHRH treatment or placebo for brain connectivity or diffusion tensor imaging (DTI) in specific individual brain regions. To explore potential interactive treatment effects between brain connectivity, DTI, and clinical measures, we implemented a machine learning (ML) approach using an R software script [24]. Specifically, we used five machine learning models (Random Forest, Neural Network, Generalized Linear Model, Partial Least Squares and Support Vector Machines with Radial Basis Function Kernel). The KNN model was less suitable for high-dimensional data and was not used during analysis. Pre and post study measures of the 272 combined clinical and DTI measures were collected from 19 subjects (3 subjects were removed due to missing variables) and were incorporated into the initial statistical model. Highly correlated values (Pearson correlation coefficient > 0.80) were filtered out resulting in a secondary set of 159 predictor variables. These variables were then incorporated into the five machine learning models to identify the top 20 predictive variables that best discriminated between groups (Table 4). These 20 predictor variables were incorporated into the five machine learning models to reduce the dataset to the top 5 predictive variables. Of the top 20 variables, the five most predictive variables for distinguishing between placebo and treatment included: 1) change in total body mass, 2) whole-brain intrinsic connectivity in the right anterior cingulate, 3) fractional anisotropy in the left superior frontal occipital fasciculus, 4) mean diffusivity in the left tapetum, and 5) mean diffusivity in the left cerebral peduncle. Repeat machine learning analysis using the reduced 20 predictor variables showed much better predictive ability and higher treatment effect (Fig. 4). Model performance was evaluated using receiver operating characteristic (ROC) area under the curve (AUC). After validation on a testing set, successful models yielded from 75% to 100% predictive accuracy with a p-value <0.05. The highest AUC value from a successful model was obtained with the Random Forests model (AUC = 0.81). (Fig. 5). Fig. 4, Fig. 5 were generated using the corrplot R package [25].

Table 4.

Highly correlated values (Pearson correlation coefficient > 0.80) were filtered out resulting in a secondary set of 159 predictor variables. These variables were then incorporated into the five machine learning models to identify the top 20 predictive variables that best discriminated between groups.

Clinical	DTI	Connectivity
Word Recall (# NOT Recalled)	Fractional Anisotropy ICP-R	LMedFronG Connectivity
Commands (# NOT Correct)	Fractional Anisotropy PTR-R	LMidN Connectivity
Ideational Praxis (# NOT correct)	Fractional Anisotropy CGC-L	LDentate Connectivity
Orientation (# NOT correct)	Fractional Anisotropy SFO-L	RACing Connectivity
Word Recognition (# NOT correct)	Mean Diffusivity CP-L	RCaudateH Connectivity
Total Mass (kg) Total	Mean Diffusivity SFO-L	RNod Connectivity
	Mean Diffusivity TAP-L	RTuber Connectivity

Fig. 4.

Panel A shows the ability to predict treatment based on all collected measures (159 variables). Panel B shows the ability to predict treatment using the reduced 20 variables. Note the reduced set of variables is more correlated within treatment groups.

Fig. 5.

ROC curve of random forests model performance on testing dataset (AUC = Area under curve, Sensitivity = True positive rate, 1-Specificity = False positive rate).

4. Discussion

We studied the effect of 10 weeks of daily GHRH supplementation in adults aged 55 to 85, who on MMSE screening were categorized over a range of cognition from normal cognition (MMSE of 27–30; n = 17) to MCI (MMSE of 23–26; n = 5). Over the 10-week course of GHRH supplementation we monitored changes expected as a result of aging including body composition, plasma IGF-1, glucose tolerance, fatigue, sleep, physical performance, and cognition (using the modified ADAS-cog). Moreover, magnetic resonance neuroimaging including brain morphometry and resting-state functional brain connectivity (rsfMRI) were assessed as the effects of GHRH on brain morphometry and connectivity have not been previously determined.

Age-related decline in growth hormone releasing hormone (GHRH), growth hormone (GH), and IGF-1 are associated with reduced executive function, impaired memory formation, and AD-like brain pathology as well as decreased lean mass and increased fat mass [3], which are also linked with neurologic decline. Interestingly, growth hormone secretion is impaired by excess adipose tissue [26], [27], an effect which is reduced by GHRH therapy [28]. We found that 10 weeks of low dose GHRH treatment did not alter IGF-1 levels (Fig. 1), body mass (Fig. 2), or peak glucose levels. It is possible that these results may support our experimental use of the lower dose of GHRH to minimize the GH secretion with GHRH treatment so that effects are more likely from direct effects of GHRH. On the other hand, the insignificant findings across these primary endpoints could reflect insufficient dosing entirely. Regardless, a lower-dose therapeutic agent seems favorable in terms of safety profile. GHRH treatment even at doses twice those used in this study are primarily associated with tissue-specific effects and systemic side effects related to the subsequent secondary increase in GH [29]. At higher doses of tesamorelin, there is evidence of glucose intolerance, fluid retention, and IGF-1 elevation [30].

Results from prior studies have revealed that both MCI and healthy patients may experience cognitive improvements after GHRH supplementation [6], [7]. Given that performance on the standard 6-min walk test has been linked to cognitive function and brain volume [18], [19], [20], [21]] we assessed physical function at baseline and throughout the duration of the 10 weeks of GHRH. We found that the average 6-min walk distance increased slightly in both groups although changes were not significantly different between groups. However, the low dose used (1 mg), compared to the standard dose (2 mg), may have impacted our ability to show a stronger GHRH effect and support the design of our study to focus on direct effects of GHRH on the brain.

The ADAS-cog is commonly used to evaluate cognitive changes in humans, but its utilization to assess the effect of GHRH treatment on cognition is unique. Although originally designed to detect AD progression, modified versions of ADAS-cog are being applied more broadly to detect pre-dementia cognitive states [22]. Importantly, we implemented the ADAS-cog 13 scoring system, which includes all tasks of ADAS-cog 11, with the addition of Delayed Word Recall and maze tasks. This modified scale allowed us to collect more sensitive data and better discriminate cognitive changes in individuals without profound impairment [22]. Specifically, the Delayed Word Recall assessment appears to be the most sensitive subscale in detecting milder levels of cognitive impairment, and the least sensitive for more severe cognitive disease like AD [22], [31]. Consistent with previous studies, we found that the Delayed Word Recall subscale was the only one associated with a trend towards improved score after 10 weeks of daily GHRH treatment. Our findings support the role of Delayed Word Recall acting as a beneficial assessment for detecting preclinical cognitive changes. GHRH and its pulsatile effects may revitalize the HPA axis allowing for beneficial effects on cognition [6], [7]. Experiments studying the neurochemical changes associated with prolonged GHRH therapy have reported associated increases in GABAergic pathways, such as those involving IGF-1, important for modulating cognitive inputs [9], [11]. Along similar lines, previous studies have also demonstrated an important link between the HPA axis and cognition, revealing improved cognition scores in traumatic brain injury (TBI) after GH treatment [32], [33].

As for the other subcategories of ADAS-cog, GHRH treatment did not result in statistically significant improvements. However, these measures should be interpreted with caution due to the small error rate (Placebo = 3 total errors combined for all subjects at baseline and 1 at 10 weeks; GHRH = 2 total errors combined for all subjects at baseline and 4 at 10 weeks). It is possible that the shorter study timeframe of only 10 weeks, compared to 20–36 weeks in similar studies, may have perturbed the therapy's full potential [6], [7], [33], [34]. Alternatively, it may be that as opposed to the Delayed Word Recall tasks, many of the remaining ADAS-cog tasks, such as those evaluating language and praxis, are more clinically useful for subjects with more severe cognitive impairment than MCI [22], [31].

In order to discover tools aside from ADAS-cog that could assist in diagnosing smaller cognitive declines, we evaluated how neuroimaging and neuronal network analysis may be useful biomarkers in the detection of cognitive changes [34], [35]. In this study, GHRH treatment did not display a statistically significant effect on measures of brain connectivity or morphometry, as measured by rfMRI and DTI. However, in efforts to explore dynamic interactions between the brain network with DTI and clinical measures, ML models were developed and trained to determine the top variables that best discriminate between placebo and treatment. The ML approach suggested the top variables associated with differences between placebo and treatment groups included total body mass, whole-brain intrinsic connectivity in the right anterior cingulate, fractional anisotropy in the left superior frontal occipital fasciculus, mean diffusivity in the left tapetum, and mean diffusivity in the left cerebral peduncle. Interestingly, most of these brain regions have displayed cognitive associations with certain diseases in the literature. For example, it has been proposed that the structural integrity of the left superior frontal occipital fasciculus is correlated with cognitive decline in 22q11.2 deletion syndrome [36]. Another study revealed that the right anterior cingulate receives more cerebral blood flow in children with attention-deficit/hyperactivity disorder (ADHD) during certain cognitive tasks [37]. Notably, the anterior cingulate region also has clinical significance in post-acute sequelae SARS-CoV-2 infection (PASC) patients. In a study of seven patients with SARS-CoV-2 encephalopathy, all displayed hypometabolism to the anterior cingulate which may contribute to the “brain fog” reported after infection [38]. Furthermore, the longitudinal data collected in that study showed that while subjects improved symptomatically after 6 months, they continued to have evidence of prefrontal hypometabolism with cognitive impairments such as attention/executive deficits [38]. The cingulate cortex is biologically plausible target for GHRH neuromodulation. Administration of GHRH is associated with increased cortical GABA levels, including in the cingulate cortex [9]. Age related cognitive decline has also been linked to reduced GABA levels in the anterior cingulate [10].

Overall, the tools utilized in this study are hypothesis generating for proposed structural changes in MCI and healthy patients that could serve as potential biomarkers of cognitive decline. These findings may demonstrate the utility of augmenting clinical tests of cognitive function with imaging data and machine learning models to understand the full range of effects of GHRH supplementation on the human body. Moreover, these data shed light on the potential therapeutic effect GHRH could have for MCI and other pathologies which negatively impact cognition.

Study Limitations: The small sample size, age heterogeneity, and predominantly cognitively normal participants all substantially affect the generalizability of results in this pilot study. Although sex was evenly distributed in this double-blind randomized placebo-controlled clinical trial, the limited number of subjects for this pilot study precluded stratification of outcomes based on sex. Although these flaws limit the conclusions that may be drawn from the current study, the results are hypothesis-generating and can inform future studies powered to detect meaningful cognitive effects in more homogenous patient cohorts.

In the current study, we did not see profound effects on the DEXA-derived body composition, metabolic, behavioral, or cognitive measures assessed here with ten weeks of GHRH treatment. We acknowledge that perhaps the lower dose of tesamorelin used could have been insufficient. The 1 mg daily injection of tesamorelin (GHRH) used in this study is half the dose typically prescribed for treating lipodystrophy, and plasma IGF-1 levels did not increase significantly in the treatment group. This dosing was chosen to test for potential direct GHRH effects while trying to minimize direct stimulation of GH secretion and IGF-1. Furthermore, the ten-week study duration may have been too brief to fully realize the therapeutic effects low dose tesmorelin.

Notably, while no statistically significant differences were detected in body composition derived by DEXA, both treatment and placebo groups trended towards increased lean mass and decreased fat mass. This may be evidence of participant mediated bias, in which enrollment into the study inadvertently lead to healthier behaviors, regardless of assigned treatment arm. Participants were not instructed to keep physical activity and food diary logs and thus we did not appropriately account for this bias in our study design.

Although ADAS-cog was developed to assess significant cognitive impairment in AD, modifications to this test have been made to make it more sensitive to assess MCI [22], [31], [39], [40], [41], [42]. However, when testing our study population of individuals with normal cognition or MCI, we found few or no errors in groups for entire subscales. This floor effect suggests that the scale may not be well suited and sensitive to monitoring cognition in a less impaired population [22]. In this study, it is also important to note that the baseline ADAS-cog scores in the placebo group were lower (indicative of better cognitive performance) compared to the treatment group. This imbalance, inherent to the small sample size, may have arbitrarily inflated the improvements in the treatment group. Similarly, the significant differences in certain subgroups (ex: improvement in orientation noted in the placebo group) is likely attributable to random variation in baseline values within an already small sample size.

Lastly, we would like to acknowledge that the machine learning approach analyzes several study parameters within a small a sample size, which may raise concerns regarding robustness and generalizability. Again, this reflects the more exploratory nature of this study design and because of this the results should be interpreted cautiously.

5. Conclusion

Baseline ADAS-cog scores were significantly correlated with MMSE scores, which shows that when combined, the two cognitive tests may rectify diagnostic sensitivity issues observed when using ADAS-cog as the sole diagnostic tool for MCI. GHRH treatment was associated with a trend towards improved scores in performance and cognition, mirroring significant changes demonstrated previously using GH treatment in patients suffering long-term cognitive and fatigue effects following TBI. The Delayed Word Recall subcategory of ADAS-cog continues to be the most sensitive indicator for cognitive change in its early stages. However, 10 weeks of GHRH treatment did not have a significant effect on longitudinal changes in cognition, brain connectivity, or brain morphometry. The shorter treatment duration may have limited the effect size. As a result, we utilized machine learning techniques to determine the most significantly correlated factors. Using machine learning techniques, we demonstrated that GHRH treatment is associated with altered brain connectivity and morphometry in brain regions associated with cognition. Changes observed in this study occurred with low dose of GHRH, and without significant increase in IGF-1 levels. Results from this study suggest that combining cognitive testing and brain imaging may increase sensitivity for monitoring longitudinal cognitive change.

CRediT authorship contribution statement

Courtney E. Stewart: Writing – review & editing, Writing – original draft, Formal analysis. Kaci P. French: Writing – review & editing, Writing – original draft, Formal analysis. Traver J. Wright: Writing – review & editing, Visualization, Formal analysis. Kayla Wilhoit: Writing – review & editing, Visualization, Software, Formal analysis. Kathleen M. Randolph: Writing – review & editing, Project administration, Investigation. Christopher P. Danesi: Writing – review & editing, Investigation. Charles R. Gilkison: Writing – review & editing, Investigation. Christof Karmonik: Writing – review & editing, Software, Formal analysis, Data curation. Lei Lu: Writing – review & editing, Resources, Methodology, Investigation, Formal analysis. Edgar L. Dillon: Writing – review & editing, Formal analysis. William J. Durham: Writing – review & editing, Resources, Methodology, Investigation, Formal analysis, Conceptualization. Randall J. Urban: Writing – review & editing, Supervision, Funding acquisition, Conceptualization. Melinda Sheffield-Moore: Writing – review & editing, Supervision, Funding acquisition, Conceptualization. Brent E. Masel: Writing – review & editing, Resources, Conceptualization.

Ethics statement

We have read and have abided by the statement of ethical standards for manuscripts submitted to eNeurologicalSci. This study was approved by the University of Texas Medical Branch Institutional Review Board (IRB) on 08/24/2015 (reference number: 15–0086). All procedures were performed lawfully and in compliance with the IRB. Informed consent was obtained for all human subjects prior to participation in this study. We complied with the privacy rights of human subjects in our study.

Funding sources

This study was supported by the Institute for Translational Sciences at the University of Texas Medical Branch, supported in part by a Clinical and Translational Science Award [UL1TR001439–06] from the National Center for Advancing Translational Sciences; National Institutes of Health and with the support of the Moody Endowment [#2014–01]; and the generous donation of tesamorelin (Egrifta®) by Theratechnologies, Inc., Quebec Canada.

Declaration of competing interest

None.