MRI as a biomarker of disease progression in a therapeutic trial of milameline for AD

Study design.

The study was a 52-week randomized double-blind placebo-controlled parallel group multicenter trial of milameline, a muscarinic receptor agonist. All patients received titrated doses up to the maximum tolerated daily dose, with a ceiling of 4 mg/d. Clinical followup was at months 1, 3, 6, and 12. The primary outcome measure was the AD Assessment Scale–cognitive subscale (ADAS-Cog). Several secondary outcome measures were employed including the Mini-Mental State Examination (MMSE), Global Deterioration Scale (GDS), and MRI measurements of the rates of atrophy of the hippocampi and the rates of enlargement of the temporal horns.10

Patients.

All patients had probable AD of mild to moderate severity and were >50 years old. MMSE scores were 10 to 27, inclusive. Patients were excluded for evidence of substantial cerebral vascular disease, non-AD dementia, major psychiatric disorders, major or unstable medical conditions, seizure disorder, PD, substance abuse, major head trauma, or tumor.

Planned enrollment was a total of 450 subjects, all from US sites with an average of 10 subjects (range 10 to 15 maximum) from each of 45 sites. The planned study duration was from June 1996 to December 1998.

MRI.

In order to avoid difficulties associated with MRI manufacturer’s proprietary data acquisition and storage protocols, only those sites with a General Electric (Milwaukee, WI) MR Imager were asked to enroll patients in the MRI portion of the trial. The precise hardware and software configuration of scanners at different sites varied. All scanners were 1.5 T, with the exception of a single site at 0.5 T.

The individual imaging sequences in the examination protocol were 1) a sagittal T1-weighted scan with contiguous 5-mm slices; 2) a three-dimensional volumetric spoiled gradient recalled echo (SPGR) scan obtained in the coronal plane with minimum full echo time, minimum repetition time, 124 partitions, and 1.6-mm partition thickness; and 3) a T2-weighted scan acquired for pathology screening purposes.

Each site went through an initial qualifying phase prior to scanning patients for the trial. The qualifying phase had two components. First, each site was required to perform the pulse sequences specified for the imaging protocol, film the examination, and send the films to the central analysis site at the Mayo Clinic. The pulse sequence parameters were checked at the central analysis site to ensure that each participating site could execute the specified imaging sequences correctly. Second, the scanner on which study patients would be imaged at each site underwent a quality control evaluation. In order to begin studying patients, signal-to-noise ratio (SNR) and geometric distortion analyses must have been completed and evaluated at the central site and must have fallen within predetermined specifications.

SNR and geometric distortion measurements were made on the specified scanner at each site both during the qualification phase and throughout the trial on a monthly basis. Each site performed and submitted to the central image analysis site an axial spin echo and sagittal, coronal, and axial gradient echo acquisitions, which were specifically designed to evaluate SNR and geometric distortion. The quality control imaging protocol was approximately 15 minutes in duration and was done with a standardized quality control phantom supplied by the manufacturer. The phantom contained spatially uniform regions where SNR measurements were obtained and also contained fiducial markers that were used to assess geometric distortion.11 The data from the quality control MR protocol was transmitted to Mayo, where an ongoing record of SNR, image artifacts, radio frequency power, and geometric distortion along all three axes was kept for each scanner throughout the duration of the trial.

All image data (both patients’ imaging data and phantom quality control imaging data) were sent by either tape or disc to Mayo and archived electronically. For each patient the total intracranial volume (TIV) was measured from the sagittal T1 sequence. The volumes of the right and left hippocampi and temporal horns were measured according to previously described criteria6,12,13⇓⇓ on the three-dimensional SPGR sequence. All the measurements were done at Mayo by a single individual over 2 years. Intrarater coefficient of variation for serial hippocampal measurements has been documented at 0.28%.6 The individual performing the MRI analyses was blinded to all clinical information including the center and the scan date (i.e., whether each scan was the first or second in the pair). The order of first and second scans was randomized.

MRI data analysis.

Age- and sex-specific elderly norms for total intracranial volume–adjusted hippocampal volume.

We next turned to hippocampal volumes in normal elderly people. The sagittal and three-dimensional SPGR pulse sequence had been performed in 181 normal elderly individuals (61 men and 120 women, mean age 80.4 years, range 62 to 100 years) as part of the ongoing clinical research protocols in the Mayo Clinic AD Research Center and AD Patient Registry. For each of these normal elderly subjects, the TIV-adjusted hippocampal volume was computed based on the relationship between TIV and hippocampal volume that had been derived for the normal young cohort as described here. Age- and sex-specific normative percentiles of this TIV-adjusted hippocampal volume value were then computed for the normal elderly cohort. The method for computing age- and sex-specific normative percentiles for hippocampal volume was identical to the method we have previously described,13 except that here normative percentiles were calculated in the elderly subjects for the TIV-adjusted hippocampal volume.

Age and total intracranial volume–adjusted hippocampal volume in milameline study patients.

The baseline hippocampal volumes in the patients with AD in the milameline study were expressed as W scores, as we have done previously.13 The W score corresponds to the percentile value in a standard normal distribution. A W score of zero corresponds to the 50th percentile of normal elderly individuals. A W score of 1.645 corresponds to the 95th percentile of normal elderly individuals, and a W score of −1.645 corresponds to the 5th percentile.

Serial MRI measures in milameline study patients.

In each study patient, volume measurements of the hippocampi were obtained at two different time points. From these, the raw change in volume (in mm³), raw annualized change in mm³/year, and annualized percent change were calculated. No relationship was found between TIV and the rates of change in hippocampal volume, and therefore no TIV adjustment was needed.6

We did not have measurements of temporal horn volume available in young subjects or elderly, normal subjects. For this reason, adjustment of temporal horn volume for TIV and the W score method could not be developed for the temporal horn measurements. We analyzed the annualized raw volumetric change and the annualized percent change in temporal horn volume in a manner identical to that described for the hippocampus.

Statistical analysis.

The variables of interest in the analyses were age, sex, education, ADAS-Cog score, MMSE, GDS, and MRI volume variables. The MRI volume variables include raw data in mm³ for the right, left, and total hippocampus at baseline, W score of total hippocampal volume at baseline, change in total hippocampal volume, raw baseline temporal horn volumes, and change in total temporal horn volume. Rank sum tests were used for skewed data, and two sample t-tests were used for normally distributed data. Spearman correlation was used to test for associations between MRI volume measurements and behavioral/cognitive measures, education, sex, and age. Univariate and multivariate modeling were used to test for associations between baseline MRI volumes and both demographic and behavioral/cognitive measures. Univariate and multivariate modeling were used to test for associations between change in MRI volumes and baseline MRI volume, demographic, and behavioral/cognitive measures. Rank transformations were used for all data in the multivariate modeling. One-way analysis of variance was used to test for differences in baseline hippocampal W score between the different centers. The degree to which serial measures, imaging and behavioral/cognitive, declined consistently over time was compared using the sign test.

Sample size calculations for clinical trials.

Because of the potential interest in using structural MRI as a biomarker of disease progression in clinical AD trials, we performed sample size calculations to detect treatment effect based on the annual change data in MMSE, ADAS-Cog, hippocampus, and temporal horn (the GDS was thought not to be a competitive metric). Sample size calculations were based on the assumption of a 50% treatment effect over 12 months—i.e., a 50% reduction in the change over 1 year in a treated vs a placebo group. Tests were one sided with power set at 90%.

Baseline demographic, behavioral/cognitive, and MRI data.

Of 453 subjects enrolled in the drug trial, 362 participated in the MRI portion. Analysis of baseline data are restricted to these subjects (table 1). Among these 362 subjects, 188 had been randomly assigned to treatment and 174 to placebo. No differences in sex distribution (roughly 60% female), education, behavioral/cognitive measures, or baseline MRI volume measures were present between the treated and placebo groups.

Women had less education (p = 0.001) and worse baseline cognitive performance on the ADAS-Cog (p = 0.02) and MMSE (p = 0.01). The raw right, left, and total hippocampal (p < 0.001) and temporal horn (p < 0.001) volumes were larger in men than women. There was a trend toward greater hippocampal atrophy (lower W score and normal percentile ranking) in women than men, but this did not reach significance.

The right hippocampus was approximately 100 to 200 mm³ (roughly 5%) larger than the left in both treated and placebo groups. However, no right-left differences were present in any of the correlations between hippocampal or temporal horn volume and baseline demographic or behavioral/cognitive measures; nor were right-left differences present in correlations between change in MRI volumes and change in behavioral/cognitive measures. Therefore, all subsequent analyses of MRI volume measures are reported for the total (sum of right plus left) volume rather than right or left volumes separately.

Correlation among baseline hippocampal volume, demographic, and behavioral/cognitive measures.

The placebo and treated groups were combined for baseline correlation analyses because the baseline MRI and behavioral/cognitive measures were completed prior to treatment initiation. Univariate correlations between baseline hippocampal W score, age, sex, education, ADAS-Cog, MMSE, and GDS were performed. No significant correlations were present with the exception of that between hippocampal W score and age (r = 0.15, p = 0.004). A series of multivariate models were constructed with performance on the behavioral/cognitive measures listed previously as the dependent variable, and baseline hippocampal W score, age, sex, and education as independent variables. None of the partial correlations between baseline hippocampal W score and baseline behavioral/cognitive performance were significant.

Normative temporal horn measurements were not available and thus a system for adjusting raw baseline temporal horn volumes of study patients receiving milameline for age and sex in subjects was not possible. We therefore did not assess correlations among baseline temporal horn volume, demographic, and behavioral/cognitive measures.

Changes from baseline.

All analyses of change (change in MRI volumes and change in behavioral/cognitive scores) are based on the 192 subjects who had both a baseline and a second MRI scan (table 2). No difference in annual change was present between the treated and placebo groups in any of the behavioral/cognitive measures, nor for either of the MRI volume variables. For each of these variables, however, the median annualized change was different from 0 (p < 0.001, for all, signed rank test). In addition, all variables in table 2 changed in the expected direction over time; behavioral/cognitive performance worsened, hippocampal volume decreased, and temporal volume increased.

To assess how consistently imaging and behavioral/cognitive measures declined over time, we computed the proportion of individuals in whom the measures in table 2 declined. Decline was defined as a decrease in hippocampal volume, an increase in temporal horn volume, a decrease in MMSE score, and an increase in the GDS and ADAS-Cog scores. The volume of the hippocampus decreased in 99% of subjects, whereas only 60.4% of subjects’ ADAS-Cog and 66.2% of subjects’ MMSE scores declined. In pairwise comparisons, the proportion of decliners was greater for the hippocampus than any of the behavioral/cognitive measures or the temporal horn (p < 0.001). The proportion of decliners was also greater for the temporal horn (85.4%) than for any of the behavioral/cognitive measures (p < 0.001).

Sample size calculations for clinical trials.

There were two outliers in the MMSE data. After deleting these values for purposes of computing sample size requirements, the MMSE and temporal horn data were not highly skewed and transformations of these data were not needed. For these variables, we computed the effect size to be 50% of the observed mean annual rate. The hippocampus and ADAS-Cog were highly skewed and transformations of these data were needed. For these we computed the effect size after transformation as follows: transformed(median) − transformed(0.5 × median).

For ADAS-Cog, we first transformed to Y′ = Y − min+1 (min = −59.9), then used Y′*10⁻², where min refers to the minimum observed value.

For hippocampus, Y′ is defined as above (min = −15.2), and we used Y′*10².

Notice that we are using a 50% reduction in mean rate of change where transformations are not required and a 50% reduction in the median rate of change where transformations are required. After deletion of the two outliers from the MMSE data, the mean annual percent change was 10.7% (19.9). The sample size required to detect a 50% reduction in this rate of change in a 1-year placebo-controlled trial with power of 90% (one-sided t-test at the 0.05 level) is 241 per arm. For ADAS-Cog, the data were highly skewed with a median of 16.4% (the associated SD of the transformed data was 65.4). In order to detect 50% reduction, n = 320 per arm are needed. For the hippocampus, the data were again highly skewed with a median of −4.9% (the associated SD of the transformed data was 2.1). In order to detect a 50% reduction, n = 21 per arm are needed. The data for the temporal horn were not skewed, with mean 16.1% (SD 14.1) and n = 54 per arm needed.

Factors that influence change in MRI volume.

In order to assess whether demographic variables or baseline MRI volume influenced the annualized change in volume, a multivariate model was constructed with the raw annualized change in hippocampal volume (mm³) as the dependent variable and with age, sex, education, and baseline hippocampal W score as independent variables. The slope (beta), standard error, partial Spearman correlation, and associated p value for each of the four independent variables in this model appear in table 3. Baseline hippocampal W scores were inversely associated with rates of hippocampal atrophy (i.e., smaller hippocampi at baseline were associated with greater volume loss over time). Age, sex, and education at baseline were not associated with the annualized rate of hippocampal atrophy.

Similar modeling was performed using the raw annualized change in temporal horn volume (mm³) as the dependent variable and age, sex, education, and baseline temporal horn volume as predictor variables. Because W scores were not available for temporal horn measurements, the raw temporal horn volume in mm³ at baseline was used as the independent baseline volume variable. Younger age at baseline was associated with a greater annualized change in temporal horn volume (i.e., a greater rate of atrophy). A larger temporal horn volume at baseline was also associated with a larger annualized change in temporal horn volume.

Similar modeling was performed using the raw annualized change in volume (hippocampal and temporal horn) as the dependent variable and age, sex, education, and baseline behavioral/cognitive variables as predictor variables. None of the baseline behavioral/cognitive variables were associated with the annualized change in hippocampal volume. ADAS-Cog was the only baseline cognitive variable associated with the annualized change in temporal horn volume (r = 0.38, p < 0.001).

Correlation between MRI volumes and change in behavioral/cognitive measures.

MRI quality control measures.

All 38 participating sites eventually met qualifying criteria from the standpoint of MRI quality control. However, over the course of the study, results of quality control analysis warranted site contact 19 times. Most of the quality control alerts occurred during the 1-month qualifying phase, which preceded enrollment of patients. The nature of the quality control alerts were gradient coil error (i.e., geometric distortion), nine alerts at seven sites; low SNR, five alerts at four sites; noise lines, three alerts at three sites; other, two alerts at two sites.

Another measure of quality control is the consistency of measured volumes across sites. The mean hippocampal W score across all sites was −1.78 (SD 0.97). Hippocampal W score did differ among sites (one-way analysis of variance, p = 0.04). We then used the Student-Newman-Keuls multiple range test to do pairwise comparisons, and no pairwise difference between sites was found. We estimated the components of variance: variance arising from difference among sites (V_A) and variance arising from difference among patients within sites (V_W). These were V_A − 0.04 and V_W − 0.90, and the percent of total variance (V_T = V_A + V_W) due to variability among sites was 4.71% (V_A/V_T*100%). Stepping down multivariate regression was used with hippocampal W score as the dependent variable, and age at MRI scan, female sex, treatment, site, baseline ADAS-Cog score, baseline MMSE score, and baseline GDS score as the independent variables. Only one site was significantly different from the others in this analysis.