Medial Prefrontal Cortex Glutamate Is Reduced in Schizophrenia and Moderated by Measurement Quality: A Meta-analysis of Proton Magnetic Resonance Spectroscopy Studies

BACKGROUND: Magnetic resonance spectroscopy studies measuring brain glutamate separately from glutamine are helping elucidate schizophrenia pathophysiology. An expanded literature and improved methodologies motivate an updated meta-analysis examining effects of measurement quality and other moderating factors in characterizing abnormal glutamate levels in schizophrenia. METHODS: Searching previous meta-analyses and the MEDLINE database identified 83 proton magnetic resonance spectroscopy datasets published through March 25, 2020. Three quality metrics were extracted—Cramér–Rao lower bound (CRLB), line width, and coefficient of variation. Pooled effect sizes (Hedges’ g) were calculated with random-effects, inverse variance-weighted models. Moderator analyses were conducted using quality metrics, field strength, echo time, medication, age, and stage of illness. RESULTS: Across 36 datasets (2086 participants), medial prefrontal cortex glutamate was significantly reduced in patients (g = −0.19, confidence interval [CI] = −0.07 to −0.32). CRLB and coefficient of variation quality subgroups significantly moderated this effect. Glutamate was significantly more reduced in studies with lower CRLB or coefficient of variation (g = −0.44, CI = −0.29 to −0.60, and g = −0.43, CI = −0.29 to −0.57, respectively). Studies using echo time ≤20 ms also showed significantly greater reduction in glutamate (g = −0.41, CI = −0.26 to −0.55). Across 11 hippocampal datasets, group differences and moderator effects were nonsignificant. Group effects in thalamus and dorsolateral prefrontal cortex were also nonsignificant. CONCLUSIONS: High-quality measurements reveal consistently reduced medial prefrontal cortex glutamate in schizophrenia. Stricter CRLB criteria and reduced nuisance variance may increase the sensitivity of future studies examining additional regions and the pathophysiological significance of abnormal glutamate levels in schizophrenia.


Contents of the Supplement
Supplemental Methods: Empirical method for identifying quality thresholds Supplemental Results: Exploratory analysis of signal-to-noise as a quality metric Table S1: Excluded studies and reasons for exclusion Table S2: Included studies of regions for which < 10 datasets are available     We examined three metrics sensitive to the quality of the glutamate measurements for each study: mean + 2 SD for Cramer-Rao lower bound (CRLB), mean + 2 SD for singlet line width (FWHM), and mean COV for glutamate. For each metric, we averaged the values for the patient and control groups. We hypothesized that glutamate measurement quality would have a moderating effect on the meta-analytic results across studies comparing schizophrenia patients to healthy volunteers. Formally, we hypothesized there was a quality threshold Q, for which the meta-analytic result would be significantly stronger in studies surpassing Q than for those falling short of Q. To identify the quality threshold Q in an unbiased manner, we first ranked the studies for each metric. We then calculated the inverse variance-weighted pooled effect sizes from a moving sample of studies (k = 7) running from the lowest to the highest quality studies for each quality metric (analogous to a moving average). A best-fitting, 4-parameter, logistic function was fit to this series of pooled effect sizes using the computational resource at https://mycurvefit.com/ using the following equation: Where Y = the pooled effect size (k=7) and X = the rank of the set of seven adjacent studies for the quality metric being examined. The best fitting four parameters (a, b, c, and d) for each of the quality metrics is shown below. Parameter "a" is the asymptote of the pooled effect size for the lowest quality datasets, and parameter "d" is the asymptote of the pooled effect size for the highest quality datasets for each metric. These best fitting parameters were used to generate a logistic transform of the ranks of each quality metric. The empirical quality threshold Q was identified as the inflection point in the logistic transform curve. The inflection point (Q) is the midpoint between parameters "a" and "d" (thus Q = (a + d)/2). This point Q was used to stratify studies into low and high quality subgroups for each metric. All studies included in a set of 7 ranked studies for which the moving pooled effect size (k = 7) was more negative than Q were stratified into the high quality subgroup for that quality metric. All other studies were stratified into the lower quality subgroup.
Supplemental Results -Exploratory analysis of signal-to-noise as a quality metric.
Spectral signal-to-noise (SNR) was not included as an a priori quality metric for testing the hypothesis that the meta-analytic result would be significantly stronger in studies surpassing an empirically identified threshold for measurement quality. We chose this approach in order to limit multiple comparisons for testing this hypothesis. It was our opinion, a priori, that SNR would be the least discriminating of the four quality metrics commonly reported (COV, CRLB, FWHM, and SNR). We reasoned that quality metrics based specifically on the glutamate measurement, such as CRLB and COV for glutamate, might have an advantage over those based on the whole spectrum, such as FWHM and SNR. With regard to the latter two, in our own lab we have consistently found low FWHM to be a better predictor of valid glutamate measurements than high SNR.
In response to a question about this issue during peer review, we searched for and extracted SNR mean and SD values from the 36 mPFC studies included in our metaanalysis. Only 23 studies reported these data, and only 20 used an equivalent method for calculating SNR (the LCModel default method). Applying the same procedure as for the other quality metrics, we identified 14 high quality datasets for SNR (mean minus 2 SD ≥ 13). Six studies were identified as having lower quality SNR values. The 16 studies not reporting SNR were included in the lower-quality subgroup. Moderator analyses showed that effect sizes were not significantly different between lower-and high-quality subgroups for SNR (omnibus model Q = 2.2, df =1, p = .13; heterogeneity: I 2 = 45, p = .002). When studies not reporting SNR were excluded altogether from the moderator analysis, there was trend for mPFC glutamate to be more reduced in the high-quality versus the lower-quality SNR subgroup, but it was not significant with our corrected alpha (omnibus model Q = 3.9, df =1, p = .048; heterogeneity: I 2 = 42, p = .02). Detailed statistics for each SNR subgroup are shown below. In agreement with our expectation, an empirical quality threshold based on SNR was less successful than the other quality metrics at identifying studies sensitive to reduced mPFC glutamate in schizophrenia.

Moderator effects
Meta-regression analysis showed no effect of either field strength or log TE. The distributions of these regressors, however, were very limited. There were no studies above 3T, and 8 of the 11 studies used TE between 30 and 35 ms. Three datasets were categorized as ≥ 80% unmedicated and 6 datasets as all medicated (2 datasets          functional magnetic resonance imaging-proton magnetic resonance spectroscopy study. Aust N Z J Psychiatry. 2020;54 (5)