The diagnostic potential of resting state functional MRI: Statistical concerns

Abstract

Blood oxygen level-dependent functional magnetic resonance imaging (fMRI) is a widely used, non-invasive method to assess brain hemodynamics. Resting-state fMRI (rsfMRI) estimates functional connectivity (FC) by measuring correlations between the time courses of different brain regions. However, the reliability of rsfMRI FC is fundamentally compromised by statistical artifacts arising from signal cyclicity, autocorrelation, and preprocessing-induced distortions. We discuss how standard rsfMRI preprocessing -particularly the widely used band-pass filters such as 0.009-0.08 Hz and 0.01-0.10 Hz- introduce biases that increase correlation estimates between independent time series. Additionally, filtering without appropriate downsampling further distorts correlation coefficients, inflating statistical significance and increasing the risk of false positives. Under these conditions, commonly used multiple comparison corrections fail to fully control Type I errors, with up to 50-60 % of detected correlations in white noise signals remaining significant after correction depending on the sampling rate, filter and duration. To mitigate these biases, we recommend adjusting sampling rates to align with the analyzed frequency band and employing surrogate data methods that better account for the statistical properties of rsfMRI signals and reduce autocorrelation-driven false positives. Additionally, we show that structured brain states-such as epilepsy and anesthesia-induced burst suppression-impose low-frequency neural activity that further amplifies these biases, distorting FC estimates. These findings indicate that accepted rsfMRI preprocessing pipelines systematically amplify spurious correlations and call for an improved statistical framework. This framework must explicitly account for autocorrelation, cyclicity, and multiple comparison biases, while excluding or correcting for structured neural activity that further distorts connectivity estimates.

Fig. 1.

The effect of sampling rate on the p-value of correlation coefficients. When the sampling rate exceeds what is needed to adequately describe the information content of the signal, it artificially inflates the statistical significance of correlations. This example illustrates three distinct signals: the first oscillates at a frequency of 1.5 cpm (black), the second also oscillates at 1.5 cpm but with a phase shift relative to the first signal (orange), and the third consists of multiple oscillations at frequencies of 0.2 cpm, 0.5 cpm, 1.4 cpm, and 1.5 cpm with randomized phases (red). At the original sampling rate of 1 Hz, all three signals (A-C) show significant correlations with each other at the Bonferroni-corrected level. However, when the sampling rate is reduced to 0.1 Hz, comparisons among the signals (D-F) reveal no correlations that pass the Bonferroni significance threshold (p ≤ 0.0167). Notably, the actual relationships between the signals remain unchanged, as the sampling rate has minimal impact on the correlation coefficients themselves in these oscillation frequencies. Instead, the inflated p-values in the 1 Hz sampling condition are due to the excessive sample size. This example underscores the importance of aligning the sampling rate with the signal’s intrinsic information content to avoid artificially inflated statistical significance.

Fig. 2.

Comparison of the theoretical studentized t-distribution against filtered, down-sampled, and surrogate data using AR(1) data. (A) Compared to the theoretical studentized t-distribution (grey line), the distribution of t-values of filtered data (red line) exhibits greater variation, substantially increasing the chance of type 1 errors. Down-sampling the filtered data (orange line) tightens the distribution of t-values offering a closer approximation to the theoretical t-distribution, however the tails are still too wide. (B) Using surrogate data (blue bars) more closely approximates the distribution of down-sampled t-statistics (orange line) compared to the theoretical t-distribution (grey line).

Fig. 3.

Proposed framework for rsfMRI preprocessing and functional connectivity (FC) estimation. This workflow explicitly incorporates key statistical corrections designed to minimize filtering-induced biases and correlation inflation. Following initial preprocessing—using only justified and clearly motivated filtering strategies—downsampling is performed whenever low-pass or bandpass filters are applied, preventing oversampling-induced artificial correlations. Surrogate datasets are generated ideally from the original, unfiltered data to accurately preserve underlying statistical (ARMA) properties. These surrogate data are then processed with identical filtering and downsampling procedures as the empirical data, creating robust statistical controls. Such matched surrogate data facilitate direct, statistically rigorous comparisons, thereby ensuring the reliability and validity of functional connectivity estimates. *By accounting for filtering and sampling biases, this pipeline significantly reduces preprocessing- driven artifacts and improves the robustness of rsfMRI analyses. Note: Standard preprocessing steps should still include controls for motion, physiological noise, and scanner artifacts. Fully documented scripts demonstrating practical implementation of this pipeline are available in the Supplementary Material (see Supplementary Methods, sections "Code for Table 1," "Code for Fig. 1," and "Code for Supplemental Tables S1 and S2").