Excessive data censoring in fMRI undermines individual precision and weakens brain-behavior associations

TL;DR

It is demonstrated that unreliable FC substantially attenuates BWAS correlations: by ~30% under optimal conditions (longer ICA-FIX scans with no censoring) but exceeding 75% in short, aggressively censored scans to maximize fidelity of FC and BWAS.

Abstract

Censoring high-motion volumes in fMRI is common practice to reduce effects of head motion on functional connectivity (FC). Although aggressive censoring removes more noise, it causes extensive data loss, creating a tradeoff that may ultimately improve or degrade FC accuracy. Here, we evaluate how censoring affects FC estimation and downstream brain-wide association studies (BWAS). Using extensively sampled participants from the Human Connectome Project (HCP) Retest dataset, we establish individual "ground truth" FC and assess the accuracy of FC estimated from 5-30 minute scans. We find that censoring degrades FC accuracy, with more aggressive censoring being more detrimental, particularly among participants exhibiting above-average motion. In these participants, aggressive censoring reduces FC accuracy by 30% for 30-minute scans denoised with ICA-FIX, an advanced denoising method, and by 3% for scans denoised with conventional confound regression. These effects reflect substantial data loss (34%) that outweighs comparatively modest noise reductions: 7% with ICA-FIX and 18% with confound regression. Compensating for this would require substantially longer scans (62% with confound regression; 76% with ICA-FIX), inflating data collection budgets. Introducing a repeated measures framework to separate motion trait from artifact, we find that standard QC metrics are dominated by motion trait and overstate motion bias, which is effectively mitigated with less aggressive censoring. Finally, using data from nearly 1,000 HCP participants, we demonstrate that unreliable FC substantially attenuates BWAS correlations: by ~30% under optimal conditions (longer ICA-FIX scans with no censoring) but exceeding 75% in short, aggressively censored scans. Our findings support the use of advanced denoising methods, limiting censoring, and collecting longer scans to maximize fidelity of FC and BWAS.

Figures (14)

• Figure 1: More aggressive censoring worsens accuracy of individual FC estimates. FC error is relative to individuals' "ground truth" based on 90 minutes of stringently censored data per participant. Root mean squared error (rMSE) is over participants and edges in panels A and B and over participants for each edge in panel C. Panels B and C show change in rMSE compared to no censoring. (A) FC error decreases with longer acquired scan duration, but more censoring leads to progressively higher error across all durations in both 36P-processed and FIX-processed data. FC error is much lower overall in FIX-processed data than in 36P-processed data. (B) Longer scans do not mitigate the detrimental effects of censoring on FC accuracy; rather, the gap between censoring and no censoring widens with increasing scan duration. This reflects an increase in the censoring rate as the scan progresses, due to higher participant motion levels later in the scan. Censoring is particularly harmful in FIX-processed data, where expanded censoring results in over 20% higher FC error for 30-minute scans. (C) Across edges, the effects of censoring are mixed for 36P-processed data, with most edges worsening but some improving ($T=10$ minutes). While the effect of expanded censoring in 36P data is mild on average across all edges, many edges show much stronger detrimental effects (indicated in dark red), particularly within-network edges and subcortical-cortical connections. Expanded censoring worsens FC error by at least $10\%$ in over $5\%$ of edges, while it improves FC error by the same amount in only $0.4\%$ of edges. For FIX-processed data, by contrast, censoring universally worsens FC accuracy across the connectome: over 71% of edges worsen by at least 10%, while effectively zero edges improve by the same amount.
• Figure 2: Censoring is more detrimental for participants who move more, due to excessive data loss. (A) For participants with above-average motion levels (see Figure \ref{['fig:motion_split']}), censoring increases FC error more than in low-motion participants. For FIX-processed data, the effects of expanded censoring are dramatic in these participants, worsening FC error by over 30% for 30-minute scans. (B) FC error is determined by the ratio of two competing factors: baseline noise and scan duration, both of which are typically decreased by censoring. (C) There is a dramatic loss in effective scan duration due to expanded censoring, particularly in high-motion participants, with data loss of 34% on average. (D) More aggressive censoring reduces baseline noise, but not as much as it reduces effective scan duration. For high-motion participants, expanded censoring reduces baseline noise by 14% on average in 36P-processed data and by 5% in FIX-processed data. One outlying participant with approximately 40% higher baseline noise in FIX-processed data is excluded for visualization purposes and from the average to avoid its undue influence. In low-motion participants, the reduction in baseline noise is mild in 36P-processed data and is non-existent in FIX-processed data. Across motion levels, any reductions in baseline noise is outweighed by more dramatic data loss, ultimately leading to higher FC error.
• Figure 3: Stringent and expanded censoring necessitate substantially longer scans to maintain FC accuracy. Values represent the percentage change in scan duration required to maintain edge-wise FC accuracy when more stringent motion censoring is used, compared with lenient censoring (see Figure \ref{['fig:durChange_illustration']} for illustration). (A) Percent change in required scan duration to maintain FC accuracy compared with lenient censoring. Mean over edges for each participant is shown, along with the mean over participants within each motion group. For high-motion participants, expanded censoring requires over 60% longer scan durations in 36P-processed data and over 75% longer scans in FIX-processed data on average. (B) Edge-wise percent change in required scan duration, averaged over participants. Color scale is truncated at $\pm 130\%$, but values range from -77% to +500%. Scan duration would need to at least double to maintain FC accuracy with expanded censoring for 13.5% of edges in 36P-processed data and 15.1% in FIX-processed data.
• Figure 4: Standard QC-FC is dominated by motion trait, while repeated-measures QC-FC isolates motion artifact, which stringent censoring is sufficient to mitigate.(A) Repeated-measures QC-FC separates between-participant effects representing motion trait (real between-participant differences in FC associated with motion tendency) from within-participant effects representing motion artifact (e.g. spin history artifacts induced by head motion leading to biased FC) and possible motion state (real within-participant fluctuations in FC associated with higher motion levels). Standard QC-FC, however, is unable to distinguish between between-participant and within-participant effects. As a result, it erroneously implies the persistence of strong motion artifacts even after stringent censoring, whereas the within-participant component of repeated-measures QC-FC shows that motion artifact is effectively mitigated after stringent censoring. (B) Negative distance-dependence of QC-FC, with more proximal connections showing greater associations with motion, is typically considered indicative of motion artifact. The within-participant component of repeated-measures QC-FC shows that in 36P-processed data, distance dependence of QC-FC is reduced by lenient censoring and removed stringent censoring. Expanded censoring, on the other hand, increases the magnitude of within-participant motion-FC associations and reverses the distance effect, suggesting that it is not simply removing motion artifact. In FIX-processed data, there is no negative distance-dependence of QC-FC. Censoring does reduce the magnitude of QC-FC in FIX-processed data, possibly due to the exclusion of epochs representing motion state.
• Figure 5: BWAS correlations are severely attenuated due to poor FC reliability and are worsened by expanded censoring and short scan duration.(A) Mathematically, BWAS correlations between imperfectly reliable brain ($X$) and behavioral ($Y$) measures are downwardly biased, and the amount of bias is determined by the reliability of both variables. We define BWAS proportional strength as the multiplicative downward bias in BWAS associations, ranging from $0$ (total attenuation) to $1$ (no attenuation). (B) Illustration of the relationship between reliability of FC and behavior and BWAS attenuation. For FC, ICC is computed based on 30-minute, 36P-processed scans with stringent censoring applied. Several behavioral variables are shown, ranging from high reliability (total cognition) to low reliability (dexterity) (see Figure \ref{['fig:ICC_demo']} for ICC values of additional behavioral variables in the HCP). (C) Empirical BWAS attenuations between FC and total cognition by scan duration and censoring level, relative to "ground truth", bias-corrected BWAS correlations (Methods Section \ref{['methods:BWAS:empirical']}). Values correspond to the mean across all edges. Total cognition is the most reliable behavioral variable in the HCP, so these results represent a best-case scenario. BWAS attenuation is severe in 36P-processed data, with well over 50% attenuation for scans up to 10 minutes without expanded censoring, and over 75% attenuation for 5-minute scans with expanded censoring. Across all scan durations, expanded censoring severely worsens BWAS attenuation in both 36P- and FIX-processed data. In FIX-processed data, BWAS is less but still substantially attenuated, with approximately 50% attenuation in 5-minute scans without expanded censoring and approximately 60% attenuation with expanded censoring. The least BWAS attenuation, approximately 30%, is seen with 30-minute FIX-processed scans with no censoring.