The Bias Associated with Movement-related Data Exclusion in Psychosis MRI Research

Magnetic resonance imaging (MRI) is now an indispensable tool in psychiatric research. It offers detailed insights into brain structure and function, with the ability to probe into deeper regions as well, compared to many other investigational tools. Over the last couple of decades, results from MRI-based studies have advanced our understanding of brain alterations in psychiatric disorders like psychosis. Psychotic illnesses like schizophrenia are now recognized as brain-based diseases with distinct alterations in structural measures like volume and thickness, functional connectivity between different cortical and subcortical areas, and white matter integrity. ¹

A fundamental methodological challenge for MRI studies in psychiatric patients is the presence of head motion during scanning. This is especially pronounced in patients with severe mental illnesses like schizophrenia and bipolar affective disorder. Excessive head motion can affect MRI data quality, potentially misleading the subsequent analysis and interpretation of the results. To mitigate this, researchers exclude scans of participants with high motion during the initial stages of analysis to obtain cleaner datasets. This practice, however, raises concerns about bias and the generalizability of results. ² In this article, we discuss the problem of motion-related exclusion of MRI Neuroimaging studies in psychosis.

Discussion

Head motion in psychotic patients during scan acquisition and its correction

Head movement in MRI scans is a problem that can affect any study population, but it is especially common in psychiatric patients. Even subtle head movements on the order of a few millimeters can have deleterious effects on MRI data. ² Patients with psychotic disorders exhibit significantly more head movement in the scanner compared to healthy controls. ³ This increased movement can be due to multiple reasons. These include difficulty following instructions due to disorganization, restlessness due to psychomotor agitation or anxiety, exaggerated discomfort inside the scanner environment due to paranoia or claustrophobia, or side effects of medication (akathisia). This can affect the quality of structural, functional, and diffusion tensor images. In functional MRI data, movement can shift the location of the signal from a given voxel to a different location, creating artifactual patterns in connectivity measures. ² In structural MRI, movement can result in blurring and ghosting artifacts that can affect various anatomical measurements. ⁴ Significant movement can similarly affect the diffusion MRI metrics of white matter microstructure. ⁵

Head movement is, therefore, a systematic artifact that leads to false results if not accounted for. Several strategies are currently followed to manage movement during scans. These include providing clear instructions during patient preparation and using physical restraints, such as foam padding around the head, to minimize movements. Other employed methods include practice mock scan sessions, providing reward incentives, and displaying media content during scan breaks. While these methods are helpful, they do not completely eliminate head motion, especially in the psychiatric population.

Recent technological advancements in MRI acquisition have facilitated real-time tracking and correction of motion during scanning. Such scanners integrate prospective motion correction during the scan by updating the slicing acquisition coordinates based on movement. ⁶ In another real-time monitoring approach, the system monitors a subject’s head motion frame-by-frame. It can signal the operator to pause the scan or extend it until sufficient low-motion data are acquired. ⁷ Despite these innovations, real-time motion correction tools are not yet in widespread use due to complexity and hardware limitations in setting them up.

Therefore, most studies rely on retrospective motion correction algorithms after data acquisition. Volume realignment is a commonly employed technique in which each volume is registered to a reference volume from the same subject. FSL’s (FMRIB Software Library) MCFLIRT tool, AFNI’s (Analysis of Functional Neuro Images) 3dvolreg command, and others are examples of realignment tools commonly used. This step in preprocessing corrects small movements between volumes. However, it cannot fully recover data lost due to larger movements or intra-volume motion (movements that occur during the acquisition of a single volume). Therefore, quality control (QC) procedures are implemented to identify and censor data segments or subjects with excessive motion. ⁸ A common practice is motion scrubbing, in which volumes with motion beyond a certain threshold are removed from the fMRI time series before analysis. Framewise displacement (FD) is a metric used for this process. FD quantifies head movement between successive fMRI volumes in millimeters. A conventional FD threshold is 0.5 mm, and volumes where FD > 0.5 mm are labeled as high-motion outliers and scrubbed from the data. After censoring bad frames, interpolation is done to account for the missing data points in statistical models. If too many volumes are scrubbed (e.g., a commonly used threshold is 20%), that entire scan or subject may be excluded from further analysis. ⁹ Another approach for scrubbing is to estimate the change in BOLD signal variance across the brain during the entire scanning window (DVARS). Volumes with a change in variance beyond a cutoff are discarded. ⁹ Some studies exclude scans with maximum head displacements larger than a threshold (e.g., 3 mm). These QC thresholds, while necessary to ensure data quality, can lead to the disproportionate loss of participants from the patient group.

In-scanner Movement as a Possible Behavioral Phenotype in Psychosis

Head movements during scan, rather than just being random noise that corrupts the data, may in fact carry important information about the study population. Patients who struggle to remain still for an MRI could represent a distinct behavioral or neurobiological subtype of psychosis. For example, marked restlessness or inability to comply with scanner instructions may be a proxy for high levels of psychomotor agitation, anxiety, or disorganization. Similarly, severe paranoia might make it challenging for a patient to tolerate the confined, noisy scanner environment, leading to increased movements. These symptoms typically indicate a more severe or acute presentation, and excluding scans from these subjects necessarily biases the study population, thereby excluding such a phenotype. ¹⁰

In healthy humans, there is some empirical evidence to suggest that in-scanner motion reflects stable individual differences that are heritable and may be related to personality characteristics.^11–13 However, in the specific context of psychotic-spectrum illness, there are not many studies that have thoroughly examined the association of illness-related variables with head movements in an MRI scanner. However, it remains plausible that specific illness characteristics can affect motion. Agitation (psychomotor excitement) is an obvious candidate since an acutely agitated or hyperactive patient will have trouble staying still. Disorganized behavior or thought might also indirectly contribute, as a disorganized patient may be less able to adhere to instructions, leading to increased movement. Poor insight might also play a role if a patient does not believe the scan is important or does not understand the need to remain still. Therefore, in all likelihood, the excluded data after quality checks will disproportionately have a larger representation of psychotic patients who are acutely ill with significant agitation and disorganized behavior.

Statistical Issues with Missing Data Exclusion that Are Not Random

The practice of excluding high-motion participant scans introduces a form of missing data that is not a random event. In statistical terms, it is referred to as missing not at random (MNAR), where the probability of data being missing is related to the underlying data itself, rather than due to random reasons. For example, suppose the most severe psychosis patients are more likely to produce unusable scans (due to motion). In that case, missingness is directly related to illness severity, and removing those patients systematically shifts the study sample to the less severe end of the spectrum. The resulting sample of data is no longer representative of the full patient population. This limits the generalizability of the findings to the original study population and biases the result toward a subpopulation obtained after excluding specific behavioral phenotypes.

From a statistical perspective, MNAR data violates the assumptions of most inferential statistical approaches. Traditional group comparisons (t-tests, ANOVAs) and regression models implicitly assume that any missing data are ignorable (i.e., missing due to random reasons). If the data are MNAR, these analyses yield biased parameter estimates and invalid inferences in hypothesis testing.

As a hypothetical example, suppose in the full schizophrenia population, the average hippocampal volume is smaller than in healthy controls by a specific value. After quality checks, suppose scans with larger head movements are discarded. If these discarded scans belong to the patient subtype with more severe and disorganized illness, exclusion of this group will bias the estimated average to a larger value. This leads to an underestimation of the true effect in the study.

Strategies for Mitigating Motion Exclusion-related Biases

Considering the potential bias that can be introduced by excluding scans from subjects with large head movements, it is crucial to be aware of this effect during subsequent data processing and analysis.

Apart from the strategies mentioned earlier to mitigate head movements, methods should be employed to retain and correct data as much as possible (Figure 1). One obvious approach is to include motion parameters as covariates in group-level statistical models. Keeping metrics such as mean FD or the number of removed frames for each subject as covariates allows for statistical adjustment for residual motion effects when comparing patients versus controls or correlating with clinical variables. ¹⁴ However, covariate adjustment assumes a linear relationship between motion and BOLD signals and cannot eliminate artifacts due to complex non-linear associations with movement (seen with larger movements). ¹⁵

Figure 1. Methods to Mitigate and Correct Head Movement-related Artifacts in Various Stages of MRI Research.

Another approach is to apply more denoising techniques that model motion artifacts and remove them, rather than scrubbing data. Independent component analysis (ICA) can be used to identify noise components associated with motion and regress them out of the data without discarding entire volumes. ¹⁶ FSL’s ICA-Based X-noiseifier (FIX) and ICA-AROMA (Automatic Removal of Motion Artifacts) are algorithms that are routinely used to remove motion-related independent components. ¹⁶ However, when head movement is large, ICA-based algorithms label many components as noise and remove them. This will effectively result in the loss of many independent time points for subsequent analysis. ¹⁷

Another recent approach that is still experimental is the use of algorithms to impute missing timepoints due to large motion in fMRI data. These algorithms are machine learning-based predictive models and are yet to be adopted into routine data processing pipelines. ¹⁸

A study systematically compared 14 retrospective motion correction pipelines and found that those combining various strategies of signal regression and volume scrubbing reduced the fraction of connectivity edges contaminated by motion to <1%. ¹⁴ In comparison, most of the edges were still biased by noise when using only a simple rigid body motion correction. These combination pipelines also recovered a clearer network modular structure and an improvement in other metrics used to evaluate movement noise contamination (e.g., QC-FC correlation), indicating that they removed structured noise rather than distorting neural signals. The study also showed that no model completely abolished motion-related variance, so scans with excessive movement remain a source of potential movement bias.

The methods discussed above are specific to movement correction in functional MRI data. Functional MRI data, due to multiple temporal sampling, allows for the modeling of noise and its removal. Structural and Diffusion tensor images lack this redundancy, making retrospective motion correction even more difficult.

Conclusions

Excluding MRI data with large head movements can bias the results by systematically removing the most severely affected individuals with psychosis. It is therefore crucial to address motion-related issues at every stage— from acquisition to subsequent proces-sing and analysis.