Table 1 Sources of dirty data.
From: Predicting the future of neuroimaging predictive models in mental health
Category | Problem | Examples |
|---|---|---|
Phenotypic measures | Measures are subjective | • Poor inter-rater reliability and high variability in gold-standard diagnostic tools and behavioral measures [33,34,35, 81] |
Measures are nonspecific | • High false-positive rate on ADOS in adults with schizophrenia [36] | |
Measures focus on the tails of behavior | • Healthy controls will be zero inflated on questionnaire data [37, 38] | |
Participants | Comorbidity | • Symptoms of psychiatric disorders often overlap across diagnoses, while the majority of predictive models in psychiatry rely on more binary classification approaches |
Medication | • Psychiatric medications have the ability to alter BOLD signal patterns. This becomes difficult to study the psychiatric phenomena of interest as signals are confounded | |
Episodic symptoms | • Symptoms change as a function of disease state. Data from scans based on one day may be vastly different in brain states relative to scans based on another day | |
Data collection | Multi-site | • Inter-scanner differences can induce significant variability [82, 83], and the complexity of the data analysis workflows could affect reproducibility [84] |
Missing data | • Subjects not completing questionnaires • Inability to complete behavioral testing or scan sessions in clinical populations [85, 86] |
