Statistical quantification of confounding bias in machine learning models

Distinct cognitive and functional connectivity features from healthy cohorts can identify clinical obsessive-compulsive disorder

Improving diagnostic accuracy of obsessive-compulsive disorder (OCD) using models of brain imaging data is a key goal of the field, but this objective is challenging due to the limited size and phenotypic depth of clinical datasets. Leveraging the phenotypic diversity in large non-clinical datasets such as the UK Biobank (UKBB), offers a potential solution to this problem. Nevertheless, it remains unclear whether classification models trained on non-clinical populations will generalise to individuals with clinical OCD. This question is also relevant for the conceptualisation of OCD; specifically, whether the symptomology of OCD exists on a continuum from normal to pathological. Here, we examined a recently published "meta-matching" model trained on functional connectivity data from five large normative datasets (N=45,507) to predict cognitive, health and demographic variables. Specifically, we tested whether this model could classify OCD status in three independent clinical datasets (N=345). We found that the model could identify out-of-sample OCD individuals. Notably, the most predictive functional connectivity features mapped onto known cortico-striatal abnormalities in OCD and correlated with genetic brain expression maps previously implicated in the disorder. Further, the meta-matching model relied upon estimates of cognitive functions, such as cognitive flexibility and inhibition, to successfully predict OCD. These findings suggest that variability in non-clinical brain and behavioural features can discriminate clinical OCD status. These results support a dimensional and transdiagnostic conceptualisation of the brain and behavioural basis of OCD, with implications for research approaches and treatment targets.

An externally validated resting-state brain connectivity signature of pain-related learning

Pain can be conceptualized as a precision signal for reinforcement learning in the brain and alterations in these processes are a hallmark of chronic pain conditions. Investigating individual differences in pain-related learning therefore holds important clinical and translational relevance. Here, we developed and externally validated a novel resting-state brain connectivity-based predictive model of pain-related learning. The pre-registered external validation indicates that the proposed model explains 8-12% of the inter-individual variance in pain-related learning. Model predictions are driven by connections of the amygdala, posterior insula, sensorimotor, frontoparietal, and cerebellar regions, outlining a network commonly described in aversive learning and pain. We propose the resulting model as a robust and highly accessible biomarker candidate for clinical and translational pain research, with promising implications for personalized treatment approaches and with a high potential to advance our understanding of the neural mechanisms of pain-related learning.

Data leakage inflates prediction performance in connectome-based machine learning models

Predictive modeling is a central technique in neuroimaging to identify brain-behavior relationships and test their generalizability to unseen data. However, data leakage undermines the validity of predictive models by breaching the separation between training and test data. Leakage is always an incorrect practice but still pervasive in machine learning. Understanding its effects on neuroimaging predictive models can inform how leakage affects existing literature. Here, we investigate the effects of five forms of leakage-involving feature selection, covariate correction, and dependence between subjects-on functional and structural connectome-based machine learning models across four datasets and three phenotypes. Leakage via feature selection and repeated subjects drastically inflates prediction performance, whereas other forms of leakage have minor effects. Furthermore, small datasets exacerbate the effects of leakage. Overall, our results illustrate the variable effects of leakage and underscore the importance of avoiding data leakage to improve the validity and reproducibility of predictive modeling.

Machine learning and artificial intelligence in neuroscience: A primer for researchers

Artificial intelligence (AI) is often used to describe the automation of complex tasks that we would attribute intelligence to. Machine learning (ML) is commonly understood as a set of methods used to develop an AI. Both have seen a recent boom in usage, both in scientific and commercial fields. For the scientific community, ML can solve bottle necks created by complex, multi-dimensional data generated, for example, by functional brain imaging or *omics approaches. ML can here identify patterns that could not have been found using traditional statistic approaches. However, ML comes with serious limitations that need to be kept in mind: their tendency to optimise solutions for the input data means it is of crucial importance to externally validate any findings before considering them more than a hypothesis. Their black-box nature implies that their decisions usually cannot be understood, which renders their use in medical decision making problematic and can lead to ethical issues. Here, we present an introduction for the curious to the field of ML/AI. We explain the principles as commonly used methods as well as recent methodological advancements before we discuss risks and what we see as future directions of the field. Finally, we show practical examples of neuroscience to illustrate the use and limitations of ML.

The effects of data leakage on connectome-based machine learning models

Predictive modeling has now become a central technique in neuroimaging to identify complex brain-behavior relationships and test their generalizability to unseen data. However, data leakage, which unintentionally breaches the separation between data used to train and test the model, undermines the validity of predictive models. Previous literature suggests that leakage is generally pervasive in machine learning, but few studies have empirically evaluated the effects of leakage in neuroimaging data. Although leakage is always an incorrect practice, understanding the effects of leakage on neuroimaging predictive models provides insight into the extent to which leakage may affect the literature. Here, we investigated the effects of leakage on machine learning models in two common neuroimaging modalities, functional and structural connectomes. Using over 400 different pipelines spanning four large datasets and three phenotypes, we evaluated five forms of leakage fitting into three broad categories: feature selection, covariate correction, and lack of independence between subjects. As expected, leakage via feature selection and repeated subjects drastically inflated prediction performance. Notably, other forms of leakage had only minor effects (e.g., leaky site correction) or even decreased prediction performance (e.g., leaky covariate regression). In some cases, leakage affected not only prediction performance, but also model coefficients, and thus neurobiological interpretations. Finally, we found that predictive models using small datasets were more sensitive to leakage. Overall, our results illustrate the variable effects of leakage on prediction pipelines and underscore the importance of avoiding data leakage to improve the validity and reproducibility of predictive modeling.

Machine Learning and Prediction in Fetal, Infant, and Toddler Neuroimaging: A Review and Primer

Predictive models in neuroimaging are increasingly designed with the intent to improve risk stratification and support interventional efforts in psychiatry. Many of these models have been developed in samples of children school-aged or older. Nevertheless, despite growing evidence that altered brain maturation during the fetal, infant, and toddler (FIT) period modulates risk for poor mental health outcomes in childhood, these models are rarely implemented in FIT samples. Applications of predictive modeling in children of these ages provide an opportunity to develop powerful tools for improved characterization of the neural mechanisms underlying development. To facilitate the broader use of predictive models in FIT neuroimaging, we present a brief primer and systematic review on the methods used in current predictive modeling FIT studies. Reflecting on current practices in more than 100 studies conducted over the past decade, we provide an overview of topics, modalities, and methods commonly used in the field and under-researched areas. We then outline ethical and future considerations for neuroimaging researchers interested in predicting health outcomes in early life, including researchers who may be relatively new to either advanced machine learning methods or using FIT data. Altogether, the last decade of FIT research in machine learning has provided a foundation for accelerating the prediction of early-life trajectories across the full spectrum of illness and health.

Predictive modeling of optimism bias using gray matter cortical thickness

People have been shown to be optimistically biased when their future outcome expectancies are assessed. In fact, we display optimism bias (OB) toward our own success when compared to a rival individual's (personal OB [POB]). Similarly, success expectancies for social groups we like reliably exceed those we mention for a rival group (social OB [SOB]). Recent findings suggest the existence of neural underpinnings for OB. Mostly using structural/functional MRI, these findings rely on voxel-based mass-univariate analyses. While these results remain associative in nature, an open question abides whether MRI information can accurately predict OB. In this study, we hence used predictive modelling to forecast the two OBs. The biases were quantified using a validated soccer paradigm, where personal (self versus rival) and social (in-group versus out-group) forms of OB were extracted at the participant level. Later, using gray matter cortical thickness, we predicted POB and SOB via machine-learning. Our model explained 17% variance (R² = 0.17) in individual variability for POB (but not SOB). Key predictors involved the rostral-caudal anterior cingulate cortex, pars orbitalis and entorhinal cortex-areas that have been associated with OB before. We need such predictive models on a larger scale, to help us better understand positive psychology and individual well-being.

Confound-leakage: confound removal in machine learning leads to leakage

BACKGROUND

Machine learning (ML) approaches are a crucial component of modern data analysis in many fields, including epidemiology and medicine. Nonlinear ML methods often achieve accurate predictions, for instance, in personalized medicine, as they are capable of modeling complex relationships between features and the target. Problematically, ML models and their predictions can be biased by confounding information present in the features. To remove this spurious signal, researchers often employ featurewise linear confound regression (CR). While this is considered a standard approach for dealing with confounding, possible pitfalls of using CR in ML pipelines are not fully understood.

RESULTS

We provide new evidence that, contrary to general expectations, linear confound regression can increase the risk of confounding when combined with nonlinear ML approaches. Using a simple framework that uses the target as a confound, we show that information leaked via CR can increase null or moderate effects to near-perfect prediction. By shuffling the features, we provide evidence that this increase is indeed due to confound-leakage and not due to revealing of information. We then demonstrate the danger of confound-leakage in a real-world clinical application where the accuracy of predicting attention-deficit/hyperactivity disorder is overestimated using speech-derived features when using depression as a confound.

CONCLUSIONS

Mishandling or even amplifying confounding effects when building ML models due to confound-leakage, as shown, can lead to untrustworthy, biased, and unfair predictions. Our expose of the confound-leakage pitfall and provided guidelines for dealing with it can help create more robust and trustworthy ML models.

Statistical quantification of confounding bias in machine learning models

BACKGROUND

The lack of nonparametric statistical tests for confounding bias significantly hampers the development of robust, valid, and generalizable predictive models in many fields of research. Here I propose the partial confounder test, which, for a given confounder variable, probes the null hypotheses of the model being unconfounded.

RESULTS

The test provides a strict control for type I errors and high statistical power, even for nonnormally and nonlinearly dependent predictions, often seen in machine learning. Applying the proposed test on models trained on large-scale functional brain connectivity data (N= 1,865) (i) reveals previously unreported confounders and (ii) shows that state-of-the-art confound mitigation approaches may fail preventing confounder bias in several cases.

CONCLUSIONS

The proposed test (implemented in the package mlconfound; https://mlconfound.readthedocs.io) can aid the assessment and improvement of the generalizability and validity of predictive models and, thereby, fosters the development of clinically useful machine learning biomarkers.