Examining meta-matching as a tool to predict Obsessive-Compulsive Disorder

561 

Abstract Submission 

Luke Hearne¹, Andrew Zalesky², Paul Fitzgerald³, Oscar Murphy⁴, Ye Tian⁵, Minah Kim⁶, Sunah Choi⁷, Jun Soo Kwon⁸, Luca Cocchi⁹

¹QIMR Berghofer Medical Research Institute, Herston, Queensland, ²The University of Melbourne, Melbourne, Victoria, ³Australian National University, Caberra, ACT, ⁴Monash University, Melbourne, Victoria, ⁵University of Melbourne, Carlton South, Victoria, ⁶Department of Neuropsychiatry, Seoul National University Hospital, Seoul, Seoul, ⁷Seoul National University, Seoul, Seoul, ⁸Department of Brain and Cognitive Sciences, Seoul National University College of Natural Sciences, Seoul, Seoul, ⁹QIMR Berghofer Medical Research Institute, Brisbane, Queensland

Luke Hearne  
QIMR Berghofer Medical Research Institute
Herston, Queensland

Andrew ZALESKY, PhD  
The University of Melbourne
Melbourne, Victoria

Paul Fitzgerald  
Australian National University
Caberra, ACT

Oscar Murphy  
Monash University
Melbourne, Victoria

Ye Tian  
University of Melbourne
Carlton South, Victoria

Minah Kim  
Department of Neuropsychiatry, Seoul National University Hospital
Seoul, Seoul

Sunah Choi  
Seoul National University
Seoul, Seoul

Jun Soo Kwon  
Department of Brain and Cognitive Sciences, Seoul National University College of Natural Sciences
Seoul, Seoul

Luca Cocchi  
QIMR Berghofer Medical Research Institute
Brisbane, Queensland

A core goal of cognitive neuroscience is to predict individual differences in cognition and psychiatric symptoms from brain imaging data. A key question is whether large predictive models trained on general population datasets can generalize to patient populations. A recent approach, termed 'meta-matching' has furthered this idea (He et al., 2022) by leveraging the observation that most phenotypes are correlated. Therefore, brain features that can predict specific phenotypes in the large, population-based dataset are likely useful in predicting new phenotypes in an independent dataset.

In the current work, we test this hypothesis explicitly by attempting to classify OCD status in a clinical sample of moderate size (N=334). We started by contrasting the predictive ability of meta-matching and baseline logistic regression models. We investigate the predictive weights generated by the model and examine them in the cortico-striatal system, a well-known biological correlate of OCD (Naze et al., 2023).

The study used data pooled across three independent clinical datasets collected in Brisbane, Melbourne and Seoul (N = 334, nOCD = 189, nHC = 145) (Kim et al., 2019; Naze et al., 2023). Brain imaging data were preprocessed using fMRIprep (Esteban et al., 2018), and functional connectivity data, used as features in the predictive models, were estimated using Nilearn. We contrasted two models in their ability to classify OCD diagnoses correctly. The first baseline model used logistic regression, whereas the second model used the openly available meta-matching model (He et al., 2022). In short, this model is a fully connected feedforward deep neural network that uses functional connectivity data to predict 67 non-brain imaging phenotypes (Figure 1A). We tested both models in a 10-fold cross-validation scheme repeated 100 times. We calculated predictive feature weights (PFWs) for each connectivity edge using the Haufe transformation. Specifically, we tested four functional connectivity pathways associated with OCD (Naze et al., 2023) with seed regions stemming from the nucleus accumbens, dorsal caudate, ventral putamen and dorsal putamen.

Both meta-matching (mean accuracy = 62.76%) and logistic regression models (mean accuracy = 62.12%) demonstrated higher performance than shuffled permutations (p < 0.0001) (Figure 1B). When compared directly, the meta-matching model performed slightly better than the logistic regression (Cohen's d = 0.498).
Our analysis revealed that predictive feature weights (PFWs) were significantly lower than would be expected from the shuffled permutations for three of the four cortico-striatal pathways of interest (nucleus accumbens p = 0.045, dorsal caudate p < 0.001, dorsal putamen p < 0.001; ventral putamen p = 1.0) (Figure 2). Moreover, the dorsal caudate pathway had larger negative PFWs than would be expected from any random set of brain regions (cortical or subcortical, p = 0.008).

The recently proposed Meta-matching framework (He et al., 2022) is a promising technique for increasing the validity of predictive models trained in small clinical datasets. Meta-matching performance was similar to a typical logistic regression model; the small prediction boost (< 1%) is not clinically significant and is substantially smaller than the performance noted in the original paper (He et al., 2022). Despite the above caveats, the performance of the meta-matching model is impressive, given the lack of OCD patients and OCD-relevant measurements (e.g., symptom scales) available in the training dataset. In line with prior work, cortico-striatal pathways were disrupted in OCD and, therefore among the best predictors for classifying group status. To improve model performance it is likely that meta-matching models will need to be trained for specific purposes, for example, models that tap into variance associated with mental health symptoms. Such models will likely increase the predictive power of this method.

Psychiatric (eg. Depression, Anxiety, Schizophrenia) ¹

Classification and Predictive Modeling ²

Connectivity (eg. functional, effective, structural)

Multivariate Approaches

Machine Learning

Obessive Compulsive Disorder

Psychiatric

Psychiatric Disorders