Authors: Qian Lv; Kristina Zeljic; Shaoling Zhao; Jiangtao Zhang; Jianmin Zhang; Zheng Wang · Research

How Can Machine Learning Help Diagnose Obsessive-Compulsive Disorder?

This article explores how machine learning applied to brain imaging data could improve diagnosis of obsessive-compulsive disorder.

Source: Lv, Q., Zeljic, K., Zhao, S., Zhang, J., Zhang, J., & Wang, Z. (2023). Dissecting Psychiatric Heterogeneity and Comorbidity with Core Region‑Based Machine Learning. Neuroscience Bulletin, 39(8), 1309-1326. https://doi.org/10.1007/s12264-023-01057-2

What you need to know

  • Machine learning applied to brain imaging data shows promise for improving diagnosis of obsessive-compulsive disorder (OCD)
  • Focusing on key “core regions” of the brain implicated in OCD may improve machine learning accuracy
  • More research is needed before these techniques can be used clinically, but they offer hope for more objective OCD diagnosis in the future

How is OCD currently diagnosed?

Obsessive-compulsive disorder (OCD) is a complex mental health condition that affects 1-2% of people worldwide. It is characterized by intrusive, unwanted thoughts (obsessions) and repetitive behaviors or mental acts (compulsions) that a person feels driven to perform to reduce anxiety.

Currently, OCD is diagnosed based on a patient’s reported symptoms and behaviors, as outlined in the Diagnostic and Statistical Manual of Mental Disorders (DSM). A clinician will ask about the types of obsessions and compulsions a person experiences, how much time they spend on them, and how much they interfere with daily life.

While this symptom-based approach has been the gold standard, it has some limitations. Different patients may describe similar symptoms in different ways. Clinicians may interpret symptoms differently. And importantly, the underlying brain changes in OCD may be present before obvious symptoms appear.

This is why researchers are exploring whether brain imaging and machine learning could provide a more objective way to diagnose OCD and track treatment progress.

How could machine learning help diagnose OCD?

Machine learning refers to computer algorithms that can learn patterns from data without being explicitly programmed. When applied to brain imaging data, machine learning algorithms can pick up on subtle patterns that may not be visible to the human eye.

The typical process works like this:

  1. Collect brain scans from people with and without OCD
  2. Extract relevant features from the brain scans (e.g. size or activity of certain brain regions)
  3. Train a machine learning algorithm to differentiate between the OCD and non-OCD brain scans
  4. Test the algorithm on new brain scans to see how accurately it can identify OCD

If successful, this could lead to a computer program that could analyze a brain scan and give a probability of whether that person has OCD. This could potentially help with earlier or more accurate diagnosis.

What brain imaging techniques are used?

The three main types of brain imaging used in OCD machine learning studies are:

  1. Structural MRI - Provides detailed images of brain anatomy
  2. Functional MRI - Measures brain activity by detecting changes in blood flow
  3. Diffusion MRI - Shows how water molecules move along white matter tracts, indicating how different brain regions are connected

Each of these techniques provides different information about the brain. Combining data from multiple imaging types may provide the most complete picture.

What have machine learning studies found so far?

Over 20 studies have used machine learning to try to distinguish OCD from non-OCD brains, with accuracies ranging from 71-95%. Some key findings include:

  • Structural differences in regions like the orbitofrontal cortex, anterior cingulate cortex, and caudate nucleus can help identify OCD
  • Altered functional connectivity, especially between frontal and striatal regions, is a key feature
  • White matter tract differences, particularly in frontal-striatal connections, also contribute to accurate classification

However, most studies have been relatively small, typically with fewer than 100 participants in each group. Larger studies are needed to confirm these findings.

What are the challenges?

While initial results are promising, there are still significant challenges to overcome before machine learning diagnosis could be used clinically:

  • Studies use different methods, making it hard to compare results
  • Most studies have not been replicated in independent samples
  • It’s unclear how well the algorithms would work on more diverse populations
  • OCD often co-occurs with other conditions, which could confuse the algorithms
  • The high-dimensional nature of brain imaging data makes overfitting a risk

Perhaps most importantly, when one large multi-site study tried to apply machine learning to structural MRI data from over 2,300 OCD patients and 2,000 controls, they achieved only 50% accuracy - no better than chance.

The core regions approach

To address some of these challenges, the authors of this review propose focusing machine learning efforts on “core regions” of the brain known to be central to OCD pathology. This could help reduce noise in the data and improve accuracy.

Evidence for core OCD regions comes from several sources:

  • Brain imaging studies consistently show structural and functional changes in regions like the orbitofrontal cortex, anterior cingulate cortex, and striatum
  • Neurosurgical procedures targeting these regions can reduce OCD symptoms in severe cases
  • Animal studies show that manipulating activity in these regions can produce or reduce OCD-like behaviors

By zeroing in on these core regions rather than looking at the whole brain, machine learning algorithms may be able to more reliably detect the key differences between OCD and non-OCD brains.

How would researchers identify core regions?

The authors suggest several approaches to identify core regions for machine learning:

  1. Literature review - Compile regions consistently implicated across many OCD studies

  2. Meta-analysis - Statistically combine results from multiple studies to find the most robust differences

  3. Animal models - Use regions identified in animal studies of OCD-like behaviors

  4. Treatment response - Focus on regions that change after successful OCD treatment

  5. Cross-species validation - Use regions that distinguish OCD-like animals to build human classifiers

Once core regions are identified, machine learning algorithms would be trained using only data from those specific brain areas rather than whole-brain data.

Beyond diagnosis: Predicting treatment response

In addition to diagnosis, machine learning may also help predict which treatments will work best for individual patients. A few studies have used baseline brain scans to predict response to cognitive behavioral therapy or medication for OCD, with moderate success.

As more data is collected on brain changes during OCD treatment, machine learning algorithms may be able to suggest personalized treatment plans based on a patient’s unique brain patterns.

A transdiagnostic approach

OCD symptoms often overlap with other conditions like anxiety disorders, eating disorders, and autism spectrum disorders. Some researchers suggest taking a “transdiagnostic” approach - looking at specific symptoms or behaviors across diagnostic categories.

For example, machine learning could be used to identify brain patterns associated with compulsive behaviors in general, rather than OCD specifically. This may lead to a more nuanced understanding of these symptoms and more targeted treatments.

Conclusions

  • Machine learning applied to brain imaging data shows promise for improving OCD diagnosis, but more research is needed before clinical use
  • Focusing on core brain regions implicated in OCD may improve machine learning accuracy
  • Future directions include predicting treatment response and taking a transdiagnostic approach to compulsive behaviors
  • While challenges remain, this approach offers hope for more objective and personalized mental health care in the future

Machine learning and neuroimaging are powerful tools, but they are not meant to replace clinical judgment. Rather, they may provide additional objective information to aid in diagnosis and treatment planning. As methods improve, they have the potential to significantly advance our understanding and treatment of complex conditions like OCD.

Back to Blog

Related Articles

View All Articles »