Rise of the machine learning algorithm: the future of diagnosing schizophrenia?

7232530118_300b9ed694_b

Papa Elf sent me a paper (Kambeitz et al, 2015) looking at a meta-analysis of a technique involving brain imaging to try and successfully classify study participants as meeting the criteria for schizophrenia or not.

The study uses a type of diagnostic screening tool using brain imaging combined with “multi-variate pattern recognition” which usually includes a “machine learning algorithm” to improve predictive accuracy. This involves looking at interconnected regions of the brain for patterns that tend to be associated with the diagnosis of schizophrenia in the participants in the study, which differentiates them from participants who did not meet the criteria. If a machine learning algorithm program is used, this will keep trying to improve its’ performance by successfully classifying the subjects.

It’s important to remember that these changes in structure are not necessarily the cause of schizophrenia, but may be the result of a process (not necessarily biological) leading to a diagnosis of schizophrenia.

This is a meta-analysis looking at the effectiveness of a diagnostic screening tool for schizophrenia using multivariate pattern recognition techniques on brain images. This is part of the largely unsuccessful quest for clinically useful biomarkers in psychiatry for “functional” disorders (i.e. no obvious “organic” cause) like schizophrenia.

Fortunately for me, the EBMH Journal recently published an article about how to appraise a meta-analysis of diagnostic screening tools, so I have used this approach to help write my blog (Takwoingi et al, 2015).

Could neuroimaging become part of the routine diagnostic process in the future?

Could neuroimaging become part of the routine diagnostic process in the future?

Methods

I will quote the paper directly as to which studies are included:

To be included in the meta-analysis a paper needed to report results of a neuroimaging-based multivariate classification model separating patients with schizophrenia from HC {Healthy Controls}. We included all available multivariate approaches such as support-vector machines, random forests, discriminant analysis, logistic regression, neural networks, as well as combinations thereof.

There is also a further condition that results must be expressed in a way to allow calculation of “true positives (TP), true negatives (TN), false positives (FP), and false negatives (FN)” (or the authors had responded to requests for this data).

Reasonably comprehensive search strategies were used and the quality of the studies was assessed using a reasonably established technique. Data extraction was done by two researchers separately and any disagreements resolved using a conference consensus. A bivariate model with random effects was used to analyse the statistics and an appropriate statistics package was used (Takwoingi et al, 2015). Studies were grouped according to the type of brain imaging used and combined for an overall result.

Studies from 1991-2013 were used, but crucially little information in the paper was given as to which diagnostic criteria were used for schizophrenia (I suspect DSM-IV for the majority of studies, but this is just an informed guess on my part).

Secondary analysis looked at various patient characteristics to see if they influenced sensitivity or specificity of the tests such as age, gender, severity of symptoms, first episode vs chronic schizophrenia, or amount of medication prescribed.   Differences in performances between different imaging techniques were examined.

Results

399 initial papers were whittled down to 38 studies with a total of 1,602 participants who met criteria for schizophrenia and 1,637 healthy controls. The data from these 38 studies was then entered into the analysis.

Sensitivity in this paper is the proportion of people that the screening test correctly identifies as meeting criteria for schizophrenia, expressed as a percentage. Specificity is the proportion of people the screening test correctly identifies as not meeting criteria for schizophrenia, expressed as a percentage.

To quote the paper:

Across all studies, neuroimaging-based classifiers separated SZ [schizophrenia] from HC {healthy controls} with a sensitivity of 80.3% (95% CI: 76.7 to 83.5%) and a specificity of 80.3% (95% CI: 76.9 to 83.3%).

In the secondary analysis, older patients demonstrated higher sensitivity; patients with a higher positive to negative symptom ratio had greater specificity; chronic schizophrenia participants had greater sensitivity than first episode participants and people on higher doses of antipsychotics had greater specificity than those on low doses. It may be that medication altered brain structure in people on it for longer or at higher doses making it easier for the test to differentiate participants with schizophrenia from healthy controls (Ho et al, 2011). Year of publication was not associated with significantly different performance.

There was overlap in confidence intervals of sensitivity and specificity but resting state functional MRI was superior to structural MRI in terms of sensitivity. Visual inspection of Forrest plots suggested publication bias was unlikely.

Conclusion

As an overall conclusion

  • 80% of participants with schizophrenia and 80% of “healthy controls” were correctly identified as such, which means:

    • 20% of participants with schizophrenia were incorrectly identified as being a “healthy control”
    • 20% of participants without schizophrenia were incorrectly classified as having schizophrenia

Strengths and limitations

  • This meta-analysis used a standardised and clear methodology to identify studies
  • The perennial bugbear of meta-analysis: were they counting apples and oranges as the same? It’s not clear how different the types of multi-variate pattern recognition types were from each other. Using the Mark 1 Eyeball and looking at individual results, there was some heterogeneity of sensitivity and specificity
  • The biggest weakness is use of different diagnostic criteria in different studies. This type of bivariate meta-analysis can only be done if the same diagnostic threshold and criteria are used in all studies (Takwoingi et al, 2015). Although there is likely to be significant overlap between patients diagnosed as having schizophrenia by related criteria such as DSM II, DSM IIR and DSM- IV, the amount of heterogeneity still makes a combined meta-analysis of diagnostic tests using different diagnostic criteria difficult to perform. It would be better to either use only the studies with a particular diagnostic criteria (e.g. DSM-IV) or perform separate meta-analysis for the different diagnostic criteria.

As in my last blog, I typed the figures from the study into www.clinicalutility.co.uk. Yet again it classified the usefulness of this type of biomarker as “Fair”. So even with clever algorithms and fancy dan brain imaging, I would suggest that this type of test is not ready for routine use. The lack of association of different performance with year of publication suggests that we may not see any improvement in performance of this technique in the future.

It looks like I don’t get to be a “proper doctor” and order diagnostic imaging scans for schizophrenia just yet.

Links

Primary paper

Kambeitz J, Kambeitz-Ilankovic L, Leucht S, Wood S, Davatzikos C, Malchow B, Falkai P, Koutsouleris N. Detecting neuroimaging biomarkers for schizophrenia: a meta-analysis of multivariate pattern recognition studies (PDF). Neuropsychopharmacology. 2015 Jun;40(7):1742-51. doi: 10.1038/npp.2015.22. Epub 2015 Jan 20.

Other references

Ho BC, Andreasen NC, Ziebell S, Pierson R, Magnotta V. Long-term antipsychotic treatment and brain volumes: a longitudinal study of first-episode schizophrenia. Arch Gen Psychiatry. 2011 Feb;68(2):128-37. doi: 10.1001/archgenpsychiatry.2010.199.

Sarpal DK, Argyelan M, Robinson DG, Szeszko PR, Karlsgodt KH, John M, Weissman N, Gallego JA, Kane JM, Lencz T, Malhotra AK. (2015) Baseline Striatal Functional Connectivity as a Predictor of Response to Antipsychotic Drug Treatment. Am J Psychiatry. 2015 Aug 28:appiajp201514121571. [Epub ahead of print]

Takwoingi Y, Riley RD, Deeks J. (2015) Meta-analysis of diagnostic accuracy studies in mental healthEvid Based Mental Health ebmental-2015-102228 Published Online First: 7 October 2015 doi:10.1136/eb-2015-102228

Photo credits

 

Share on Facebook Tweet this on Twitter Share on LinkedIn Share on Google+
Mark as read
Create a personal elf note about this blog
Profile photo of Samei Huda

Samei Huda

Dr Samei Huda has been working as a Consultant Psychiatrist for over 20years including a year in Australia. He also completed a Diploma in Cognitive Therapy (though he doesn’t see patients for therapy) and a Master of Science in Psychiatry. He has been working with an Early Intervention in Psychosis team since 2007. Over his career he has directly assessed, managed and treated hundreds of patients with psychosis of different types. He recommends people with an interest in psychosis to read www.EASEnet.dk (he is not affiliated with them). He has written a book evaluating the evidence for diagnosis and treatment in psychiatry called “The Medical Model in Mental Health” published by Oxford University Press. He has recently completed an MA in Philosophy of Mental Health from UCLAN.

More posts - Website

Follow me here –