Matlab scripts for DELCODE project 162's support vector classifications and statistical comparisons of their performance. The scripts rely on Joram Soch's ML4ML toolbox. The results of the study will be published under the title
Konrad F. Waschkies (1,2), Joram Soch (1,3), Margarita Darna (4), Anni Richter (4), DELCODE study group, Jens Wiltfang (2), Björn H. Schott (1,2,4), & Jasmin M. Kizilirmak (1,5)
- Cognitive Geriatric Psychiatry, German Center for Neurodegenerative Diseases, Göttingen, Germany
- Department of Psychiatry and Psychotherapy, University Medical Center Göttingen, Göttingen, Germany
- Bernstein Center for Computational Neuroscience, Berlin, Germany
- Leibniz Institute for Neurobiology, Magdeburg, Germany
- Neurodidactics and NeuroLab, Institute for Psychology, University of Hildesheim, Hildesheim, Germany
Background: Alzheimer’s disease (AD) is the most common cause of dementia and represents a serious health issue, especially in aging societies. Over the past two decades, machine learning approaches to classify people into healthy, increased AD risk, and AD have gained popularity. Their main goal is the identification of valuable predictors for valid classification, prediction of conversion, and automatization of the process. While biomarkers from cerebrospinal fluid (CSF) are the best-established predictors, other less invasive candidate predictors have been identified that show considerable association with AD and increased risk for developing said disease. Candidate predictors include imaging markers like resting-state fMRI, and self-report measures like personality traits or affective symptoms such as anxiety and depression.
Methods: Here, we evaluated the predictive value of such less invasive, easy-to-assess predictors separately and in different combinations for classification of healthy controls (HC, N = 189), subjective cognitive decline (SCD, N = 338), mild cognitive impairment (MCI, N = 132), and mild AD (N = 74) in multi-class support vector machine (SVM) classification. Participants were recruited from the multi-center DZNE-Longitudinal Cognitive Impairment and Dementia Study (DELCODE). In a subset of the sample, we also assessed the predictive performance of CSF markers. The different predictors and predictor combinations were further compared regarding their predictive performance.
Results: We found that HC were best predicted by a feature set comprised of personality, anxiety, and depression scores, while participants with AD were best predicted by a feature set containing CSF markers. Both feature sets had equally high overall decoding accuracy. However, all assessed feature sets performed relatively poorly in the classification of SCD and MCI.
Conclusion: Our results highlight that SCD and MCI are heterogeneous groups, pointing out the importance of optimizing their diagnosis criteria. Moreover, CSF biomarkers, personality, depression, and anxiety indicate complementary value for class prediction, which should be followed up on in future studies and extended by assessing the predictive value of the latter three regarding conversion rates.
Keywords: Alzheimer’s disease, subjective cognitive decline, mild cognitive impairment, biomarker, cerebrospinal fluid, personality, fMRI, resting-state, support vector machine, machine learning
The manuscript will be published as a preprint. URL follows.