Data used in AI to diagnose depression and predict course of illness

AI has the technical ability to automatically analyze and learn from previous data to the prediction of naturalistic courses of depression based on psychological, biological, and clinical data.3 For the AI system, common biological data will include genetic basis of mood disorders, EEG, tissue biomarkers (e.g., tumour necrosis factor-alpha, interleukin-6, interleukin-1β, C-reactive protein, brain-derived neurotrophic factor, BDNF)6, genetic (e.g., single nucleotide polymorphisms of BDNF, serotonin-2A-receptor, and serine/threonine-protein phosphatase genes)7, or combined (e.g., neuroimaging and phenomenological; genetic).8 Metabolic syndrome variables included triglyceride level, high-density lipoprotein cholesterol level, systolic and diastolic blood pressure, and fasting glucose level.9

AI utilizes neuroimaging techniques to classify and predict likely treatment outcomes of major depressive disorders. Previous research found that changes of gray matter volume are able to predict response of ECT.5 Functional magnetic resonance imaging (fMRI) studies suggest that the pattern of brain activation to neutral faces in depressed patients is not consistent with the pattern observed in healthy people to the same stimuli.10 Novel neuroimaging techniques include functional near infrared spectroscopy (fNIRS)11 or optical topography12. For fNIRS, the paradigm is verbal fluency test.

Phenomenological data include psychometric, neurocognitive, sociodemographic, psychiatric history, depression rating scale13. Chekroud et al (2016) identified most important psychological predictors of treatment outcome of depression for machine learning.14 These predictors include Initial symptom severity, current employment, psychomotor agitation, energy level, ethnicity, years of education, loss of insight, anxiety, impact of traumatic event, delayed insomnia, avoidance of reminders of traumatic event, somatic complaints, suicidal ideation, depressed mood most of the day, panic attacks and number of previous depressive episodes.14

Overview of AI applications and research in depression

As the medical literature on the applications of AI in depressive disorder has rapidly expanded in recent years, a research group headed by Professor Bach Tran from the Hanoi Medical University applied bibliometric analysis, which objectively evaluates the productivity of global researchers or institutions in this field of AI and depression.3

The rise of AI could assist physicians and psychiatrists in shortening processing times and improving the quality of care in clinical practice. The potential of AI in medicine has attracted more attention from researchers, which makes the year 2018 an important year for AI research.3 In terms of global network of AI Research in depression, the US ranked first at the number of publication and collaboration networks. In Europe, countries like the United Kingdom, Germany, France, Belgium, and Denmark, which possess the most advanced digital technologies in the Europa, had strongly collaborated to promote research in AI for managing depression.3 In Asia, Japan and South Korea have also joined this research field by collaborating with highly developed Western AI researchers. India and South-East Asian countries, including Singapore, Malaysia, and Indonesia were highly connected in AI research.

For the current research interests on AI application in depression, the most well-studied application of AI in depression appeared to be the utilization of machine learning to predict clinical characteristics, which accounted for more than 60% of total cases.3 Researchers have also paid attention to a wide range of issues and fields where AI can foster innovation, such as diagnosis accuracy, structural imaging techniques, negative symptoms of certain disorders and diseases, gene testing, and drug development. AI assisted diagnosis by neuroimaging and electroencephalography (EEG)-based diagnosis were by far more researched than other areas.3 There are AI models developed for specific types of depression. Wang et al used electronic health records and machine learning to predict postpartum depression.15 Predictors include race, obesity, anxiety, depression, different types of pain, antidepressants, and anti-inflammatory drugs during pregnancy. The best prediction performance achieved was 79%.15

A meta-analysis was conducted to explore how, and to what extent, previous studies have applied machine learning algorithms to inform treatment selection and personalize therapy in patients suffering from depression.8 This meta-analysis found that pooled estimates of classification accuracy were significantly higher in predictive models integrating multiple data types (pooled accuracy = 93%) when compared to models with single data type (pooled accuracy was between 68% to 85%).

Specific machine algorithms and examples

Lee et al (2018) reported the following methods of AI in depression. Most AI systems trained a machine learning algorithm using a labelled dataset (i.e., by therapeutic outcome) to iteratively evaluate, compare, and select variables that would differentiate treatment responders vs. non responders with the highest accuracy in an unlabelled (i.e., validation) dataset.8 Linear algorithms can avoid complexity and overfitting.16 Examples of commonly applied linear algorithms in the identified literature are linear kernel-based support vector machines (SVM), L1-regularized logistic regression, logistic regression with elastic net regularization, and linear artificial neural networks.8 SVM is the most popular method to classify depression because of its strengths on including a reliable theoretical foundation in real world data16 and provides flexible response to high‐dimensional data.5 For the Kernel-based SVM, the general task of pattern analysis is to find and study general types of relationships between components in datasets. Nonlinear algorithms are more flexible than linear algorithms.16 Examples of commonly used non-linear algorithms include, alternating or hierarchical multi-label decision trees, and multilayer perceptron artificial neural networks.8

Logistic regression is a regression model with a binary dependent variable.15 L1-regularized logistic regression tunes and generalizes the model in order to balance the bias variance tradeoff.17 Logistic regression was used to detect depression based on patients’ speech.18 For neuroimaging data, the Gaussian Process Classifier (GPC), a machine learning approach that assigns a predictive probability to an individual pattern of brain activation based on the confidence of a classifier computed from pre-processed fMRI scans.10 If the classifiers’ predictive probability to a stimuli would be below a validated threshold, there would be evidence that the patient’s pattern of brain activation was different from the healthy controls pattern which would indicate that the patient suffers from depression.10 Principal Component Analysis (PCA) is used to reduce a larger set of variables to a small set that still contains most of the information in the larger set. Partial least squares (PLS) regression reduces the predictors to a smaller set of uncorrelated components and performs least squares regression on these components. PCA and partial least squares discriminant analysis (PLS-DA) were used to analyze the metabolic profiles of depressed patients and healthy people.19

Validation measures are used to assess how well a model developed by a learning method will perform on new unseen data.16 One method of cross validation is 10-fold cross validation. In 10-fold cross-validation, the original sample is randomly partitioned into 10 equal sized subsamples. It was found that 10‐fold CV provides a more stable performance across different data.5 Most AI system adopts a binary classification of clinical response was adopted by 92% of identified studies applying to define therapeutic outcomes using a pre-specified cut-off value (e.g., 50% reduction in or <10 endpoint HAMD total score) in a supervised learning algorithm.8

Limitations of AI

There are several limitations. Firstly, machine learning algorithms are limited by the type of data captured, the quality of available data, the conceptual frameworks algorithms are applied to, and the underlying assumptions.8 Secondly, some researchers proposed to analyse negative events, negative emotions, symptoms, and negative thoughts posted by patients with depressions on the web. Such method without biological data had lower accuracy.20 The concepts of privacy and confidentiality were under-researched, which indicates a lack of attention on these particular issues in AI and depression research.3 Thirdly, some studies were not able to replicate their findings in an independent dataset due to different characteristics of data from different samples. Fourth, this paper mainly discusses the application of AI in diagnosing depression and predicting the course of illness. There are other applications such as psychological AI which offers psychotherapy but not covered here. Nevertheless, patients mentioned that there is a lack of personal touch by psychological AI.21 Fifth, the applications mentioned in this paper mainly applied to adults with depression although there are preliminary study to apply AI and wearable devices to children with depression.22

Future directions of AI

The research gaps in the application of AI is to develop diagnostic algorithm which can handle the complexity and diversity of clinical data and establish reasoning models for specific clinical tasks. There are several research directions for AI. Firstly, developing countries, especially those from the south-east Asian region, could seek investment from developed countries in the field of information technology.3 Secondly, the next focus of development will include artificial neural network, genetic algorithms, and natural language processing.3 Third, new policy and research on the use of AI in management of depression should be made to achieve a balance between the beneficial uses of clinical data and personal privacy.3 Fourth, future studies will try to leverage multi-site dataset to minimize missing and erroneous data points.

Conclusion

The last five years have witnessed the exponential growth in applications of AI in depression research. Developed countries, such as the US, England, Germany, and Japan, have made numerous attempts to innovate the current AI system and conduct various research studies to confirm its effectiveness.3 In the future decade, the research objectives will involve the integration of various clinical data, neuroimaging, and biochemical parameters and genotypes from different centres into the AI system, which helps diagnose depression and predict treatment outcomes. With around half of patients with depression experiencing treatment response to first-line antidepressants23, machine-learning algorithm have the potential to significantly reduce the duration of patient suffering.

 References

1. Lim, G. Y.; Tam, W. W.; 

Lu, Y.; Ho, C. S.; Zhang, M. W.; Ho, R. C., Prevalence of Depression in the Community from 30 Countries between 1994 and 2014. Scientific reports 2018, 8 (1), 2861-2861.

2. Ho, R. C. M.; Mak, K.-K.; Chua, A. N. C.; Ho, C. S. H.; Mak, A., The effect of severity of depressive disorder on economic burden in a university hospital in Singapore. Expert review of pharmacoeconomics & outcomes research 2013, 13 (4), 549-559.

3. Tran, B. X.; 

McIntyre, R. S.; Latkin, C. A.; Phan, H. T.; Vu, G. T.; 

Nguyen, H. L. T.; Gwee, K. K.; Ho, C. S. H.; Ho, R. C. M., The Current Research Landscape on the Artificial Intelligence Application in the Management of Depressive Disorders: A Bibliometric Analysis. International journal of environmental research and public health 2019, 16 (12), 2150.

4. Tran, B. X.; Vu, G. T.; Ha, G. H.; Vuong, Q.-H.; 

Ho, M.-T.; Vuong, T.-T.; La, V.-P.; 

Ho, M.-T.; Nghiem, K.-C. P.; Nguyen, H. L. T.; Latkin, C. A.; Tam, W. W. S.; Cheung, N.-M.; Nguyen, H.-K. T.; Ho, C. S. H.; Ho, R. C. M., Global Evolution of Research in Artificial Intelligence in Health and Medicine: A Bibliometric Study. J Clin Med 2019, 8 (3), 360.

5. Gao, S.; Calhoun, V. D.; Sui, J., Machine learning in major depression: From classification to treatment outcome prediction. CNS Neurosci Ther 2018, 24 (11), 1037-1052.

6. Liu, Y.; Ho, R. C.-M.; Mak, A., Interleukin (IL)-6, tumour necrosis factor alpha (TNF-α) and soluble interleukin-2 receptors (sIL-2R) are elevated in patients with major depressive disorder: a meta-analysis and meta-regression. Journal of affective disorders 2012, 139 (3), 230-239.

7. Mak, K. K.; Kong, W. Y.; Mak, A.; Sharma, V. K.; Ho, R. C. M., Polymorphisms of the serotonin transporter gene and post-stroke depression: a meta-analysis. Journal of neurology, neurosurgery, and psychiatry 2013, 84 (3), 322-328.

8. Lee, Y.; Ragguett, R.-M.; Mansur, R. B.; Boutilier, J. J.; Rosenblat, J. D.; Trevizol, A.; 

Brietzke, E.; Lin, K.; Pan, Z.; 

Subramaniapillai, M.; Chan, T. C. Y.; Fus, D.; Park, C.; 

Musial, N.; Zuckerman, H.; Chen, V. C.-H.; Ho, R.; 

Rong, C.; McIntyre, R. S., Applications of machine learning algorithms to predict therapeutic outcomes in depression: A meta-analysis and systematic review. Journal of affective disorders 2018, 241, 519-532.

9. Dinga, R.; Marquand, A. F.; Veltman, D. J.; Beekman, A. T. F.; Schoevers, R. A.; van Hemert, A. M.; Penninx, B. W. J. H.; Schmaal, L., Predicting the naturalistic course of depression from a wide range of clinical, psychological, and biological data: a machine learning approach. Transl Psychiatry 2018, 8 (1), 241-241.

10. Oliveira, L.; 

Ladouceur, C. D.; Phillips, M. L.; Brammer, M.; Mourao-Miranda, J., What does brain response to neutral faces tell us about major depression? evidence from machine learning and fMRI. PloS one 2013, 8 (4), e60121-e60121.

11. Lai, C. Y. Y.; Ho, C. S. H.; Lim, C. R.; Ho, R. C. M., Functional near-infrared spectroscopy in psychiatry. BJPsych Advances 2017, 23 (5), 324-330.

12. Ho, C. S. H.; Zhang, M. W. B.; Ho, R. C. M., Optical Topography in Psychiatry: A Chip Off the Old Block or a New Look Beyond the Mind-Brain Frontiers? Frontiers in psychiatry 2016, 7, 74-74.

13. Ho, R. C. M.; Chua, A. C.; Tran, B. X.; Choo, C. C.; 

Husain, S. F.; Vu, G. T.; McIntyre, R. S.; Ho, C. S. H., Factors Associated with the Risk of Developing Coronary Artery Disease in Medicated Patients with Major Depressive Disorder. International journal of environmental research and public health 2018, 15 (10), 2073.

14. Chekroud, A. M.; Zotti, R. J.; Shehzad, Z.; Gueorguieva, R.; Johnson, M. K.; Trivedi, M. H.; Cannon, T. D.; Krystal, J. H.; Corlett, P. R., Cross-trial prediction of treatment outcome in depression: a machine learning approach. Lancet Psychiatry 2016, 3 (3), 243-250.

15. Wang, S.; Pathak, J.; Zhang, Y., Using Electronic Health Records and Machine Learning to Predict Postpartum Depression. Stud Health Technol Inform 2019, 264, 888-892.

16. Patel, M. J.; Khalaf, A.; Aizenstein, H. J., Studying depression using imaging and machine learning methods. Neuroimage Clin 2015, 10, 115-123.

17. Zhang, X.; Hu, B.; Ma, X.; Xu, L., Resting-state whole-brain functional connectivity networks for MCI classification using L2-regularized logistic regression. IEEE Trans Nanobioscience 2015, 14 (2), 237-247.

18. Jiang, H.; Hu, B.; Liu, Z.; 

Wang, G.; Zhang, L.; Li, X.; Kang, H., Detecting Depression Using an Ensemble Logistic Regression Model Based on Multiple Speech Features. Comput Math Methods Med 2018, 2018, 6508319-6508319.

19. Zhang, F.; Wu, C.; Jia, C.; 

Gao, K.; Wang, J.; Zhao, H.; 

Wang, W.; Chen, J., Artificial intelligence based discovery of the association between depression and chronic fatigue syndrome. Journal of affective disorders 2019, 250, 380-390.

20. Tung, C.; Lu, W., Analyzing depression tendency of web posts using an event-driven depression tendency warning model. Artif Intell Med 2016, 66, 53-62.

21. Fulmer, R.; Joerin, A.; Gentile, B.; Lakerink, L.; Rauws, M., Using Psychological Artificial Intelligence (Tess) to Relieve Symptoms of Depression and Anxiety: Randomized Controlled Trial. JMIR Ment Health 2018, 5 (4), e64-e64.

22. McGinnis, R. S.; 

McGinnis, E. W.; Hruschak, J.; Lopez-Duran, N. L.; Fitzgerald, K.; Rosenblum, K. L.; Muzik, M., Rapid Anxiety and Depression Diagnosis in Young Children Enabled by Wearable Sensors and Machine Learning. Conf Proc IEEE Eng Med Biol Soc 2018, 2018, 3983-3986.

23. Tran, B. X.; Ha, G. H.; Vu, G. T.; 

Nguyen, L. H.; Latkin, C. A.; Nathan, K.; McIntyre, R. S.; Ho, C. S.; 

Tam, W. W.; Ho, R. C., Indices of Change, Expectations, and Popularity of Biological Treatments for Major Depressive Disorder between 1988 and 2017: A Scientometric Analysis. International journal of environmental research and public health 2019, 16 (13), 2255.