Review articleApplications of machine learning algorithms to predict therapeutic outcomes in depression: A meta-analysis and systematic review
Introduction
The functional and psychosocial deficits associated with depression are pervasive, often chronic, progressive, and highly disabling (Vigo et al., 2016). The Sequenced Treatment Alternatives to Relieve Depression (STAR*D) trial reported that patients who remit with the first trial of an antidepressant experience significant reductions in work-related disability; in contrast, patients who remit with subsequent treatment trials or strategies exhibit residual functional impairments (Trivedi et al., 2013). Notwithstanding the opportunity afforded by intervening earlier along the disease trajectory, most individuals with mood disorders do not receive timely and accurate diagnoses (Trivedi et al., 2013). Moreover, despite the availability (and, in many cases, accessibility) of effective and multi-modal treatments for mood disorders (e.g., pharmacotherapy, psychotherapy, neuromodulation), current treatment selection and sequencing methods are largely a process of trial-and-error, further delaying effective treatment (Trivedi et al., 2006).
Precision-based approaches that integrate quantifiable neurobiological substrates of psychopathology have been proposed to inform treatment selection and refine disease models in psychiatric disorders (Insel and Cuthbert, 2015). The National Institute of Mental Health’s Research Domain Criteria (RDoC) proffers a dimensional approach to parse clinical heterogeneity across current diagnostic classifications into homogenous, data-driven categories that are more proximate to the etiology of psychiatric disorders. For example, a recent study characterized behavioural phenotypes along a continuum–ranging from functionally adaptive to maladaptive groups–informed by a data-driven analysis of behavioural and psychiatric assessments and resting-state functional magnetic resonance imaging (rs-fMRI) measures in a community-based sample (Van Dam et al., 2017). In addition to examining evidence from populations across symptom-based diagnostic categories, there is a need to integrate findings using a ”systems biology” approach across units of analyses (e.g., genetic, molecular, cellular, circuits, behaviour) and data sources (e.g., pre-clinical, clinical, epidemiological) (Barabási, 2007, McIntosh, 2013).
Recent advances in information processing and computational resources present a novel opportunity to conceptualize and analyze human behaviour (Huys et al., 2016). While conventional statistical methods sequentially analyze the relationship between variables, machine learning approaches can iteratively and contemporaneously analyze multiple interacting associations between variables and/or variable sets (Orrù et al., 2012). Many machine learning methods are capable of identifying which variables in a given dataset are relevant or irrelevant to an outcome of interest, whereas conventional statistical methods rely on investigator input to specify variables relevant to a particular analysis (Bzdok and Meyer-Lindenberg, 2017). Thus, machine learning approaches are capable of agnostically devising data-driven models and algorithms with predictive capability. Moreover, the precision and accuracy of the models’ predictions (i.e., classifications) can be quantified and compared across models and/or iterations within-sample (i.e., cross-validation), as well as with external validation datasets, for greater out-of-sample generalizability (Goodfellow et al., 2016). Machine learning methods are additionally capable of integrating multiple data types (e.g., neuroimaging, behavioural) from multiple sources while correcting for heterogeneity within pooled datasets (Bzdok, Meyer-Lindenberg, 2017, Friston, Redish, Gordon, 2017, Rohart, Gautier, Singh, Lê Cao, 2017).
Toward the aim of improving treatment outcomes in mood disorders, recent studies have attempted to identify replicable and quantifiable predictors of therapeutic outcomes. Attempts to characterize biological and phenomenological characteristics capable of predicting illness trajectory and therapeutic recovery have shown promise; however, no single variable or modality has emerged as a robust predictor of therapeutic outcome (Rosenblat, Lee, McIntyre, 2017, Trivedi, Morris, Grannemann, Mahadi, 2005). The successful application of machine learning algorithms in consumer data analytics and other areas of medicine (e.g., oncology) demonstrates its predictive capability and the potential for machine learning approaches to inform predictive models in mood disorders (Huys, Maia, Frank, 2016, Rakesh, 2017, Ray, Britschgi, Herbert, Takeda-Uchimura, Boxer, Blennow, Friedman, Galasko, Jutel, Karydas, Kaye, Leszek, Miller, Minthon, Quinn, Rabinovici, Robinson, Sabbagh, So, Sparks, Tabaton, Tinklenberg, Yesavage, Tibshirani, Wyss-Coray, 2007, Rumshisky, Ghassemi, Naumann, Szolovits, Castro, McCoy, Perlis, 2016).
Notwithstanding the growing interest in, and opportunities afforded by, the application of machine learning algorithms in mechanistic and therapeutic research, no previous study has comprehensively reviewed the extant literature reporting on the use and application of machine learning algorithms to predict therapeutic outcomes in populations with depression. Herein, we aim to explore how, and to what extent, previous studies have applied machine learning algorithms to inform treatment selection and personalize therapy in clinical populations with depression.
Section snippets
Methods
We conducted a meta-analytic and scoping systematic review informed by recommendations from the Cochrane Handbook for Systematic Reviews of Interventions, the Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) statement, and the Joanna Briggs Institute’s Manual for Scoping Reviews (Cochrane Collaboration, Moher, Liberati, Tetzlaff, Altman, PRISMA Group, 2009, Peters, Godfrey, Khalil, McInerney, Parker, Soares, 2015). We evaluate extant studies that have applied machine
Results
The initial search yielded a total of 639 unique records (Table S1). After title and abstract review, 564 records were excluded; 75 full-text publications were assessed for eligibility and after full-text review, 44 records were excluded (Fig. 1). Twenty-seven studies investigated predictors of a therapeutic outcome in depression but did not apply a machine learning algorithm. Six studies applied a machine learning algorithm to differentiate depressed subjects from controls but did not report
Evidence source
Approximately 96% of studies included in the qualitative review were published in peer-reviewed medical journals, with a recent trend towards higher impact factor journals. Results from Egger’s test and funnel plot asymmetry prior to trim and fill additionally support the presence of publication bias. Although our search included grey literature sources, studies that applied machine learning algorithms with negative results may have been less likely to be published, abstracted, or made
Conclusions and future directions
We conducted a comprehensive systematic review and meta-analysis surveying the use of machine learning algorithms to inform predictive models in mood disorders. Classification algorithms were able to predict therapeutic outcomes among subjects of previously published prospective interventional trials () with an overall accuracy of 0.82. Pooled estimates of classification accuracy were significantly different between models informed by a single data type (i.e., neuroimaging,
Disclosures
All authors contributed to the development of the research hypothesis and study design. YL conducted the search, data extraction, and data analysis and wrote the first draft of the manuscript. All authors contributed to the interpretation of results and manuscript writing. This work did not receive any specific grant from funding agencies in the public, commercial, or not-for-profit sectors. The authors have no conflicts of interest to declare.
References (72)
- et al.
Predicting tDCS treatment outcomes of patients with major depressive disorder using automated EEG classification
J. Affect. Disord.
(2017) - et al.
Responders to rTMS for depression show increased fronto-midline theta and theta connectivity compared to non-responders
Brain Stimul.
(2018) - et al.
Predisposition to treatment response in major depressive episode: a peripheral blood gene coexpression network analysis
J. Psychiatr. Res.
(2016) - et al.
Cross-trial prediction of treatment outcome in depression: a machine learning approach
Lancet Psychiatry
(2016) - et al.
The binomial distribution of meta-analysis was preferred to model within-study variability
J. Clin. Epidemiol.
(2008) - et al.
Combining clinical variables to optimize prediction of antidepressant treatment outcomes
J. Psychiatr. Res.
(2016) - et al.
The combined effect of genetic polymorphisms and clinical parameters on treatment outcome in treatment-resistant depression
Eur. Neuropsychopharmacol.
(2015) - et al.
A machine learning approach using EEG data to predict response to SSRI treatment for major depressive disorder
Clin. Neurophysiol.
(2013) - et al.
Magnetic resonance imaging measures of brain structure to predict antidepressant treatment outcome in major depressive disorder
EBioMedicine
(2015) Toward a systems biology of mood disorder
Biol. Psychiatry
(2013)
Clustering of depression and inflammation in adolescents previously exposed to childhood adversity
Biol. Psychiatry
Preferred reporting items for systematic reviews and meta-analyses: the PRISMA statement
J. Clin. Epidemiol.
Using support vector machine to identify imaging biomarkers of neurological and psychiatric disease: a critical review
Neurosci. Biobehav. Rev.
A neural network model for combining clinical predictors of antidepressant response in mood disorders
J. Affect. Disord.
Inflammation and clinical response to treatment in depression: a meta-analysis
Eur. Neuropsychopharmacol.
Data-driven phenotypic categorization for neurobiological analyses: beyond DSM-5 labels
Biol. Psychiatry
Estimating the true global burden of mental illness
Lancet Psychiatry
Network medicine–from obesity to the “diseasome”
N. Engl. J. Med.
Operating characteristics of a rank correlation test for publication bias
Biometrics
Machine learning for precision psychiatry: opportunites and challenges
Biol. Psychiatry
Reevaluating the efficacy and predictability of antidepressant treatments: a symptom clustering approach
JAMA Psychiatry
The use of confidence or fiducial limits illustrated in the case of the binomial
Biometrika
Neural correlates of sad faces predict clinical remission to cognitive behavioural therapy in depression
Neuroreport
Resting-state connectivity biomarkers define neurophysiological subtypes of depression
Nat. Med.
A nonparametric “trim and fill” method of accounting for publication bias in meta-analysis
J. Am. Stat. Assoc.
Trim and fill: a simple funnel-Plot–Based method of testing and adjusting for publication bias in meta-analysis
Biometrics
Bias in meta-analysis detected by a simple, graphical test
BMJ
A cognitive-emotional biomarker for predicting remission with antidepressant medications: a report from the iSPOT-D trial
Neuropsychopharmacology
Transformations related to the angular and the square root
Ann. Math. Stat.
Computational nosology and precision psychiatry
Comput. Psychiatry
Peripheral biomarkers of major depression and antidepressant treatment response: current knowledge and future outlooks
J. Affect. Disord.
Deep Learning
Testing the predictive value of peripheral gene expression for nonremission following citalopram treatment for major depression
Neuropsychopharmacology
Computational psychiatry as a bridge from neuroscience to clinical applications
Nat. Neurosci.
Medicine. brain disorders? precisely
Science
Cited by (167)
Prediction of non-suicidal self-injury in adolescents at the family level using regression methods and machine learning
2024, Journal of Affective DisordersPrognostic models for predicting insomnia treatment outcomes: A systematic review
2024, Journal of Psychiatric ResearchPerformance of advanced machine learning algorithms overlogistic regression in predicting hospital readmissions: A meta-analysis
2023, Exploratory Research in Clinical and Social Pharmacy
- 1
University Health Network, 399 Bathurst Street, MP 9–325, Toronto, ON M5T 2S8 Canada.