The use of machine learning techniques in trauma-related disorders: a systematic review
Introduction
Trauma-related disorders such as Posttraumatic stress disorder (PTSD) and acute stress disorder (ASD) are considered to be debilitating conditions, developed from exposure to traumatic events including war, mass violence, natural disasters, and accidents. The DSM-5 (American Psychiatric Association, 2013) lists 20 diagnostic criteria for PTSD divided into four clusters of symptoms: re-experience of the traumatic event; avoidance; persistent negative thoughts or feelings; trauma-related arousal and reactivity. The WHO World Mental Health Survey conducted across 24 countries found a lifetime prevalence of any traumatic event of 70.4% (Benjet et al., 2016), suggesting that constitutional and sociocultural factors are also involved in the development of the disorder, besides the magnitude of trauma (Yehuda, 2004). The prevalence of PTSD in a lifetime is 11% for women and 5.5% for men (Kessler et al., 1995). It is postulated that a dose–response relationship exists between exposure to traumatic events and the subsequent development of PTSD, indicating that prior trauma and/or multiple traumatic event exposures increase the risk of the disorder (Ozer et al., 2008; Kilpatrick et al., 2013).
Establishing the diagnosis of PTSD and ASD has always been a challenge in clinical practice, as well as in academic research. As indicated by its numerous risk factors, the etiologies of trauma disorders are multicausal and complex. In addition, the development of diagnostic criteria for classification systems (such as DSM-5) has been elaborated from research with chronic populations and in tertiary care settings; such phenotypic expressions may not reflect the instability and nonspecific nature of the phenomenology of the disorder in its development (McGorry et al., 2006). Evidence-based, trauma-focused therapies with the most support are cognitive- and exposure-based approaches, with prolonged exposure and cognitive processing therapy being the most investigated (Charney et al., 2018). Notwithstanding, establishing first-line psychotherapies may be difficult because of—among other aspects—the burden to patients and patient profiles (Nash and Watson, 2012). Some statistically significant results provided by evidence-based medicine may not represent a real benefit for an individual patient; subjects in clinical trials do not always reflect the multimorbidity profile of “real-life” patients (Greenhalgh et al., 2014). This may be particularly true in the field of PTSD, where clinical heterogeneity can be a very important factor, not always taken into account in research.
Machine learning, a field of computer science and a part of artificial intelligence, refers to the science and engineering by which machines (i.e., computer systems) can analyze and acquire information from data (Liu and Salinas, 2017). Machine learning can help develop sophisticated data models using advanced mathematical techniques and handling complex data sets with heterogeneous distribution. The ‘learning’ method is usually made by a supervised or an unsupervised approach (Bishop, 2006). In supervised learning, the user feeds the machine with input data and expected outcome: the machine learns a mapping from the input to the outcome target, through classification (where the output variable is a category, such as ‘disease’ or ‘no disease’) or regression (where the output variable is a numeric variable) methods. Common examples of supervised learning algorithms are Logistic Regression, Support Vector Machines and Neural Networks. Supervised learning is often used to estimate prediction and risk: the Framingham Risk Score for coronary heart disease may be one of the most famous uses of supervised learning in medicine (Deo, 2015; Kannel et al., 1975). Unsupervised learning does not depend on previous associations and output variables: the goal is to model the underlying data structure to learn more about the data. It can be performed by discovering groups of similar cases (clustering) or determining the distribution of available data (density estimation). Network analysis allows visualization of the connectivity among symptoms and clusters of symptoms providing knowledge about the strength and quantity of relationships (Sullivan et al., 2018), taking into account regression and clustering techniques. A revision of the relevant principles of machine learning and its limitations can be found elsewhere (Schultebraucks and Galatzer-Levy, 2019; Librenza-Garcia et al., 2017; Deo, 2015).
Machine learning techniques can be applied to improve classification of disorders, to predict risk factors and treatment outcomes, and to improve person-specific treatment selection (Hahn et al., 2017). Since PTSD and ASD are disorders that present clinical and biological heterogeneity, which may constitute a barrier to understanding the causative mechanisms and to developing optimal treatments and diagnostic tools, machine learning is a suitable approach to better achieve this understanding. The present study aims to systematically review data in which PTSD and ASD were assessed through machine learning techniques regarding classification, prognostic, and treatment selection studies. Furthermore, we proposed a method of quality measurement of these studies.
Section snippets
Methods
This study followed the Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) reporting guidelines (Liberati et al., 2009) and is registered on the International Prospective Register of Systematic Reviews (PROSPERO identifier CRD42019115850). We searched PubMed, Embase, and Web of Science for articles published between January 1960 and May 2019 using terms associating machine learning techniques with PTSD and ASD. The complete filter is available in the supplementary
Results
We found a total of 806 potential abstracts and included 49 articles in the present review. Fig. 1 shows the study selection process. A list of the included articles as well as the most relevant characteristics and findings are presented in Table 1 (Prognostic studies), Table 2 (Classification studies), and Table 3 (Network analysis and unsupervised studies).
Thirty-three articles assessed prognosis, most in order to predict risk factors related to the development of PTSD or to identify its
Discussion
We evaluated 49 articles that used machine learning techniques to assess trauma-related disorders, including ASD and PTSD. We identified several studies designed to aid in the prediction and diagnostic classification of PTSD. No study aimed at the treatment of the disorder was found.
One of the most immediate uses of robust techniques such as machine learning in psychiatry is developing predictive models of mental health disorders, especially from risk factors. Thirty-three of the 49 articles in
Declaration of competing interest
The authors declare the following financial interests/personal relationships which may be considered as potential competing interests: ICP is supported by CAPES – Brazilian Federal Agency for Support and Evaluation of Graduate Education and by SENAD/Brazil – National Secretariat for Drug Policies. The other authors declare that they have no conflicts of interest.
Acknowledgements
This study was financed in part by the Brazilian Federal Agency for Support and Evaluation of Graduate Education (CAPES, Brazil, Finance Code 001) and by the Conselho Nacional de Pesquisa (CNPq, Brazil).
References (81)
- et al.
Psychometric properties of the PTSD checklist (PCL)
Behav. Res. Ther.
(1996) - et al.
Leveraging big data to model the likelihood of developing psychological conditions after a concussion
- et al.
Quantitative forecasting of PTSD from early trauma responses: a Machine Learning application
J. Psychiatr. Res.
(2014) - et al.
Early identification of posttraumatic stress following military deployment: application of machine learning methods to a prospective study of Danish soldiers
J. Affect. Disord.
(2015) - et al.
The PRISMA statement for reporting systematic reviews and meta-analyses of studies that evaluate health care interventions: explanation and elaboration
- et al.
The impact of machine learning techniques in the study of bipolar disorder: a systematic review
Neurosci. Biobehav. Rev.
(2017) - et al.
Ensemble machine learning prediction of posttraumatic stress disorder screening status after emergency room hospitalization
J. Anxiety Disord.
(2018) - et al.
Identifying a clinical signature of suicidality among patients with mood disorders: a pilot study using a machine learning approach
J. Affect. Disord.
(2016) - et al.
Predicting posttraumatic stress disorder following a natural disaster
J. Psychiatr. Res.
(2018) - et al.
Blood-based gene-expression biomarkers of post-traumatic stress disorder among deployed marines: a pilot study
Psychoneuroendocrinology
(2015)