Functional brain-imaging methods provide rich datasets that can be exploited by machine-learning techniques to help assess psychiatric disorders. A recent study uses this approach to identify patients with suicidal thoughts, and to distinguish those who have attempted suicide from those who have not.
Psychiatric disorders, in spite of extraordinary research over many years, still remain difficult to diagnose and treat. Many investigators have pointed out that a major problem affecting diagnosis is its dependence on behavioural symptoms rather than tests that focus on biological mechanisms1. It has been argued that because of gene–environment interactions, genetic analysis has not developed a convincing pathway towards understanding the underlying mechanisms1. Rather, brain imaging seems to be emerging as perhaps a more likely route that will enable diagnosis and assessment of treatment to be based on neurobiological mechanisms. In particular, there have been recent advances in applying machine-learning methods to brain-imaging data that have allowed researchers to decode the patterns of neural activity that represent specific thoughts of the scanned individual. Writing in Nature Human Behaviour, Just and colleagues present remarkable findings from a functional magnetic resonance imaging (fMRI) study of psychiatric patients who have suicidal thoughts2. They demonstrate that they can distinguish these patients, based on the neural representations of such thoughts, from control participants. Most impressively, Just et al. show that this approach can also classify suicidal patients who had attempted suicide from those patients who had not.
fMRI generates signals related to the amount of neural activity in a large number of brain areas with a spatial resolution of a few millimetres. When regional neural activity increases, oxygenated blood flows into the region, and fMRI can measure that change because oxygenated and deoxygenated blood cells have slightly different magnetic properties. In early fMRI studies of patients, researchers tried to find brain areas in which there was a statistically significant difference between the patients and a control group (often a group of healthy individuals). The difference could be in the amount of fMRI signal (that is, functional activation)34, or more recently, in how activity in one or more brain areas is correlated with activity in other areas (that is, functional connectivity)56. However, such differences most often were not large enough to classify individuals as being a patient or a control, and thus could not be used as a functional biomarker for diagnostic or treatment efficacy purposes. The machine-learning analysis method pioneered by Just and co-workers overcomes some of the limitations of the more traditional fMRI data analytic techniques.
In the current study, each individual was presented with three sets of ten words each related to suicide (for example, death), expressing negative effects (for example, gloom) and expressing positive effects (for example, carefree), respectively, while undergoing an fMRI scan. Six of the words and five brain locations were found to best discriminate between the suicidal patients and controls. They went on to train a machine-learning classifier to use these 30 (6 × 5) features in each participant to identify whether the participant was a patient or a control. The results were dramatic: the classifier correctly identified 15 of 17 patients as belonging to the suicide group and 16 of 17 healthy subjects as belonging to the control group. They went on to investigate just the suicidal patients, which they divided into those who had attempted suicide (9 participants) and those who had not (8 individuals). A new classifier was trained that resulted in a correct identification of 16 out of 17 patients (94% accuracy).
What information corresponding to thought processes is in the brain signals (what Just and colleagues call the neurosemantic signatures) that enables such high accuracy in patient identification? In a previous study7, the brain’s patterns of activation to specific emotions were evaluated. Thus, Just et al. were able to determine the emotional signatures of each of the six discriminating concepts (for example, death, trouble, carefree), and compare these signatures between the patients and controls, and between suicide attempters and non-attempters within the patient group. One conclusion based on their findings was that negatively valenced discriminating concepts resulted in more sadness and shame but less anger in the suicidal group than in the controls.
As emphasized by Just et al., the clinical importance of this study for psychiatry is that the focus was not on how suicidal individuals responded to a task either verbally or with button presses, but rather on how these individuals thought about the various concepts; that is, the focus was on the direct experience of the participants.
What can we learn from this study, and what are the potentially important consequences that follow? First, it is important to note that the neurosemantic signatures associated with the discriminating concepts were not localized to a single brain region, but rather were distributed in multiple brain regions (see their Fig. 1). Because of this, the novel data analysis techniques, such as the machine-learning method used by Just et al., now becoming more widely utilized are multivariate; they are not trying to localize a cognitive function to a single brain area. This has become one of the main sources of diagnostic power of functional neuroimaging: data from multiple parts of the brain are simultaneously generated, allowing investigators to employ multivariate/network analysis methods8. Second, the Just et al. study explicitly shows, in a small group of patients, that one can use the brain representations of specific thoughts to classify suicidal patients from healthy controls, and more specifically, suicide attempters from non-attempters. As the authors point out, this latter distinction is one that few risk factors can make. If subsequent studies show that (1) the particular findings in this study with its small sample size can be replicated, and (2) that the machine-learning method used by Just and colleagues can yield comparable results in other psychiatric populations, then a case can be made that functional neuroimaging has potential to become a major medical tool for diagnosis and/or evaluation of treatment efficacy of psychiatric disorders.