Background
Melancholia is frequently conceptualised as a biological disorder encompassing disturbances of mood, motor function, thinking, cognition and perception [
1,
2]. Whilst cognitive impairments in melancholia have been investigated in detail [
3,
4], definitive identification of selective neurocognitive impairments has not been achieved. Given the pressing need to examine underlying perceptual and inferential processes in heterogeneous illnesses such as depression [
5], it is increasingly recognised that a range of methodological approaches should be utilised in the analysis of neurocognitive data to more accurately capture the nature of disturbances across differing depressive syndromes. Such refined approaches have direct utility in enhancing understanding of group-specific psychophysical processes across sub-types of depression.
There is typically great inter-subject variability on tests of neuropsychological function in the major psychiatric illnesses [
6]. Meaningful interpretations of brain function in specific disorders is difficult given such variability. This is further compounded by summarising an individual’s position on a performance continuum (as with summing errors on a task) in order to infer the presence or absence of cognitive impairments. Furthermore, commonly utilised neuropsychological tests in those with depressive disorders typically rest upon broad construct-level approaches (e.g. tests of ‘executive function’ or ‘attentional control’) that do not facilitate the development of theories regarding specific psychophysical disturbances in individuals. Despite such drawbacks in assessing cognition in psychiatric illnesses, significant advances have been made over the past 20 years in explaining human perceptual inference and action [
7‐
9] using probabilistic statistical principles such as those developed through a Bayesian-based approach [
10]. The Bayesian statistical modelling approach has been applied to individual and group cognitive data across multiple cognitive domains, including signal detection that is viewed as encompassing the processes of attention, decision-making and executive functioning [
11,
12]. Formally, the signal detection capacity of an individual can be influenced by prior beliefs (or internal models of the world) and the incoming sensory stream, generating that individual’s response profiles. This, in turn, provides an ideal platform through which to measure perception and inference. In the analysis of cognitive data, signal detection theory or SDT [
13] allows modelling of the optimal detection of stimuli, through estimating discriminability and bias [
14]. SDT gives rise to measures of discriminability – how easily signal (response) and noise (non-response) trials are distinguished – and bias, reflecting how well the decision-making criterion relates to the optimal criterion. Both constructs reflect an individual’s internal model of the sensorium and their prior contextual beliefs. Signal and noise trials of a task can be represented along a perceptual strength construct in SDT, referring to the strength of inference made to a particular stimulus – that is, the probability that a conclusion (decision/action) is true given its premises. Inferences during streams of trials are continuously monitored through sensory experience and evaluation, and may then be used to update decision criteria for subsequent task performance. Rouder and Lu [
15] suggest it is reasonable to expect that on such tasks there will be significant participant-level variability in signal detection sensitivity, creating a need for statistical models that capture individual processes.
Inter-subject variability is rarely modelled in neuropsychological studies of depressed individuals. Moreover, commonly used aggregation methods have the potential to lead to statistical effect estimates that may poorly represent group heterogeneity [
15]. Bayesian statistics offer the ability to pre-specify prior knowledge through the specification of priors [
16]. A Bayesian approach to data analysis is also appealing in the setting of decision- making in the face of uncertainty because it embodies the same type of assumptions – and hence represents the same constructs – as emerging models of human decision-making [
17,
18]. When considering group data using SDT, individual subject variability can be modelled using hierarchical Bayesian techniques [
15], allowing estimation of data-driven posteriors of mean bias and discriminability as well as their variance or precision (the inverse of variance) [
11,
12]. When cognition is variably disrupted, as arguably is the case in depression, inter-subject estimates of bias and optimal judgement may be influenced, which can ideally be modelled through hierarchical Bayesian SDT analyses. There are several reasons as to why such an approach may offer significant benefit.
In health, cognitive ‘priors’ can be viewed as personal beliefs that are optimised towards the most likely value of a given percept [
19]. In depression, however, such processes may be suboptimal in different ways across individuals, extending variously across perceptual, inferential and performance domains. It has been suggested that depression is associated with distorted inference (e.g. “arbitrary inference”) at certain levels of severity (e.g. psychotic depression [
20]), yet despite recent theoretical research with Bayesian modelling in depression [
5,
21] no studies have employed Bayesian statistics to model cognitive capabilities in depressive illnesses such as melancholia. Most studies to date have attempted to delineate underlying mechanisms of negative cognitive biases [
22] based on the notion that depressed individuals have a characteristically negative view of the self, world and future [
20,
23]. Several studies have shown that attention is selectively drawn to negative information (e.g. [
24]), and that memory of negative information is enhanced [
25]. However, few studies have provided a formal quantitative framework for modelling individual level disturbances from empirical psychophysical data. While some studies (e.g. [
26]) have established evidence for neurobiological correlates of response bias, it remains to be seen whether cognitive biases extend across depression as a whole or whether they are specific to given individuals or sub-set diagnostic groups. From the findings to date it is evident that there is an unmet need in elucidating basic mechanisms of neurocognitive dysfunction across individuals with depression.
We propose that biases in emotional stimulus processing in depression can be accurately captured through investigation of different depressive sub-types using a hierarchical Bayesian emotional SDT framework. Employing an emotional word ‘go/no-go’ task, which requires responding and inhibition of responding to serially presented, randomly sequenced positive, negative and neutral words, we hypothesised that each depressed sub-set would show less optimal responding (poorer discriminability) across emotional signal conditions as compared to the control group, but that the melancholic sub-set would show more difficulty in detecting true signal trials from noise trials, particularly on emotional signal conditions (i.e. lower sensitivity) compared with non-melancholic and control participants.
Discussion
The hierarchical Bayesian SDT model implemented in this paper revealed that signal detection processes in melancholic and non-melancholic depression are significantly influenced by stimulus type and individual subject variability. Our modelling approach allowed interrogation of the neuropsychological data at two levels: the mean results across individuals in specific groups, and the heterogeneity of the groups themselves, across the psychophysical constructs of bias and discriminability. In terms of mean differences, we observed that the melancholic participants overall were less sensitive to detecting emotional signals, while non-melancholic participants displayed more liberal responses to emotional signal blocks. This provides support for our predictions that the melancholic group would display difficulty in detecting signal trials from noise trials on emotional word blocks. Also, optimal responding was found to be reduced in the melancholic group compared to the control group, as evidenced by decreased mean discriminability on neutral signal blocks. In terms of subject heterogeneity, we found that there was greater inter-subject variability of the bias estimates for the emotional signal conditions in the melancholic group, which indicates divergent bias estimates across individuals. Visualising the range of individual participant responses (Figure
5) argues against this effect being driven by outliers. Further, changes to the precision of prior distributions had little impact on the observed mean difference findings, suggesting these findings are highly robust. Despite this, however, when differing noise conditions were examined there was a moderate effect, with some previously significant differences for specific signal conditions no longer remaining significant. The observed impact of differing noise conditions thus warrants further consideration in future psychophysical studies using differing emotional and non-emotional stimuli. The main findings allow for specific neurocognitive models to be advanced with regard to depression and its sub-typing, namely the potential to gain insight into underlying psychophysical mechanisms across individuals and whether the depression is melancholic or non-melancholic in type, and highlight several important issues in the analysis and interpretation of neurocognitive data.
Our findings align with the commonly held notion (e.g. [
3]) that those with melancholic depression exhibit cognitive deficits that are observable during task performance. However, we add the observation that those with non-melancholic depression may also be impaired in their ability to perform ‘optimally’ on cognitive tasks such as the AGN. The observed trend of less optimal responding in non-melancholic depression did not, however, reach significance but may benefit from a focus in future studies. While research using the AGN task in depression [
43] claims it as a measure of ‘inhibitory control’, the analytic methods previously employed often prevent interpretation beyond a continuum of impairment (e.g. number of errors on a task). The current study is the first, to our knowledge, to utilise an affective go/no-go task in sub-types of depression. In doing so – and through analysis of the data using hierarchical Bayesian SDT – the findings offer an increased understanding of the sensitivity and discriminability capacity of individuals with differing types of depression, and highlight the importance of examining for apparent dysfunction with more refined models. Recently, Schulz and colleagues [
44] examined the convergent validity of emotional and non-emotional go/no-go tasks and concluded that, together, they offer “moderate capacity” for probing the neuropsychological construct of behavioural inhibition. Those authors also emphasized the need for testing emotional and non-emotional signal detection mechanisms in affective disorders to clarify underlying cognitive-emotional contributions. The diverging sensitivities across emotional and non-emotional conditions in the depressed groups in the current study, along with a lack of discrimination to neutral signals in melancholia, suggests that a range of cognitive mechanisms may be involved in responding to differing stimuli.
Across neuropsychological studies of depression it is evident that no single cognitive deficit model can be applied to specific groups, due to the inherent heterogeneity of the depressive domain. However, in light of the current findings of discrepancies in bias between signal conditions, it might be possible that set-shifting impairments – as previously reported in depression [
43], and more specifically in melancholic depression [
45] – play an important role. While not explicitly assessed in the current study, it is conceivable that neuropsychological constructs such as disrupted attention set-shifting and perseveration underlie the observed effects. The melancholic sub-type has been shown to be differentiated from non-melancholic depression on the basis of response selection performance [
46], where performance on compatible and incompatible trial types (e.g. stimulus–response compatibility tests) reflect cognitive strengths and weaknesses. As a rule these studies have not specified psychophysical functions of stimulus sensitivity and discrimination capacity, and have tended to report broader metrics of performance such as numbers of hits and misses across subjects. From a cognitive standpoint there is likely to be significant benefit in modelling performance-related psychophysical mechanisms (i.e. through SDT and similar analyses) in depressive disorders and then next establishing whether this provides insight towards any neurocognitive disease mechanisms. We argue that the cognitive deficits observed in different types of depression can be conceptualised in such a way so as to explain impairments in emotion-bound optimal decision-making.
Prior research on sensory processing offers further insight into the findings of decreased sensitivity to emotional signals and poorer discrimination to neutral signals in melancholia. Knill and Pouget [
9] suggest that perception of one’s environment is influenced by the likelihood of the presence or absence of relevant stimuli given an individual’s past experience (i.e. perceptual priors) with that stimuli. These factors contribute to the
relative uncertainty over one’s environment, and allow inference regarding the causes of percepts. We propose that the low sensitivity to emotional signals and lack of discriminability observed across melancholic participants may be a result of inefficient sensory integration – possibly resulting from constructs such as inefficient cognitive control mechanisms (see [
47]). This in turn may be a function of ‘inflexible’ priors (e.g. negative cognitive biases) and unsuccessful updating (e.g. such as perseveration due to a failure of emotional inhibition [
48]). The observation that the non-melancholic participants responded more frequently to noise trials on emotional signal blocks also suggests that they too are less sensitive to fluctuations in the emotional environment. Such erroneous judgements could be due to emotional processing biases in depression, a factor that has been acknowledged in accounting for decision-making impairments [
5]. Research into the probabilistic nature of decision-making [
21,
49,
50] suggests diverse mechanisms underlying optimal judgement. Neurobiologically, probabilistic learning paradigms have been used to examine human cognition [
51], with the findings pointing to distinct roles of serotonin and noradrenaline in learning and inhibition. Both neurochemicals have long been implicated in depressive disorders [
52] and may be of relevance to understanding the differences observed between and within depressive sub-types.
In addition to the behavioural changes within individuals, research using Bayesian inference has also highlighted the importance of perceptual variation between individuals. Recent theoretical work on perceptual uncertainty advocates the utility of modelling trial-by-trial updating across individuals in a Bayesian framework [
53], which is argued to be of significant benefit in conceptualising the current findings from the signal detection task. The hierarchical modelling using MCMC in the current study yielded estimates of the standard deviation of both bias and discriminability performance (i.e. the extent of the differences between measured individuals). The increased standard deviations in the melancholic group on emotional bias suggests differential individual performance profiles when compared to non-melancholic and control groups. This could be due to a range of non-cognitive factors in an otherwise homogeneous group or may alternatively reflect divergent cognitive strengths and weaknesses, which we now consider.
Several lines of research have indicated that melancholic depression is associated with diurnal variation of mood [
54], with such variation thought to impact on neuropsychological performance across the day [
55]. Clinical depression with diurnal variation has also been found to result in differential performance on accessibility and recall of positive and negative (self-related) experiences [
56], with positive memories more likely to be retrieved when depression is less severe in the afternoon/evening. Varying biological influences such as cortisol hypersecretion – shown to be specific to melancholia [
57] – may play a pivotal role in modulating cognitive function in depression as previously suggested [
58]. Such factors are important considerations with respect to the current findings given study participants were not all tested at the same time of day. Furthermore, inconsistent medication regimens across individuals within and between groups may contribute to individual differences in emotional processing biases [
59], thus possibly dampening the association between the depressed state and cognitive impairments. If such factors were found to be unrelated – and individual differences were indeed evident upon replication – it is conceivable that melancholia (due to the observed variation) may be able to be portrayed as comprising several distinct sub-types as suggested by Parker and Hadzi-Pavlovic [
1] (e.g. functional and structural melancholia). Factors such as family history, age-of-onset, presence/absence of neuropathological changes and cardiovascular disease would need to be clarified for any such sub-typing model to be advanced within the current context. Given the increased age of our sample, neuropathological changes in some individuals cannot be excluded. Despite these possibilities, the utility of the current findings lie first and foremost in their ability to inform psychophysical models of depression, with several caveats.
As indicated above, there were several study limitations. Firstly, the analysis did not interrogate trial-to-trial variability. Thus, the supposed dynamics of sensory integration (stimulus–response updating), as previously put forward [
53], could not be quantified in this sample. In addition, the signal detection task used did not allow for further analysis of aspects of emotional decision-making (i.e. specific decision-making rules) beyond bias and discriminability functions. Such specific limitations, if overcome in future studies, would provide a more refined model of decision-making and clinical conditions themselves. It is therefore proposed that future work should attempt to examine the utility of dynamic models of decision-making in light of changing emotional environments, along with key clinical variables, to further establish the mechanisms by which perception and action interact in depression.
Competing interests
There are no competing interests for any of the authors.
Authors’ contributions
MH undertook neuropsychological testing of participants, undertook the analyses and wrote the final version of the manuscript. GP held primary responsibility for overseeing clinical diagnoses and contributed to writing the manuscript and was the chief investigator on the study from which the data arose. MB contributed to the statistical modelling and interpretation of the results and assisted in the drafting of the manuscript. All authors read and approved the final manuscript.