Semantic Representation of Fixation Sequences and the Classification of Bipolar and Unipolar Patients Presenting in a Depressed State

Chung, Jonathan 1 ; Rakita, Uros 4 ; Giacobbe, Peter 4 ; Eizenman, Moshe 1, 2, 3

1. Department of Electrical Engineering; 2. Institute of Biomaterials and Biomedical Engineering; 3. Department of Ophthalmology and Vision Sciences; 4. Centre for Mental Health

Bipolar Disorder (BD) is characterised by episodes of depression and elevated mood (manic state). In the depressive states, there are no pathognomonic clinical differences between patients with BD and patients with Major Depressive Disorder (MDD, unipolar). As such, during depressive episodes more than 20% of patients with depression are wrongly diagnosed. To minimise the risk of inappropriate diagnosis, there is a need for an objective test that helps to disambiguate BD and MDD. We used differences between the visual scanning behaviour of MDD and BD patients on images of emotionally valenced faces to develop a novel test to distinguish between individuals with BD or MDD during a depressive state.

Twenty-six patients with a diagnosis of BD and 47 patients with a diagnosis of MDD participated in the study. All the participants were tested in a depressed state (i.e., HAMD =20). Patients naturalistically viewed slides containing faces displaying a range of emotions and their eye movements and visual scanning patterns were analysed. Differences between the visual scanning behaviours of the two groups were evaluated by comparing features that describe average visual scanning behaviour (e.g., the total fixation time, number of fixations) on emotional faces. Furthermore, the ordered sequences of fixation on regions of interest within and between emotional faces were used to construct novel semantic representation of fixation sequences. The semantic representation of fixation sequences for patients with BD or MDD were learnt by a neural network that was used to construct a novel maximum likelihood classifier.

Patients with a diagnosis of BD had significantly longer fixation times (t = 2.66, p = 0.01) and significantly more visits (t = 2.39, p = 0.02) to faces with happy expressions than patients with a diagnosis of MDD.  Patients with a diagnosis of BD also had significantly less fixation (t = 2.15, p = 0.03) on faces with dysphoric expressions than patients with a diagnosis of MDD. When the novel maximum likelihood classifier that we developed was used to classify the two groups of patients, the accuracy of the classifier was 80.8% and the area under the curve of the receiver operating characteristics was 0.90.

The results demonstrated that traditional matrices that are commonly used to describe visual scanning behaviour can help to characterise differences between groups of patients but they are not sensitive enough to characterise individual patients. Using a novel classifier that is based on semantic representation of fixation sequences, we were able to capture more fully the temporal and spatial differences between the scanning behaviour of the two groups and to improve significantly the performance of the classifier. An objective method of differentiating BD and MDD in a depressed state can better inform treatment decisions and prevent the prescription of ineffective therapies.