Seeing compassion through a child’s eyes

Use of avatars to help patients with depression overcome feelings of high self-criticism and low self-compassion was demonstrated by Professor Chris Brewin, UCL Division of Psychology and Language Sciences, London, United Kingdom. In his demonstration, 43 highly self-critical female patients became embodied as avatars and were asked to comfort an upset avatar child. The child responded to the patient’s avatar and was reassured by the compassion shown by the patient. The patient was then embodied in the child avatar and allowed to see and hear their compassionate intervention from the child’s perspective.

Assessment scales specifically developed to measure changes in self-compassion and self-criticism indicted that just a single avatar experience helped patients to change the degree of harshness with which they judged themselves. Repeated exposure helped reinforce this finding.

A single experience of avatar self-compassion helped patients with depression to change the degree of harshness with which they judged themselves

Interestingly, patients had no fear of expressing compassion for others. In real-life, they were spontaneously showing compassion to others even though their ability to show self-compassion was still outside the healthy range. With practice using the avatar, this fear of self-compassion was trending upwards.

Patients don’t like hearing their own voices

Interestingly, patients didn’t like the sound of their own voices – which was heard while they, in the guise of avatar child, were being comforted. However, Professor Brewin suggested that becoming comfortable with their own voices might be an integral part of the therapy. It may be that patients will find use of virtual reality programs, such as this, a means of experiencing situations they fear or long for in reality.

Depressed patients see things differently

A machine learning algorithm could also prove valuable in antidepressant prescribing, as Gerard Dawson, P1vital LTD, Psychiatry, Oxford, United Kingdom outlined. Usually, it can take between 4-6 weeks to determine whether a patient is deriving benefit from antidepressant therapy. However, research has shown that patients with depression process emotions differently from non-depressed controls. Where a control sees a face expressing happiness, a patient with depression might see underlying sadness; particularly, if the face is caught not wholly expressing an emotion.

Antidepressants, however, induce changes in this perceived negative emotional bias in patients with depression very rapidly post-commencement of therapy; suggesting that the monitoring changes in negative bias, and thus emotional processing, might afford a quicker means of predicting whether or not a therapy will be effective.

Monitoring changes in emotional processing might afford a quicker means of predicting whether or not antidepressant therapy will be effective

A machine learning-derived algorithm was developed by combining changes in emotional processing with information derived from the Quick Inventory of Depressive Symptomatology (QIDS) questionnaire and was incorporated into an app. Patients could use this app to regularly send completed questionnaires back to their treating physician for evaluation.

A proof of concept study using 58 patients recruited from 10 general practitioners’ (GP) practices predicted which patients would fail to gain benefit from antidepressant therapy with a 75% accuracy. Would GP knowledge of changes in negative bias in their patients facilitate early switching antidepressant therapies (instead of waiting longer) and attainment of better patient outcomes?

PReDicT - but will GPs act on the information provided?

To determine whether this was the case, a randomized controlled trial - the PReDicT Test (the Predicting Response to Depression Treatment Test) study – was undertaken in 900 patients. The study compared patients given treatment as usual to those in whom negative emotional bias was monitored after 1 week of antidepressant therapy. It should be noted that general practitioners could choose to use the information to change therapy or not, based on changes in emotional bias observed, as they saw fit. Cost-effectiveness and patient acceptance of the technology were also assessed.

The results gained from the study suggest that there were indeed early behavioral changes in clinicians in the PReDicT arm of the study that led to better outcomes for patients. 

However, it should be noted that; in over 50% of cases where the algorithm suggested the patient would gain no benefit from their current antidepressant therapy, clinicians did not switch medication and chose to ignore the prediction. Furthermore, because the proof of concept study was based on use of a single antidepressant therapy (whereas in the PReDicT study patients had been given many different therapies), the algorithm was over-predicting positive responses to medication. Both factors did lead to lower levels of patient benefit in PReDicT than could otherwise have been obtained.

In over 50% of cases where the algorithm suggested patients would gain no benefit from their current antidepressant therapy clinicians chose to ignore the prediction

Accurate patient reporting and good acceptance of technology

What was clear was that patient self-reporting using the questionnaire was as accurate as that clinician-based reporting; and levels of compliance over the course of the 8 weeks study was high. Patients valued the technology – they believed that the clinician was monitoring their test results and would step in if something was wrong – even those patients in the control arm of the study. Clearly, Dr Dawson concluded, patients with depression are really interested in getting better.

Patients valued the technology – they believed that the clinician was monitoring their test results and would step in if something was wrong

Professor Andreas Meyer-Lindenberg, Central Institute of Mental Health, Mannheim, Germany, presented machine and deep learning methods, focusing essentially his presentation on algorithms and artificial neural networks. Systems have moved on considerably from uni-layered neural networks - or perceptrons - due to three discoveries that have led to the explosion in their applicability.

Hidden layers of neural networking

Artificial neural networks need to be multilayered in order to work and the most powerful systems, deep neural networks, have lots of hidden layers. If information is fed-back into the network, such recurrent networks are even more powerful. Back propagation of errors in predicting outputs allows networks to be trained. Thirdly, use of efficient hardware in the form of graphics processors enables machines to be efficient multitaskers.¹

To train networks, lots of data - or BIG data are needed. And this is not something that psychiatry is short of.  In fact, the amount of data being gathered is increasing exponentially. However, hardware is now functioning at maximum capacity and many organizations are seeking alternatives to conventional computer architecture.

By integrating, for example, networks handling imaging and genetic data, useful phenotypes can be uncovered. Similarly, smart phones can collect data that can be utilized to identify and determine patient grouping. Essentially, think of an area of psychiatric practice and the chances are that artificial neural networks will aid further research.