Diagnosing Depression in Casual Conversation? Neural Networks Lead The Way

Asiket Singh Dhillon

Neural Networking and pattern recognition are a rising force in technological sector. Once we can teach machines to ‘learn’, the avenues that machines can be used in are ever expanding. Improved neural networks could possibly yield higher efficiency rates in automatic screening of individuals who may have cognitive and behavioural abnormalities.

Earlier this year Tulka Anhanai, Mohammed Ghassemi and James Glass came together to present their paper titled Detecting Depression with Audio/Text Sequence Modeling of Interviews. It puts forward a model of detecting depression in individuals which rather than relying on traditional iterations of uses, what they call “context-free” detection. Traditional models may analyse a set of answers to predetermined questions. For example, “How often do feel low/depressed?” / “Does your family have a history of any mental diagnoses?”. The model knows what a depressed person would say and compares the individual’s answers to see where they fit.

Al Hanai’s model analyses the content as well as the manner of expression of answers. Therefore, it is not constrained to the specific conversation but can utilise patterns recognised in other individuals that may exhibit similar content and exhibition to reach its assessment. Depressed individuals may speak slower and with lower voice modulation in comparison to healthy individuals. The model recognises certain patterns which emerge and then presents its diagnosis. This is known as sequence modelling which allows it to not be constrained to a case-by-case analysis and use recognisable patterns and apply those to a variety of contexts. Sequencing allows for the analysis to not be restricted to pointed questions and direct answers but rather for entire conversations and detect the severity of depression as well.

Out of a total of 142 subjects that the system analysed it labelled 28 individuals as those with depression. In terms of precision the system scored 71%. Out of the 142 some were used as training for the model and others for development which left 47 (25%) of the subjects which the model had to assess. While this is far from ideal, it is a learning step. The implications of such a model means that it would be possible to not rely only on self-reports and manual analysis which may be hindered by the assessor’s biases. The system will be able to analyse casual conversation and provide assistance. Furthermore, individuals that cannot physically be present at the practitioner’s office will also fall under the purview of the system. Clinicians are then, no longer restricted to asking specific questions and direct answers within specific context.

What the researchers found upon poring over the results was that the system requires over 4 times more audio interactions (30) to identify patterns as compared to text (7).  I believe that this may be due to the fact that audio would require far more categorisations than text. Such as modulation, tone, pauses between words and tempo. The textual analysis is fairly constrained to the content of the text as well as arrangement. While the parameters that the system works on are understood, “How” the system reaches its conclusions as well as what patterns it recognises are hard to discern. This is due to the nature of neural networks and their functioning, i.e the individual nodes and their function in the larger network is largely unknown. So we know the input which is the audio and transcripts and the output which is the model’s assessment.  Figuring out how it reached that conclusion is the next step forward.

The researchers hope that this model paves the way for different, specialised versions which can diagnose other impairments such as Dementia. The way forward for the model is a symbiotic relation between the clinician and the model. Improving the accuracy of the model as well as increase the understanding that we have of the inner workings of the neural networks.

Matheson, Rob. (2018, August 29). Model can more naturally detect depression in conversations. MIT News. Retrieved from http://news.mit.edu/2018/neural-network-model-detect-depression-conversations-0830

Al Hanai, T., Ghassemi, M., Glass, J. (2018) Detecting Depression with Audio/Text Sequence Modeling of Interviews. Proc. Interspeech 2018, 1716-1720, DOI: 10.21437/Interspeech.2018-2522.