Facilitating cognitive research in clinical environments: Automated scoring of the Autobiographical Memory Task

This project aims to train a machine learning model to identify different types of text-based autobiographical memories, that is, memories of an individual's past experiences. An impaired ability to retrieve 'specific' memories such as "I went swimming last night" is a key predictor of the course of depression. Identifying memory deficits in individuals, and how these memory difficulties might influence response to psychological and pharmacological treatments, can help improve outcomes for mental illness.

The gold standard for assessing memory retrieval is the Autobiographical Memory Task (AMT). This is a cued-recall task in which individuals type out a memory in response to a cue word. Currently, categorising individual's reported memories is reliant on human reading and scoring of the response using a coding manual – with each AMT taking five minutes to score. This reliance on human scoring limits how widely we can use the AMT along with the accuracy of data.

In this project we will work with MDAP to produce a machine learning model able to score text-based memory responses on the AMT. The model will be trained on a training dataset (9,000 text-based memories) and test dataset (8,400 text-based memories), both of which have been annotated by human-raters.

Developing this model will allow us to embed the memory task in the thousands of assessments taken by nation-wide treatment services. The model will also support digitisation of our memory-based intervention which has demonstrated treatment efficacy in paper-format, thus providing a platform for planned grant applications, and an intervention prototype to attract seed funding.