Artificial Intelligence at Ieso - Part 2: The Present - Valentin Tablan
On March 1st 2017 I gave a presentation at the summit dedicated to Deep Learning in Healthcare, organised by ReWork in London. I used that opportunity to talk about Ieso’s work on using artificial intelligence to improve mental health provision in the UK. This blog post is a companion for the content of my presentation.
Part 2: The Present
Last time I was talking about moving from paper and pencil to an internet platform and how that allowed us to treat more than 10,000 patients. Today we’ll have a look at what one can do with all the experience accumulated while providing that service.
Every time a patient refers themselves to Ieso Digital Health's online talking therapy service, and completes the questionnaires, when they receive treatment from their therapist, or when they exchange messages with their therapist between sessions, they do that via Ieso’s bespoke technology platform, which captures and stores these interactions. We also know which of these patients recovered, and which didn’t, which patients dropped-out of treatment and how many treatment sessions were needed to bring each diagnosable condition to below the clinical threshold. Having treated 10,000 patients, there's an exciting opportunity to use the historic data accumulated and develop a data-driven deep understanding of the therapy process. This unique dataset can tell us what works best in therapy for each individual sub-group of patients: males or females, younger or older people, those pregnant or those suffering from a chronic medical condition.
To confirm the usefulness of this dataset, we ran an experiment and created a deep learning computer model that could be used to read the patients’ description of their problem, and then use that description to come up with a hypothesis about what their principal presenting condition may be. The model was built by looking at the self-provided descriptions from all our previous patients, and associating the linguistic features present in those descriptions to the presenting problem as diagnosed by each patient’s individual therapist. Because this model uses statistics, it is able to generate numeric scores indicating the strength of association between textual description and a particular condition. These scores can be useful when identifying potential comorbidities (for example when two conditions get very similar scores), or situations where the model is uncertain (when multiple conditions get similar scores, and no particular value dominates the others).
When we measured the accuracy of this statistical model, we found it to be very similar to the level of agreement between two different therapists. This means that the top-scoring diagnoses (provided by the system) agree with those given by a therapist roughly the same number of times as another therapist’s diagnoses would agree.
This experiment has proved that the dataset we've collected does indeed have value, and can be used to build computer models with good predictive performance.
What next? you may ask... Well, you’ll have to wait for the third and final part of this blog series Part 3: The Future.