Dr Clare Matthews and Dr Rabia Aziza’s poster presentation at the Child Health Technology Conference 2022 showcased a fascinating project that Aparito have worked on with University College London to explore “Feasibility of facial expression analysis as an objective palatability assessment of paediatric medicine”.
This project utilised Aparito’s innovative video capture & analysis technology via the Atom5™ clinical trial platform to enable study participants to upload videos capturing their responses to taste. The videos were analysed by our data science team and their findings were disseminated via the poster.
Below is the transcript from Dr Matthews’ presentation along with a link to download the poster in PDF format!
Dr Clare Matthews, Senior Data Scientist at Aparito
“The palatability of drug formulations is an important factor to consider when developing paediatric medicines as it has a strong impact on treatment adherence and clinical outcomes. At Aparito we specialise in collecting data from studies and trials within a home setting. This reduces the burdens associated with clinical assessments for both patients and caregivers, but can also provide more accurate data, as children especially are likely to behave more naturally in a home environment. As part of this study, participants recorded videos of themselves testing four different taste strips.
The videos, along with a five-point smiley-face-based rating of the taste, are uploaded using Aparito’s Atom5™ app.
Dr Clare Matthews, Senior Data Scientist, Aparito
We’re looking at the feasibility of analysing facial expressions in the videos using pose estimation and facial recognition software to explore how facial points move in response to children’s reactions to different tastes, specifically, the tastes of sweet, sour and bitter.
All participants tasted a blank control strip first. The remaining three strips were then tested in one of three different orders, which were assigned randomly via the app. From the five-point hedonic scale rating that the participants submitted for each strip, we found the somewhat expected result that children like sweet, dislike bitter and have a varied response to sour. However, this result validates our methodology by illustrating that we can get meaningful data from a decentralised approach.
From the videos, we see a lot of variation in the magnitude, duration and timing of the reactions to the taste trips. Our approach has been to initially identify single frames from the videos to illustrate a baseline, or no reaction, and a best reaction facial expression. We generate quantitative measures based on the relative position of specific landmarks in these facial expressions.
These measures are then the input features on which our machine learning models are trained and we’re developing models to classify both the strip taste and the hedonic ratings.”