AI for Future Digital Health workshop - December 2020

The AI for Future Digital Health workshop was held, on Wednesday 8 December 2020, as part of the British Computer Society Specialist Group in AI's 40th annual AI conference.

The event is usually held in Peterhouse college in Cambridge, but this year, for obvious reasons was online. There were 412 registrations for the main conference, of which 84 attended the AI for Future Digital Health workshop. The workshop featured guest speaker Kerstin Bach from the Norwegian University of Science and Technology and showcased work in progress at the University of Liverpool under the umbrella of Liverpool University's Doctoral Network in AI for Future Digital Health. The workshops was chaired by Nirmalie Wiratunga from the Aberdeen Robert Gordon University, and Vitaliy Kurlin and Frans Coenen both from the University of Liverpool.

The guest speaker, Kerstin Bach presented a device, hosted on a smart phone and founded on Case-Based Reasoning (CBR), for prescribing self-management exercise routines for individuals suffering from lower backpain; the main cause of disability in Europe. It was interesting to note how well CBR can work with only a few examples. The current case base features 300 patients but many more are anticipated, obtained using the feedback loop that is a feature of the system.

The remainder of the workshop was divided into two session, four presentations per session. The first session featured talks on Leukaemia, mobility devices for young children, biomarkers for liver cancer and corneal disease by Muizdeen Raji, Peter Wright, Mohamed Elhalwagy and Weiqiang Chen respectively. The first talk was directed at using machine learning in the context of cytometry (the measurement of the characteristics of cells). This was couched as an unsupervised learning problem. The data was clustered using a self-organising map technique which was then used as a binary classifier. Muizdeen Raji observed that a particular challenge of using cytometry was data acquisition, especially given the current pandemic restrictions. The second presentation focused on children with mobility disabilities. Peter Wright emphasised the need for early mobility support and proposed a nobility device, fitted with sensors, to help disabled nine-month-olds to crawl. Great video simulating a toddler using the mobility device. The talk raised some interesting questions concerning the ethics of testing the device, is it OK to use real babies? Mohamed Elhalwagy noted that ultrasound techniques for detecting liver cancer did not work well with respect to tumours that were 5cm or less in length; the stage when intervention would be most effective. Instead, a mechanism using bio-markers was proposed. The presentation investigated a range of bio markers. In the end a combination of markers was proposed, the GALAD score, built into a linear regression model for predicting liver cancer. There was much discussion afterwards as to whether the assumed linearity between the dependent variable and the independent variables was valid. The final presentation, in this first session, was directed at the detection of corneal disease using Optical Coherence Tomography (OCT) scan technology. Weiqiang Chen observed that there were a range of challenges associated with this technology involving the nature of the hardware used, segmentation of the resulting 3D images and their sub-sequent reconstruction. The presentation focused on AI solutions with respect to these challenges.

The second session featured presentations on predicting influenza epidemics, prioritising pathology data, sleep pose recognition and cycling to improve public health by Yanhua Xu, Jing Qi, Omar Elnaggar and Aidan Watmuff respectively. The first presentation, the prediction of influenza, seemed particularly apt given the current circumstances (COVID-19 pandemic). The central theme of the presentation was that intermediary influenza strains from animals can provide a useful predictor of influenza in humans. Yanhua Xu proposed a mechanism founded on the BLAST (Basic Local Alignment Search Tool) algorithm for comparing human-animal protein sequence information to predict Influenza A epidemics. Jing Qi presented work on prioritising pathology results in the absence of a ground truth. Two solutions were presented, a point-based approach and a time series approach with the second argued to be the better. Much discussion at the end of the presentation as to whether a ranking approach would be more beneficial than the proposed binary approach, priority versus non-priority. The third presentation was directed at sleep pose recognition as a means of supporting the hypothesis that Plantar fasciitis (pain on the bottom of the heel and/or sole of the foot) is affected by sleep posture. Omar Elnaggar proposed a Support-Vector Machine (SVM) model to classify twelve alternative sleep poses. The final presentation considered the benefits that cycling can bring to public health. Aidan Watmuff, who confessed to being a keen cyclist himself, presented an unsupervised analysis of a 1,208 cyclists in the Liverpool City Region. There was lengthy discussion concerning the clustering algorithm used, k-modes, and whether alternative forms of unsupervised learning might be more appropriate.

Overall, the workshop demonstrated the extensive scope of the many health related applications that can benefit from AI. Access the workshop presentation slides.