Symposium 1 - Application of Artificial Intelligence to Medical Practice

Generative AI represents a sophisticated branch of artificial intelligence technology, renowned for its ability to produce new and original content encompassing images, text, audio, and video. Unlike conventional systems that simply fetch existing information, generative AI exhibits creative flair, drawing upon self-supervised learning techniques. It undergoes a "training" phase, during which it is fed vast collections of data, and harnesses the power of deep neural networks along with probabilistic models to decode and assimilate patterns observed in pre-existing creative outputs. After this intensive training regimen, generative AI systems acquire the proficiency to craft new images, videos, text, and various forms of media that bear a close resemblance to the input data they were trained on.


The presentation will commence with a concise overview of generative AI, highlighting diffusion models and large language models as prime illustrations. It will proceed to explore the cutting-edge developments within the realm of generative AI, subsequently introducing various scenarios. These scenarios will serve as a platform for the presenter to engage with the audience, collaboratively seeking solutions to contemporary challenges.

Intracranial haemorrhage (ICH) is a diagnosis that needs to be detected and managed promptly, and failure to do so can have serious consequences. An automated workflow incorporating vendor artificial intelligence (AI) software was devised for the detection and management of patients with ICH in the Accident & Emergency (A&E) Departments.
A validation study was first performed on approximately 800 CT brain scans to ascertain the accuracy of ICH detection and segmentation by vendor AI software. The sensitivity, specificity and accuracy were approximately 95% or higher for the detection of ICH.
A pilot project was then initiated to establish proof of concept (POC) of detecting patients with ICH and facilitating timely management of such patients in A&E. To this end, an automated workflow incorporating the vendor AI software was established, whereby the AI results could be accessed by doctors on a web browser based interactive portal set up on the hospital's intranet, and this was made accessible in all cubicles and clinical bays in the A&E, as well as PACS workstations in the Radiology Department (RD). The portal included a dashboard with a notification system which served as a mechanism to alert doctors when patients were detected to have ICH. This workflow acted as an 'assistant' to A&E doctors, and was described to reduce 'stress' in detecting ICH in patients, especially in patients with subtle ICH, or at times of constrained manpower or during overnight shifts.
This workflow was subsequently scaled for use in the A&E of 17 hospitals by joint collaboration between A&E, Radiology, IT and administrators. The process of validation, establishment of POC, and workflow including network infrastructure, data and cyber security, user access control, and means to subsequently scale up to cover 17 hospitals will be discussed.