parameters, run the workflows, and then capture and analyse the images. ‘But using Imagence, one single biologist can now run the complete workflow from image production to the interpretation of results,’ Dr Steigele says. ‘It saves time, money, resources, and a multidisciplinary workforce. Labs can now effectively scale-up throughput and data output massively; and whereas classical approaches to HCS require major computing clusters to analyse huge sets of images, we use some essentially very primitive hardware setups,’ he adds. Imagence is the result of collaborations

between Genedata and leaders in the biopharmaceutical industry. It was the basis for a Genedata project with AstraZeneca, Deep Learning for Phenotypic Image Analysis, which won the 2018 Bio-IT World Best Practices Award. ‘We wanted to remove the burden of classical analysis, in which humans have to think about, recognise and then handcraft features covering relevant phenotypic information from the cell images,’ Dr Steigele notes. ‘For AI-based approaches, the biologist just provides the algorithm with a curated training set, and the software can then learn by itself which features to extract and are best suited to differentiate between the existing trainings classes (eg, images of dogs vs cats or different cellular states in the applied example of high-content screening). It’s the expertise of the biologist that drives the whole process, but the underlying complexity is hidden.’ Efficient generation of training data is

key. As Genedata explains, the process involves generation of interactive, human readable maps from images in which similar responding cells present in the images co-cluster and thereby allow a very fast exploration of the phenotypic space and the collection of training data for all phenotype classes of interest; a process typically applied during assay development and via deep learning, which is capable of detecting even very subtle differences. It’s then over to the biologist, who has the expertise to understand which phenotypes are important for the assay in development. Relevant training data are then curated in just a very few hours to train networks for production application on quantities of screening data (typically 50,000 to 2.5 million tested substances per screening campaign).

Advancing healthcare The ability to apply deep learning techniques to typically workforce- intensive tasks in discovery and preclinical R&D will have important knock-on effects | @scwmagazine

for personalised medicine and other areas of clinical practice, Dr Steigele believes. Large-scale assays using low-volume clinical samples, coupled with more consistent data extraction and analysis, will help to identify which and how drugs impact on subsets of patients, and make it faster and more cost-effective to apply technologies, such as HCS, for the fast- developing field of personalised medicine. Come out of the lab and into the clinic and healthcare environment, and artificial intelligence is being harnessed to aid and speed diagnosis at the patient bedside. Transformative AI is harnessing artificial intelligence and analytical techniques developed at CERN and Cambridge University to generate predictive monitoring tools that they hope will ultimately save patients’ lives. ‘We believe that predictive analytics hold the key to transforming healthcare,’ states Dr Marek Sirendi, CEO and co-

“We believe that predictive analytics hold the key to transforming healthcare. We aim to transform the emergency medicine paradigm from rapid response to personalised preparation and prevention”

founder at Transformative AI. ‘We aim to transform the emergency medicine paradigm from rapid response to personalised preparation and prevention.’ The firm’s first product, designed for

use in hospitals, analyses data from monitors to warn doctors in advance that the patient may be likely to suffer deadly cardiac arrhythmias. The algorithm detects tiny changes in physiology that are predictive of sudden cardiac death, triggering an alarm that gives doctors the opportunity to prepare for, and potentially prevent the episode, explains Dr Sirendi. Perhaps surprisingly, the Transformative AI algorithm used for predictive monitoring in hospital settings is based on decision algorithms developed and used at CERN’s large hadron collider, which detect exotic proton-proton collision events in real time. ‘The algorithm employs a number of state- of-the-art deep learning models along with other machine learning frameworks. We present it with examples, and ask it to learn to recognise the cases of interest.’ This basic AI approach can thus be

used in multiple fields, from business analytics, to physics or even predictive medicine. ‘AI is best thought of as an optimiser and a mathematical learning machine,’ Dr Sirendi continues. ‘It can streamline many processes encountered in the modern workplace, which can boost the productivity of existing business activities. But AI can also reveal unexpected insights from large messy datasets, which opens up entirely new product categories. Our AI is designed to accomplish both tasks. It will make alarms more actionable and clinically relevant, while identifying something the human eye is unable to spot – subtle changes in human physiology that precede the onset of sudden cardiac arrest.’ And that predictive capacity can save

lives, the company believes. Although doctors can give lifesaving CPR and defibrillation to patients already in cardiac arrest, the arrest occurs suddenly and there is no warning, so treatment is started only after blood has stopped flowing to the brain. This increases mortality rates and can have damaging effects on long-term neurological function among survivors, Dr Sirendi states. AI could be used to warn doctors in advance of a cardiac event. ‘Telemetry monitoring only identifies arrhythmias after they have started. Our algorithm can potentially predict deadly arrhythmias up to an hour before they begin, giving doctors a chance to proactively manage this life-threatening condition.’ Importantly, the algorithm also reduces the number of false alarms. ‘Nurses are frustrated by an abundance of alarms, rather than a lack of them,’ Dr Sirendi adds. We cut out 54 per cent of irrelevant alarms while making the remaining more clinically relevant. The key is to make technology smarter, so that people (the nurses, doctors, technicians) can rely on it to a greater extent.’ Ultimately, the decision on whether to act on a machine-derived prediction is up to the clinician, who will also have blood test and potentially other clinical data and results to inform their course of action. ‘We’re working with a number of hospitals, cardiologists and electrophysiologists,’ Dr Sirendi notes.

Engaging clinicians and healthcare

providers will be key if AI is to be accepted into mainstream healthcare, the company believes. ‘To get healthcare providers excited about integrating AI into healthcare, new tools shouldn’t just replicate tasks that human healthcare workers are capable of. Rather, AI tools should provide novel insights that elevate the standard of clinical care.’

October/November 2018 Scientific Computing World 23

Page 1  |  Page 2  |  Page 3  |  Page 4  |  Page 5  |  Page 6  |  Page 7  |  Page 8  |  Page 9  |  Page 10  |  Page 11  |  Page 12  |  Page 13  |  Page 14  |  Page 15  |  Page 16  |  Page 17  |  Page 18  |  Page 19  |  Page 20  |  Page 21  |  Page 22  |  Page 23  |  Page 24  |  Page 25  |  Page 26  |  Page 27  |  Page 28  |  Page 29  |  Page 30  |  Page 31  |  Page 32  |  Page 33  |  Page 34  |  Page 35  |  Page 36