Working to bring Artificial Intelligence into life-saving medical care

We are continuing our salons on the intersection of health and tech, as we explore and develop our health thesis. At the same time we want to identify players who can help start-ups we invest in. We’re sharing our learnings and would love any feedback.

As early investors in Zebra Medical Vision, we believe Zebra will continue to revolutionise radiology diagnostics and improve healthcare. But what are the next frontiers? We think pathology is interesting given its core to diagnosis and monitoring of a wide range of conditions, in particular cancer. However, we have just learned through our most recent salon — digital pathology is incredibly challenging.

So what is the potential for AI in pathology diagnostics? What are the potential barriers to adoption in the NHS? We were delighted to welcome two of the UK’s foremost experts in this area whose day job is finding ways to bring the exciting new technology and diagnostic techniques into mass healthcare.

To learn more about pathology, we asked Dr Clare Craig, Consultant Pathologist with an interest in AI, to come to Platoon on a Friday afternoon to talk about her work. We were also joined by Prof Chris Holmes, newly appointed National Director for Health at the Alan Turing Institute who told us about their new health programme, their partnership with Health Data Research UK, and we discussed how they can play a role in acting as a secure broker to NHS and population health data.

Dr Clare Craig works for Genomics England (100,000 Genomes Project), as part of the cancer team. One aspect of her role is responsibility for Digital Pathology — in the future this will be incorporated into the cancer patient’s clinical record, sitting alongside the associated genomic data.

Clare’s interested in how we can create a high quality standardised digital pathology library including as much clinical and imaging data as possible which would provide the best value to artificial intelligence companies wanting to build diagnostic or prognostic algorithms.

Current digital pathology use cases include second opinions of frozen sections, counting, and quantifying; but recent exciting developments include using AI to improve the prediction of the overall survival of patients diagnosed with brain tumours from microscopic images of tissue biopsies and genomic biomarkers. As well as improving prognostic predictions AI has potential to better classify patients for treatment for example identifying patients who won’t benefit from neoadjuvant chemotherapy, so their surgery isn’t delayed.

Challenges to AI development include accessing the material and data to build libraries (not trivial, as unlike imaging data, clinical data comes from multiple organisations and require consent/ethics approval); the quality of those libraries, and, as we see across the board, IP ownership. Pathology is particularly challenging. Digital pathology scanners and slides are extremely variable — both between laboratories and within an individual scanner even to the point of the vibrations of a bus going by and impacting the angle of refraction of the lenses resulting in different RGB values. The relative importance of the many variables is not yet fully understood. NHS pathology laboratories also face typical challenges of funding and regulation, and there is potentially resistance from pathologists, who may fear being replaced.

Beyond that, Clare highlighted the need for a gold standard for AI pathology. When training AI algorithms there needs to be an outcome that they are working towards. This could be a consensus pathology opinion; a survival outcome or a response to treatment outcome. Libraries for AI are often built up by pathologists who select images from areas of solid tumour with a textbook appearance which introduces bias. One technology alternative may be Maldi (matrix-assisted laser desorption/ionization) otherwise known as imaging mass spectrometry. This could allow automated labelling of cell types within an H&E image for accelerated AI learning. In recent years, a great many advances in the practice of imaging mass spectrometry have taken place, making the technique more sensitive, robust, and ultimately useful.

We were also thrilled to hear from Chris Holmes, who is spearheading the health initiative at The Alan Turing Institute. The Alan Turing Institute was founded in 2014 to ensure Britain leads the way in big data and algorithm research. It’s a network of charity, industry and government partners in collaboration with a network of university partners and strategic government funding. Based in the British Library, around the corner from us, they house over 200 Fellows, PhD students, data scientists, interns and business team members. The institute aims to advance data science and AI to revolutionise health care, among seven other key themes.

The Health programme has been established in partnership with Health Data Research UK and scientific priorities are:

1. Actionable Health Data Analytics

Structured and unstructured (e.g. imaging, text) data for derivation of new or deep phenotypes

Adding value at scale to existing world-leading cohorts in the UK

Demonstrating system-wide opportunities for research that improves quality of care

2. Precision Medicine

Enable large scale, high-throughput research that combines genomic data with electronic health records

Genomics, epigenomics, statistical and complex genetics, population genetics, cancer ‘omics’, molecular epidemiology

3. 21st Century Trial Design

Transform Phase II — Phase IV clinical trials including ‘real world evidence’ studies

4. Modernising Public Health: towards prevention and early intervention

Ability to link health and administrative datasets across multiple environments

New technologies, from sensors to wearable devices to artificial intelligence

Some of their existing health research projects include improving cystic fibrosis healthcare, cancer pre-diagnostic analytics with AI, diagnosing mental health disorders and British Heart Foundation and University College London Hospitals partnerships.

For companies looking for NHS and population data, Prof Chris Holmes was keen to continue the discussion as to how the Alan Turing Institute can play a role as secure broker to this data. Please do get in touch with any considerations or hurdles.

We love companies that solve incredibly difficult problems that can impact many lives using technology. If that’s you, we’d love to hear from you. As a starting-point, we believe AI Diagnostics companies need:

— Team: balanced teams with deep relevant life science and tech expertise

— Data: A lot of quality training data, annotation of that training data, AI Framework

— Route to market: Good integration with hospital workflow and path to commercialisation