Things & Thinks XIX
This is a busy edition! I have curated some headline healthcare AI news and Big-Tech’s splendid headway through 2020, followed by an interesting perspective on hedging for failures. As usual, the newsletter ends with a Chart of the Fortnight and a Tweet of the Fortnight. Happy Reading!
State of AI and healthcare
Nathan Benaich and Ian Hogarth published fantastic overview of progress in artificial intelligence in 2020 with their compilation-State of AI report. I reviewed the report’s healthcare angle and here are the nuggets-
- The overarching ‘big ticket’ advancement for 2020 seems like the several advances we have had in drug discovery. The report lists several of these, in research phase and also on the industry side. Research advances include deep learning on cellular microscopy, designing ‘logistically practical’ drug candidates using reinforcement learning, ML generated drug candidate synthesis plans and graph networks to guide antibiotic drug screening. Moving further away from research to practical industry applications, several other advances happened including Exscientia advancing AI designed molecule in Phase I, several start-ups receiving funding for their AI-first drug discovery platforms and launch of Project Melody as the privacy-preserving machine learning platform for drug discovery.
- After diagnosis of ‘wet’ age-related macular degeneration (exAMD) in one eye, a computer vision system predicted whether a patient’s second eye will convert from healthy to exAMD within six months. The system used
3D eye scans and predicted semantic segmentation maps (ref). - The COVID Symptom Study app collected and analyzed the health of over 4 million global contributors to discover new symptoms, predict COVID hotspots and using AI, eventually predicting COVID-19 without a physical test. ZOE is running the world’s largest clinical study to validate the prediction model (ref).
- Prospective testing began for privacy-preserving AI applied to medical imaging. While the pooling of medical data should lead to improved medical knowledge and clinical care, it is also an area with strong safeguards around privacy. New techniques will enable privacy-preserving innovation; e.g. the 5P Project demonstrated federated learning and encrypted inference on paediatric chest X-rays in a clinical setting.
- A review of 20,000 recent studies in the field found that less than 1% of AI based medical imaging studies had sufficiently high-quality design and reporting. New international guidelines were drafted for clinical trial protocols (SPIRIT-AI) and reports (CONSORT-AI) that involve AI systems in a bid to improve both quality and transparency. (ref)
- Causal reasoning is a vital missing ingredient for applying AI to medical diagnosis. To overcome this, diagnosis can be reformulated as a counterfactual inference task that uses counterfactual diagnostic algorithms. e.g. when compared to the standard associative algorithm and 44 doctors using a test set of clinical vignettes, one counterfactual algorithm placed in the top 25% of doctors, achieving expert clinical accuracy. In contrast, the standard associative algorithm achieved an accuracy placing in the top 48% of doctors.(ref)
- Viz.ai was granted a New Technology Add on Payment of up to $1,040 per use in patients with suspected strokes. This is the first reimbursement approval for a deep learning-based medical imaging product granted by the Centers for Medicare and Medicaid Services (CMS) in the USA. (ref)
Digital Health and Big Tech in 2020
Instead of the regular round-up of digital health news, here is a rapid-fire review from big tech’s healthcare activities in 2020-
- Announced enhanced performance of DeepMind in identifying breast cancer from mammograms-reduction in both false positives and false negatives
- Received US FDA 510(K) clearance for Study Watch wearable’s arrhythmia detection for use in healthcare facilities (and not for consumer purchase) (Verily)
- Partnering with Mass General and clinical trial platform ProofPilot to create a new research platform for academic research
- Launch of Google Cloud Healthcare API, platform standardizing data exchange between Google Cloud-based digital tools
- Launched Coefficient Insurance Company, a tech-enabled payer targeted at self-funded employers (Verily)
- Deal with Amwell to be each other’s preferred cloud and telehealth providers ($100mil)
- Deal with US Dept of Defense for a prototype AI Pathology System to improve cancer diagnosis accuracy by providing centers with augmented reality microscopes that overlay the AI system (Google Cloud)
- Releases research on mapping biomarkers for depression (Moon shot subsidiary X)
- Launched two AI tools designed to help healthcare and life science organizations scan and analyze large volumes of unstructured text — the Healthcare Natural Language API and AutoML Entity Extraction for Healthcare
- Launched Healthcare Interoperability Readiness Program, a consultancy offering designed to help healthcare organizations as they gear up for Cures Act compliance in 2022
- DeepMind’s AlphaFold cracked a decade-old protein folding challenge
- Partnership with HHS’ Agency for Healthcare Research and Quality (AHRQ)for an online tool that helps patients build a visit plan for their upcoming healthcare visits (Google Cloud)
- Launched Google Health Studies platform, to streamline research study recruitment
- Launched Amazon Care, a virtual primary care offering for its Seattle-based employees
- Entry into wearables with Amazon Halo, on-wrist health tracker with accompanying app (activity and sleep tracking, body fat % measurement, and tone of voice)
- Alexa’s Care Hub, helping U.S. users check in on an aging family member from afar
- Start of Amazoon Pharmacy, that lets customers order and manage their prescription medications online and get them delivered at home
- Launch of HealthLake, a new cloud storage and analysis service available to healthcare and life sciences organizations
- Closure of Haven!
- Partnered with CDC with COVID-19 information dissemination, symptom tracker and trace-and-track activities
- Launched Heartline study, in partnership with Johnson & Johnson and Evidation Health, to measure the impact of Apple’s devices and a study-specific engagement program on cardiovascular outcomes among the Medicare population.
- Partnership with University of California, Los Angeles for a three-year investigation on depression and anxiety, using h Apple devices, such as an iPhone or Apple Watch, and a Beddit sleep monitor, as well as a downloaded research app to collect behavior data from participants.
- Launched several other studies including Digital Asthma study (with Anthem, CareEvolution and the University of California, Irvine), managing heart failure using metrics such as blood oxygen (with the University Health Network and the University of Toronto) and a study on how heart rate and blood oxygen can signal an early onset of respiratory conditions like the flu and COVID-19 (with the Seattle Flu Study and the University of Washington School of Medicine)
- Launched new fitness and health capabilities in Apple Watch (including sleep tracking, blood oxygen and so on!).
Healthcare and Negative thinking
I love reading Tim Harford because of his contrarian thinking. In this blogpost about ‘a way of thinking which anticipates that things sometimes go wrong and then plans accordingly’. He vouches for the value of thinking which anticipates that things sometimes go wrong and then plans accordingly.
Case in point-UK’s test and trace system for Covid-19- 15,841 positive test results had neither been published nor passed on to contact tracers in October 2020, as announced by Public Health England. Here’s how Tim see the case and provides his thinking-
The proximate cause of the problem was reported to be the use of an outdated file format in an Excel spreadsheet. Excel is flexible and any idiot can use it but it is not the right tool for this sort of job. It could fail in several disastrous ways; in this case, the spreadsheet simply ran out of rows to store the data. But the deeper cause seems to be that nobody with relevant expertise had been invited to consider the failure modes of the system.
What if we get hacked? What if someone pastes the wrong formula into the spreadsheet? What if we run out of numbers?
I totally resonate with this such thinking, especially in healthcare settings. While there is a lot of emphasis, by leaders and organizations, about ‘out-of-the box’ thinking and optimism, sometimes it makes sense also to think about critical weak links in our plans and try to build appropriate safeguards so that when the worst-case scenario actually happens, we are already prepared to face it!
Tidbits
- Healthcare Chart of the Fortnight: According to data from Bloomberg, the latest global vaccination rate is 4,253,851 doses per day, on average. At this rate, it will take 7.4 years to cover 75% of the population with a two-dose vaccine.
- Healthcare Tweet of the Fortnight: this tweet is intriguing! ‘The more prepared you were for the pandemic (according to JHU experts in October 2019), the more deaths you recorded from covid-19. If experts were right this line should have been *downward* sloping big time.’
