AI is here today. Not with the broad, deep, and subtle intelligence that we associate with human intelligence, but with the ability to deliver billions of dollars’ worth of value, and real aid to clinicians in the field in medicine.
Global management consultancy McKinsey & Company estimates the potential annual savings of AI in health care to be 0.7 percent of gross domestic product, or $300 billion US dollars in the United States and £3.3 billion British pounds in the United Kingdom. According to ABI Research, AI will save the health care sector $52 billion US dollars in 2021, with $21 billion US dollars in savings in North America alone.
It’s not just better, faster, cheaper — it’s different. AI allows us to do things that humans just couldn’t do before, like consider your entire genomic profile before making a recommendation.
In 2017, Google’s Verily Life Sciences released DeepVariant, a deep convolutional neural network open-source AI tool. Single nucleotide polymorphisms (SNPs) are the most common type of genetic variation and are used to predict a person’s disease-risk and sensitivity to environmental influences on gene expression such as food, drugs, or toxins. DeepVariant was able to identify SNPs with 99.9587 percent accuracy, garnering the “Highest SNP Performance” award in the 2016 PrecisionFDA Truth Challenge.
A number of ground-breaking AI startups have entered healthcare. For example, in 2017 Cloud DX won the XPRIZE Foundation‘s “Bold Epic Innovator” award. Cloud DX is applying machine learning with a large sound dataset to identify respiratory infections and diseases based on the sound of a patient’s cough.
AI is being deployed in predictive analytics in patient monitoring devices, imaging and diagnostics, drug discovery, and oncology. Recent FDA approvals for AI-based health care solutions include Imagen’s OsteoDetect for X-ray image analysis; IDx-DR for diabetic retinopathy detection; and Viz.AI Contact for early stroke detection.
In May of 2018, the Annals of Oncology published a landmark German study where a deep learning convolutional neural network (CNN) trained on 100,000 images outperformed an international group of 58 dermatologists from 17 countries in diagnosing malignant melanomas. Hurdles to overcome in AI in health care include access to large data sets for training deep learning algorithms, and curating electronic health records (EHR) that lack data labeling and are mostly unstructured text.
The greatest challenge is to convert the whole of healthcare to be data driven — AI is part of that transformation. The recent bloom in AI is largely due to advances in pattern recognition with deep learning algorithms. In the future, Jacobstein anticipates more convergence of different types of AI, such as model-based reasoning, predictive analytics, and simulation models, in order to go beyond basic pattern recognition.
AI for Oncology
Equity-funded startups in AI oncology include companies such as Freenome, Globavir Biosciences, CureMetrix, Notable Labs, Cyrcadia, Enlitic, Entopsis, Insilico Medicine, OncoraMedical, Pathway Genomics, Proscia, and SkinVision. Flatiron Health, a New York headquartered AI technology company focused on oncology backed by Google Ventures and other investors, was recently acquired by Swiss pharmaceutical Roche for $1.9 billion in 2018 for its massive curated datasets for decision support and electronic health record platform.
ABI Research estimates that the number of patient monitoring devices for training AI predictive analytics will grow to 3.1 million in 2021 with a CAGR of 176%. Venture capital funded startups in this space include Sense.ly, Sentrian, Babylon Health, and AiCure.
AI and Drug Discovery
A number of companies such as Calico, BenevolentAI, and Deep Genomics are seeking to apply machine-learning to big data sets for drug discovery and development. Additional startups in AI drug discovery include venture-capital funded twoXAR, Numerate, Atomwise, and Numedii.
Pattern-recognition, an area of recent advances in AI, is a natural fit in health care imaging and diagnosis. Radiology is an area with large, structured data sets that can be used to train deep learning algorithms. For example, in 2017 the National Institutes of Health (NIH) released over 100,000 chest x-ray images from more than 30,000 anonymized patients for academic and scientific research institutions.
At the Exponential Medicine conference, Jacobstein cited an example of a recent AI healthcare breakthrough in 2017, where a team of computer scientists at Stanford trained a deep convolutional neural network to classify skin lesions. Using a database of 129,450 clinical images, the deep learning algorithm had an accuracy on par with human dermatologists in detecting malignant carcinomas and melanomas.
[ Infographic courtesy of and copyright Publicis Health ]