The AI Landscape in Health

AI is making its way into every facet of the healthcare and life sciences from research and drug discovery to diagnostics and disease management. The use of AI in these areas dramatically increased in 2019. In Optum’s survey, 62% of who report having implemented an AI strategy—in 2018, only 33% reported implementation. According to same survey of healthcare executives, fifty percent of companies that choose to invest in AI are seeing return on investment in cost savings. Hospitals and health plans are expected to a payback period of three years on AI investment while biotech executives anticipate it taking five years or longer.

In the healthcare and life sciences landscape, AI is no longer a differentiater. It is an investment that many companies are taking to reap benefits (revenue, customers, savings) in a relatively short period of time. Also, given that AI tech is so widely available, if you can find the very in-demand data scientists required to build products, the use case and business model matters more than the AI algorithm.

AI in Health Research

There is a strong need to accelerate research to better understand mechanisms of disease, establish biomarkers, and develop disease models. To do so, scientists are challenged with keeping up with the tsunami of biomedical literature that flows in daily. Anecdotes abound with wasted time and research dollars repeating work that has already been done. Scientists are generating unprecedented volumes of data, and simplified access to the biological context found in scientific literature is needed. We need to empower scientists with AI tools that quickly assimilate, integrate, and present new (and old) knowledge.

Efforts go from general to vary focused use cases. Quertle is an improved search engine meant to simplify the discovery process by integrating, organizing, and presenting biomedical literature. Genomenon’s Mastermind improves variant interpretation by searching millions of scientific articles. Ingenuity Pathway Analysis attempts to predict downstream effects and identify new targets and biomarkers. Each offers a partial solution to very large problem.

Natural language processing is at peak usefulness. It feels like everyday a new language model is pushed out to the world. OpenAI released their coveted GPT-2 model in the last few weeks. This model was trained on 40 terabytes of text and can answer questions, translate, and write stories with a bit of prompting. Without specifically being trained to do so (though the accuracy is not great), GPT-2 will summarize given text when prompted with “TL;DR;”.

Summarization will be a key use case in the biomedical field. A number of efforts have focused on the use of NLP to mine PubMed,, and patents. The focus has been to find the right data and put in front of the user. The best insights to drive health research will come from the unification of many sources and types of data, including published literature, genomics, gene expression, microbiome, electronic health records, social determinants, social media, wearables, and more. Natural language processing will be an essential component of this strategy to dynamically model evolving entities and relationships among many disparate forms of data.

AI in Drug Discovery

AI is supporting critical use cases to find novel drug candidates, cut costs, de-risk the drug discovery pipeline and accelerate time to market. Drug discovery and development is a top focus of AI in life sciences. This is a complex area of AI and there are 100s of startups focusing in this area, along with significant commitments by established pharma leaders. The efforts span everything from using AI to to design novel drug candidates, identify opportunities for drug repurposing,and in silico screening for efficacy and toxicity much faster than traditional R&D. The use of AI to stratify patients to recruit and optimize clinical trials and analyze real world evidence will also change the speed and success rate of drug development.

Goole company, Calico Labs is using AI amongst other techniques to tackle aging. One of their more unusal project is studying the age defying naked mole rat that everyone should know about if only because evolution is so weird (here and here). Former Calico Labs chief computing officer, Daphne Koller, launched a startup of her own, insitro, to lower the cost of drug development. In the same spirit, Atomwise provides ab initio drug screenings that boast to deliver accuracy comparable to wet lab experiments.

Although decades away, quantum computing will transform the search for new drug candidates. It will enable us to design new drugs that aren’t possible today, and will accelerate and de-risk the drug pipeline by rapidly solving for computationally intensive molecular interactions for toxicity and efficacy. In the case of precision medicine, it will make some of these complex algorithms to compute in real time, possibly while the patient is still in the office.

Cloud Platforms for AI. Sometimes it feels like everyone has an algorithm, proof of concept, and a partnership to repurpose a drug or find that key disease target. The drug process can require lengthy partnerships. The true power will be providing easy to use tools that let everyone do their own target and drug discovery. A number of companies are shifting to that paradigm. UK start-up, Lifebit, provide an enterprise platform to scale analysis within their own secure data environments. San Francisco-based BlackthornRX has a neuroscience focused cloud-based computational platform for multimodal data collection, integration and large scale analysis.

AI in Diagnostics and Disease Management

AI has the potential to accelerate and improve diagnostic accuracy for many disease processes. Deep learning has been used to create applications to identify the patterns of disease in images andscans. Applications have been developed based on diagnosing issues in retinal images, tissue slides, CT images, MRI scans, x-rays, and facial features. The primary challenge to build these applications is the assembly of large data sets to train and validate models with images or scans that represent both disease and non disease. Most of these applications have been targeted to the medical community, for use in hospitals and clinics.

A number of new applications are targeted directly to the patient. These deep learning models are being trained on speech, video, voice, social media activity, wearables sensor data, eye tracking, even keyboard usage patterns. Applications are already in use to detect abnormal heart rhythms, alerting of concerns that require medical intervention. The hope for the future is to turn our smartphones into personal medical assistants, with passive, unobtrusive monitoring, and alerts of concerning patterns. To enable this next generation functionality, we will need models that can learn from individual user patterns, synchronizing signals from multiple sources, to identify abnormalities in real time. To do so, the ability to train deep learning models on a smartphone will become common.

The challenges we see in introducing AI applications in high consequence environments, such as clinical decision support, are in user adoption. For a physician to trust and act on recommendations coming from an AI application, they need to know that the application has been validated prospectively via clinical trial, and they must have some explanation for why a prediction is generated for a specific patient. The FDA has defined the clinical trial requirements for AI based clinical decision support applications, considered Software as a Medical Device. To address the user adoption challenge, we need to make predictions more interpretable and understandable. Explainable AI is a necessary focus to drive adoption among physicians.

Cartoon from Loren Fishman

Bottom Line

The AI in healthcare is expected to grow from $2 billion (2018) to $36 billion by 2025. AI has started to bring significant value to the life sciences and health care industries. The face of healthcare will completely change in the next 5 years because of AI and automation. Continued investments in natural language processing, quantum computing, deep learning, and explainable AI are needed to realize the full potential of AI to maximize human health and reduce human suffering.

Written by

Angela Wilkins

I like science, machine learning, start-ups, venture capital, and technology.