AI opens up countless new possibilities for businesses and individuals. However, AI is yet to show its full potential in the healthcare sector. In this article we will explore AI's potential to transform healthcare now and in the future.
How does AI already deliver in healthcare
To date, AI has seen patchy success in the broad health sector. There are three notable success stories though. These are protein folding, drug discovery, and synthetic data.
Protein folding
For the first time in history we now have a complete map of the human genome. The full instruction book to build a human being. This will open up a whole array of new treatments for chronic conditions and diseases. However, there are many steps between knowing the genome and understanding how it will express itself. One of the most important is a process known as protein folding. A given gene simply determines which amino acids are joined in which order to make a specific protein. The actual function of that protein is determined by the shape it folds into as it is formed. If you can predict that shape, you can understand how it interacts with viruses or molecules.
This is a hugely complex problem because of the trillions of possible answers. AlphaFold from DeepMind is able to tackle this problem thanks to AI’s unrivaled ability to crunch numbers. At the end of 2020, the AlphaFold team announced that they were able to correctly predict the structures of over 365,000 proteins produced in the human body and other well-studied organisms. This is bound to have a significant research impact as the graph below shows.
Drug discovery
Drug discovery is typically a mixture of trial and error, luck, and inspiration. New drug discoveries are made in many ways. For instance, adapting existing drugs, or using existing drugs to treat different conditions. Other times, you identify naturally-occurring bioactive chemicals and explore how they affect the body. Or come up with a target molecular structure and then synthesize a chemical of the correct shape. AI can help in various ways:
- Better mining of -omics data (e.g. genomics, lipidomics, and proteomics)
- Understanding protein structures (see protein folding above)
- Modeling how proteins interact and bind with other molecules
- Predicting and modeling how molecules will behave within the human body
- Analyzing scientific papers to identify chemicals with desirable characteristics.
- Identifying existing drugs that may help a different disease or condition.
Synthetic data
Clinical trials are a vital part of the clinical research process. However, they usually take many years to complete (unless there is an exceptional event like a global pandemic). Researchers don’t want to recruit a huge number of trial participants in the early phases, as it is expensive and time consuming. However, the only way to get access to the required medical data is to do precisely that. Step forward synthetic data.
Synthetic data uses an approach called conditional-GAN. GAN stands for generative-adversarial network. In conditional-GAN, you need a generator, which will try to create fake data, and a discriminator, which compares it adversarially against the real data. Both the generator and discriminator are given the correct labels for the data (unlike in a classical GAN system). The system then repeatedly updates its generator model until the resulting fake data is statistically indistinguishable from the real data. This gives you extremely realistic data for your research without revealing any personal data. That means you can use this synthetic data in any clinical trial without getting explicit consent for each one.
The riddle of the missing data
Synthetic data leads us on to the next issue with the current healthcare system. Data. Without data you won’t have good AI. Yet using health data for AI is hard: it is patchy, siloed, and receives strong legal protections.
Patchy or incomplete data
A typical AI project will take a set of labeled data and use this to forecast some missing label. A good healthcare example might be trying to predict when a patient can be safely discharged from hospital based on their clinical observations. This helps the hospital manage bed occupancy and helps patients to return home faster. But health data is seldom complete or continuous. Nurses only record patient observations periodically. Patients routinely use over-the-counter medications that aren’t listed on their notes. Treatments you receive away from home may not be added to your medical record. As a result, it becomes much harder to create the usual types of AI model.
Health data lives in siloes
Electronic health records (EHR) have existed within the US and many European countries for a couple of decades. However, there is no common standard, the data is often not centralized, and even the data subject has only limited control over it. HIPAA was designed to partially address this issue in the US. But only in the context of helping people to transfer between health insurers. Even where there is a central EHR system, not all data will be correctly added to a patient’s own record. And getting access to that data is a legal headache as discussed below. For AI to work, you need all that health data to be in one central location so you can process it and use it to train your AI model.
Health data receives strong legal protections
Laws like HIPAA and GDPR grant strong protection to health data. It is one of the most sensitive categories of personal data. This means there are significant barriers to accessing and using this data for research. Typically, a patient must give clear and explicit consent for the data to be used in that way. Often, that consent has to be specific to each piece of research. But certain groups in society are more likely to give consent for this than others. As a result, the available data has real potential for bias.
A healthcare revolution waiting to happen
So, what can AI deliver if you could solve all these problems? What new healthcare revolutions might arise? Let’s look at three potential ideas.
The smart hospital
One current field of research is exploring the so-called smart hospital. Here, AI would be used to improve patient care in numerous ways. These range from improving the flow of patients through the hospital to delivering personalized care pathways. This would have a particularly strong impact on patients needing critical care or with complex care needs.
Holistic healthcare
Millions of us now wear smart watches to monitor our fitness and wellbeing. We track everything from pulse and oxygen saturation to calories burned and quality of sleep. Just imagine if that data could be used to improve and personalize our healthcare. Holistic healthcare envisages combining fitness and wellness data with medical data to deliver complete healthcare.
Smart home care
Globally, the average age of the population is increasing. Improved healthcare has led to people living far longer, but often with much greater (and ongoing) care needs. AI can be used here in several ways. This includes using smart home devices to detect falls or other medical emergencies. AI can also power chatbots to provide support for patients with dementia, Parkinson’s, or similar degenerative diseases.
How we can solve this puzzle
No one company or institution can solve this puzzle to deliver better AI outcomes in healthcare. Rather, it takes a mix of new technology, government intervention, and changing attitudes.
Data sharing
Some countries have taken explicit steps to enable data sharing. For instance, in Finland all health and social care records are transferred to a central system, FinnData. Researchers can then apply for permission to use the anonymized and processed data for their research project.
Specialized health data platforms
AI relies heavily on data. As a result, most AI ecosystems like SAIBRE provide a data platform. This allows you to perform the necessary data engineering to prepare the data for AI. Specific health data platforms are now being built in order to provide the enhanced security and privacy requirements mandated by HIPAA and GDPR.
Novel approaches to AI
The unusual nature of health data calls for some novel techniques. For instance, one Dutch firm, Pacmed, has developed models to help predict when a patient can be discharged from ICU. As this article explains, they had to use new techniques to extract usable features from the complex medical observations.
Conclusions
AI promises an amazing revolution in healthcare. It will usher in new gene therapies, custom drugs, and novel treatment pathways. However, these may take many years to emerge. We need to start treating our health data as the valuable resource it is and do everything we can to help share it for research purposes. This needs action by governments, hospitals, health insurers, and others. But it also needs each and every one of us to buy into this vision of a better future.