How artificial intelligence is changing healthcare [Q&A]
Artificial intelligence is having an impact on more and more areas of our lives. One of the areas where it has most potential is in healthcare, allowing professionals to make faster and better decisions, and applying innovative problem solving.
We spoke to Eric Landau, founder and CEO of Encord, to find out more about the benefits and challenges of using AI in this sector.
BN: What are the main applications of AI in healthcare today?
EL: Most headlines surrounding AI focus on how it will replace or replicate the judgment of doctors. A more sensible goal is not to replace doctors, but rather to augment them. AI will mainly serve to expand the reach of doctors to a much wider set of data and patients.
Within medical AI, there are obvious applications like diagnostics, however, there are more niche applications like the automation of patient record creation. These mundane tasks take up a lot of doctors' time but AI helps reduce this by providing them with the right tools so that they can spend more time with patients and doing diagnostic work.
We are also seeing many AI applications to facilitate nursing. Many countries around the world are suffering from nursing shortages where ultimately, nurses are stretched for time. Monotonous tasks like patient monitoring are time-consuming so the implementation of AI-powered monitoring tools help them spend more time treating the patients rather than just observing them.
BN: What impact does machine learning and medical AI have on medical professionals' day-to-day responsibilities?
EL: AI can act as a clinician's second pair of eyes by providing automated second opinions or observations. Doctors, especially surgeons, can be overworked and are ultimately human. Having an additional modality to assist them helps them to double-check imagery or video feeds. The objective is for AI to augment medical professionals and focus their time in a more effective way.
Medical AI can also assist medical professionals with the more tedious day-to-day tasks such as filling out documentation. Automated documentation can be manifested in different ways from documenting the steps of a procedure like creating timestamps during a colonoscopy to filling out paperwork. Doctors then focus on their primary activity rather than documentation, enabling them to work in a focused way.
BN: What challenges are preventing medical AI from reaching its full potential?
EL: The main challenge for medical AI is the data labeling problem. Data is the biggest bottleneck but there are several sub-components to it.
The first is the sourcing of data, machine learning (ML) teams need to have enough of a specific type of data to build an AI model for a certain disease or certain function.
Secondly, there's the actual annotation of the data which addresses the need to label, structure and cleanse the data. These steps are difficult to carry out and made even more challenging by privacy and regulatory requirements around the data. Medical data is not as easy to process as other applications, such as autonomous vehicles. More thought and care has to be put into any steps that you go through with regard to the data layer.
Lastly, there's making sure that the data is properly balanced to the real-world application that it will be released on. Solving the data layer will unlock functionality and higher performance in these models than we have seen so far.
BN: What are some of the data problems that machine learning teams encounter when developing medical models?
EL: A problem ML teams encounter is the constraints of domain experts. Sometimes, when training a model on a very specific or niche use case, it can require outsourcing it to or using people that aren’t as highly qualified as a medical professional so the probability of error is greater on labels.
There's also a lot of disagreement between medical professionals on the best way to label things. As the expertise threshold goes up, the process towards generating ground truth becomes more difficult, ambiguous and uncertain. These are problems that you don't have when you’re training a model to detect human faces or animals. It's an additional wrinkle to the medical field.
BN: What's a data-centric approach to AI and what are its benefits?
EL: A data-centric approach to AI means that when you're building an AI application, the first place to look at is the data, not the model. Data quality is imperative to improving models, so ML teams must figure out what is the best dataset to feed their model? Are they labeling things correctly? Are there errors in the labels? Is it specified correctly? Are there model failure modes within specific segments in slices of the data?
Thinking about the data first is the best way to actually get high-performing models to work in the real world.
BN: What is the end goal for AI?
EL: There are three main end goals for Medical AI. The first is essentially giving medical professionals superpowers. The best applications give medical professionals time back and provide them with additional tools that can expand their judgment and capability. Farmers used to break and till fields manually but they now use machinery like tractors to help them become more efficient at scale. Medical AI's aim must be to provide medical professionals with the machinery to give them superpowers.
The second end goal is to expand the scope of medical capabilities and help democratize healthcare globally. This will give people in developing countries access to a higher standard of medical facilities and treatments.
Lastly, AI must get to the point where it helps medical professionals focus on preventative care versus reactive care. Unfortunately, there are instances where people only get treatment when it's too late, and you find it only after your body figures it out. AI must enable clinicians to give diagnoses sooner when diseases are easier to treat.