In this article, we are going to cover some of the ethics, ‘good’ and ‘bad’ related to the use of AI in healthcare as well as the challenges current infrastructures present to the deployment of these AI technologies.
As we’ve seen in part I, AI has a great many applications in healthcare. Its use cases in radiology, oncology and pharmacogenomics, amongst other medical fields, signal the transformative weight it carries. And its ability to perform tasks with superhuman precision, speed and efficiency (in a wide variety of iterations) legitimises the anxiety surrounding the future of healthcare practitioners. That being said, I argue that the technologies are a supplement rather than a substitute to medical staff’s work.
WHY DOCTORS WON’T DISAPPEAR
The set of observable characteristics of an individual, resulting from her genotype interacting with the environment, are called phenotypes. Phenotypes are measurable attributes; they can be straightforward visible characteristics like height and eye color, but also overall health, disease history, and even behavior and general disposition. Teaching AI to make an exact diagnosis based on phenotypes is very complicated. Think back to a WebMD search of your cold symptoms and the sinister list of doomsday conditions it toggled back. That is because phenotypes are on a spectrum. They can have many confounding variables and come in millions of different combinations with overlapping characteristics between them.
And so, when presented with a case (unlike algorithms), doctors use intuition and real world experience to integrate all the data and say that one person has a stroke and another rheumatoid arthritis. But how do you teach a machine to do that? We’ve seen that using deep learning to synthesize large amounts of data lends great results. That is however not enough to make a perfect diagnosis. Those kinds of almost more intuitive data; the elements of the phenotype that aren’t captured in the EHR, are very important in making decisions for treatment. The breadth and contextual knowledge of human pathologists (and all other physicians) to verify and rectify the AI’s diagnosis process and set the threshold to say whether someone has a disease or not, are essential and cannot be replaced. I believe that the same goes for nursing – whilst AI can take on the low-level skilled tasks nurses perform, it cannot detract from the intuitive checks, the human and humane interactions and the comprehensive caregiving they provide.
PREVENTING HUMAN ERRORS
AI has the potential to do away the negligence, inefficiency and the biased treatment that many overworked or irresponsible doctors inadvertently subject their patients to. A 2018 study (consult the study here) showed that a pathologist with unlimited time and not overwhelmed with data, has a 94% rate of sensitivity in finding the tumors in large amounts of slides. When you add time constraints on the pathologist, her sensitivity drops. And she will start overlooking where little metastases may be. Negligence and inefficiency are a matter of fact and cost patients dearly. That is unsurprising given the amount of information pathologists have to process. One of the slides in the experiment (which mimics real life situation) if digitized, is about 10 gigapixels – a literal needle in a haystack.
The second interesting finding in the study is that pathologists can find 73% of the cancers if she spends all her time looking for them with zero false positives per slide. A model trained to perform this task finds about 95% of the cancer lesions and has eight false positives per slide. So clearly, an ideal system is one that combines both forces, on the AI to detect the cancers and on the pathologist to look over the false positives.
Incorrect prescriptions of medication are also a serious problem. Whether it is because a doctor lacks training or makes the occasional error of choosing the wrong medication (sometimes as trivial as selecting the wrong option in a pull-down menu), patients end up severely harmed. AI can reduce these error rates by analysing the patient’s records, her history of drug prescriptions, and flagging anything that seems fishy. An Israeli startup, MedAware, has already developed this type of software, with promising results. Of the 747,985 patient records that were analyzed, 15,693 were flagged – that is an error rate of roughly 2%. And from a sample of 300 of these alerts, roughly 75% were validated by doctors, proving that the tool could potentially save the industry billions per year.
ADDRESSING PUBLIC HEALTH PROBLEMS
At the interstice between medicine and big data is ‘Public Health 2.0’. The term has become a bit nebulous today, but it initially referred to the use of big data to follow public health trends such as the spread of Ebola. If we pool geo-tagged social media posts that talk about ‘illness’ or ‘getting sick’, we can imagine being able to track the spread of a particularly bad strain such as H1N1. Given the funds invested in public health by governments worldwide, using this information may allow resources to be utilised more effectively to address disease outbreaks faster. Concretely, it would allow public health experts to spend less time localising and isolating disease. So far, private medical technology and the use of wearable devices has grown at a rapid pace, and consumers with buying power may be seeing health improvements. However, despite having more data, public health outcomes are still insufficiently addressed and remain a field with significant growth potential.
AI carries all this promise of revolutionising healthcare towards best practices across the board but we must remember that it is driven by historical data and current practices. As we’ve seen in previous blogposts, if deep learning is extrapolating from people that are biased, it can exacerbate these biases.
Here’s a thought experiment to illustrate the above: A drug is estimated to work for about 20% of people. We have two groups. In group A, 75 people receive the drug and in group B only 3 can afford to do so. For the group B, there’s greater than a 50% chance it doesn’t work on any of them. So the model might learn that the drug doesn’t work on that group of people. This is an instance of bias from a small sample. The algorithm might do worse than not recommending it, it might say it’s a bad recommendation. A real world example of this is for those groups (usually minorities) that do not have medical insurance (often the gateway to receive the treatment). In the US, the European population represents about 80% of the genetic tests that have been performed and indexed for researchers to work with, the genetically African group only represents 2% of the genetic tests that are available to researchers and the genetically Native American represent less than 1%.
Another real world instance of bias in the training population of the algorithm is the government of New Zealand’s 2016 computer vision algorithm blunder.
They developed an algorithm to recognise people’s faces to determine whether their pictures were of adequate quality for passport photos. When Richard Lee, a man of Asian background, wanted to submit his passport photo on the government portal, the algorithm rejected it on the grounds that his eyes “were closed”.
PATIENT DATA PRIVACY
Another major ethical challenge facing healthcare AI is patient data privacy. In 2016, The Royal Free London NHS Foundation Trust received a slap on the wrist from the UK Information Commissioner’s Office for its mishandling of sensitive NHS data after it was found that it had supplied AI firm DeepMind with data on 1.6 million British patients without their consent. Despite the fact that this wasn’t an inherent fault of AI, it did a lot to erode patients’ trust in the software. It seems obvious that patients need to be able to share as much detail as necessary with their clinicians without the anxiety that their data will wind up in an algorithm for someone else to look at, or otherwise be used inappropriately without their consent. But many commentators fear concerns about data privacy could set back vital research projects. Healthcare privacy regulations often don’t cover the way tech companies wind up using health data. The US Health Insurance Portability and Accountability Act (HIPAA), for example, only protects health data when it comes from organisations which provide healthcare services, like insurance companies and hospitals.
TIDBITS TO THINK ABOUT
AI will inevitably exacerbate the global healthcare inequalities between the haves and the have nots. A single da Vinci Surgical Robot System costs $2 million, research on deep learning algorithms can cost a hundred times that amount and take decades to complete, and embedding AI systems in clinical facilities is also a huge financial commitment most countries cannot undertake.
Other ethical questions are on medical liability, who is liable when apps like the “skin cancer apps” (mentioned in part I) make a mistake? Finally, what happens if someone comes up with a generative adversarial network (an AI technique to generate new data with the same statistics as the training set) to generate fake cancer diagnoses?
..AND THE CHALLENGES
At present, the greatest barriers to the adoption of big data solutions in healthcare is the lack of a unified view of patient data. Patient data is warehoused in a way that is unique to a healthcare organisation and the medical bodies it uses to benchmark with. We still do not have consistent laws/rules around the interrogation of anonymised patient data and it may be a while until we are able to wrangle data sets with identifiable data. Actual clinic deployment of these models and algorithms also requires large infrastructures and staff training, whilst optimising and renewing existing infrastructures is incredibly costly. And so, until we can change infrastructures and incorporate AI models in our current workflows, there will always be barriers to widespread AI in many areas of medicine.
The process of integrating information across systems (EHRs, Claims Systems, etc.) is anything but straightforward for the reasons mentioned above. We know that bringing new technologies and new algorithms into healthcare is critical, and we can only do so by harnessing the intellectual capabilities and the data from as many key stakeholders as possible. Interoperability is that concept that refers to the ability of health information systems to work together within and across organizational boundaries in order to advance the effective delivery of healthcare for individuals and communities. While the health IT sector has made a big impact in terms of facilitating cooperation between electronic health records (EHRs), it still has a way to go with regards to simplifying healthcare interoperability. Imagine how efficient treating patients could be in a world where academic health science centers, biotech companies, insurance companies, pharmaceutical companies, hospitals, and tech giants collaborated in a way that is safe and effective.
Fortunately, research organisations are paving the way by partnering together to share research data and insights. Verily Life Sciences, for one, collaborates with Google as well as other partners like Nikon and Optus to bring cutting edge technology to the visual imagery recognition space.
Written by Nada Fouad