Ever used the internet to look up a health concern? If so, you’re not on your own – more than 5% of search engine queries are health related. But you’ll also know that the hundreds and thousands of results you get can be overwhelming and scary. Is that sore toe a sprain, or something more serious like gout? Is your persistent dry cough a cold that won’t go away or could it be interstitial lung disease? It’s so common that they’ve even coined a new term for the phenomenon: cyberchondria.
Luckily, developments in artificial intelligence could make that experience less frightening and a lot more useful, for patients and doctors alike. Very soon we could build highly personalized health question-answering systems that will be able to handle much more than simple queries like “what is measles?”
To explain how this will be possible and what it will look like, let’s take a closer look at how today’s search engines would answer that pretty basic question. First, it would strip away “what is” and then it would do an easy search for an entity that best matches “measles”. In search engine talk, an entity is a person, place or thing. Every disease, symptom, chemical, treatment and physician has an associated entity. The relationship between these different entities – X drug contains X chemical, or X protein group always occurs in tumours of X cancer in X organs – is the scaffolding on which facts are based. So if your search engine has access to a fact base of known drug interactions, it could very easily answer something like “can I take Advil if I am using Simvastatin?”
The larger that fact base, the more detailed and personalized the answers can get. How about if I asked “is there somewhere near work where I can go for a knee check-up?” First, the search engine would have to check my personal fact bases to figure out which knee problem I’m referring to. It would also help if it looked in my medical history for any treatments that had in the past been suggested. Second it would have to establish which medical institutes deal with that type of knee problem, and whether the treatments they offer are suitable for me and my condition. Then there is still the question of insurance – does the medical institute accept my coverage – and of course proximity to my office.
But where it really gets interesting is in cases of uncertainty. Many relationships in a healthcare fact base will be probabilistic. People with fair skin have a 15% chance of developing rosacea at some point. A person suffering from pancreatitis has an increased chance of a yellow colour in the whites of their eyes. These facts can help answer questions such as “my feet are numb – what should I do?”
In medical areas with small entity sets and small fact bases, we’ve already had decades of success with machine diagnostics of underlying causes for symptoms. I expect to see question-answering systems trying (cautiously at first) to provide answers to these questions. Initially we will see little more than extensions of what search engines currently provide: a list of possible causes that have the given symptoms. But, as happened with maps and highway navigation, things will accelerate as increasing numbers of people start using these systems. The diagnosis systems will benefit from knowing personal information such as gender or previous conditions in reweighting the likely causes, and in some cases will benefit from full genomic information. The fact bases will include known genomic indicators for specific diseases, but more importantly relationships between personal indicators, treatments and outcomes. I am confident that by 2020, patients will be asking their search engines questions like “what should we try next for my tendonitis?”
And I’m not the only one confident of the progress we will make and the potential for such diagnostic systems to save hundreds and thousands of lives. That’s why we’re seeing big projects and collaboration in this field. Just last month, Carnegie Mellon University, the University of Pittsburgh and Pittsburgh’s largest hospital system, UPMC, announced an alliance that they think will transform healthcare through big data. The plans involve fielding a robust and measurable health question-answering system by the end of the decade. Healthcare software providers such as IBM and Cerner, along with major ecommerce and search engine companies, are also racing to become major suppliers of backend technology for personalized healthcare systems.
While I’m confident highly personalized healthcare systems, powered by artificial intelligence and big data, are just around the corner, there are still some unanswered questions. First, it’s not yet clear whether all this new information will be integrated into hospital information systems and the doctor-patient consultation process. If there is little integration then it could be that we simply see patients arriving at doctor’s offices with iPads full of suggested diagnoses and tests. If the two are integrated, we may see much more powerful answers to patient questions, and better-planned use of time during consultation with human doctors. The second question concerns privacy and data breaches. The system will only work if individuals disclose as much personal information as possible. But as we have seen with online shopping, when that data is hacked, the consequences can be grave. A breach of a patient’s medical history could be catastrophic.
If we manage to resolve some of these outstanding issues, I believe we have the power to revolutionize healthcare.
The Annual Meeting of the New Champions 2015 will take place in Dalian, China, from 9-11 September.
Author: Andrew Moore, Dean, School of Computer Science, Carnegie Mellon University
Image: A robot is attached to a patient’s arm as part of a rehabilitation program at Sant’Anna hospital in Crotone, south of Italy, December 16, 2014. REUTERS/Max Rossi