Revision as of 14:54, 8 May 2010

Humanity's medical resources are overworked and completely overwhelmed, as five minutes conversation with any doctor will suffice to verify. Doctors often work 36 hour shifts, there are not enough beds to hold patients, annual insurance premiums for doctors are in the tens of thousands. This describes the situation in a developed country, the sort that has one doctor per 300 people; the situation in a country like Malawi, which has one doctor per 50,000 people ^[1], is worse beyond description.

The number of medical studies published is growing exponentially and no doctor could be expected to keep up with even a fraction of research. Unbelievably, it takes 10-15 years for a drug to go from development to availability. ^[2][3] (some authorities say 20 years ^[4]). Our scientific understanding of disease has grown explosively in the past decade, but there has been no change in the rate of adoption of this data. This is a problem.

So what is the alternative? Imagine for a moment that every time a patient walks into a clinic, modern information technologies whirr into action and pull from a distributed global network the patient's medical record, family history and perhaps even genetic code. (This is all securely stored, of course, and cannot be accessed without the patient's consent.) The patient descibes his symptoms to the doctor, who inputs them into a handheld computer. The computer processes the symptoms and cross-references them with a database of known diseases. It compares this patient with all similarly-presenting cases – no matter where in the world those other cases happened. It uses all this information to create a list of possible diagnoses, then uses optimized artificial intelligence algorithms to calculate the likelihood of each one, based on the patient's history, race, age, sex and so forth. The doctor's computer displays the list of possible conditions and possible courses of action, including treatments or further testing. If a scan is done, say an X-ray, the computer can identify patterns in it and use this to inform its diagnosis. Such a system never overlooks a piece of data such as a drug interaction or allergy, and intelligently updates itself based on the entirety of mankind's medical knowledge.

The technology to do this exists now. All that is required is a large database of medical data, standardized electronic medical records (EMRs) and – to analyze all this data - an artificial intelligence program known as a Clinical Decision Support System (CDSS). Such a system would also be of immense benefit to people in places with no medical professionals, or for low-level treatments that did not require a visit to a hospital.

But can modern artificial intelligence really diagnose disease as effectively as human doctors? The surprising answer is that even mid-1970s artificial intelligence could. In the 1970s, a CDSS called MYCIN proved itself able to perform medical diagnosis as well as, or better than, human doctors. Several things have changed since then. Firstly, the raw computing power available has increased by a factor of well over a billion. Secondly, programmers are able to make smarter AI which is better able to recognize patterns and learn from experience. AI can now autonomously make diagnoses from an ECG, for example ^[5]. Thirdly, we have much more medical data. This data cannot possibly be remembered and analyzed by a human doctor, but can be accessed by a computer. Fourth, the amount of data we can pull out of a human body with various scanning modalities has increased a billionfold. Fifth, the Internet raises the possibility of clinical data being shared on a massive, global scale, allowing for extremely quick learning and limitlessly broad education for the AI.

So why hasn't such a system come into place? Of 70 commercial attempts to bring CDSS into hospitals, none have really caught on, because the task of generating a database of medical knowledge large enough to be really useful is outside the capacity of any one company ^[6]. Any database of clinical information becomes drastically limited as soon as it is made private. The only way to create such a system is as free and open-source software — allow anyone to contribute to the database, get college students, government departments and data entry companies to pile the results of medical research into the database, get software programmers to improve the AI, get clinicians to put the system into practise and record clinical outcomes, and you soon have a database approaching the ideal descibed above. It is achievable within a few years, but only with global open collaboration.

EgaDSS is just such an open-source CDSS and (unlike commercial CDSS) it integrates with electronic medical records.

Another problem holding back the adoption of medical AI is that many have had difficult interfaces ^[7]. This is exactly the sort of end-user problem that free and open-source software is good at remedying.

See also

• http://www.phoenix.tc-ieee.org/016_Clinical_Care_Support_System/Open_CIG_9_19_sanitized.ppt - A slideshow showing the need for open-source CDSS AI systems
• AI in medicine