From bench to bedside - small issues which compound to bigger failures
From bench to bedside - small issues which compound to bigger failures
Bringing solutions from the laboratory to the hospital is often filled with small issues which add up to bigger failures. Sometimes, the failure is not bringing the solution in the first place.
Progressive and good healthcare is all about bringing tools developed in the lab to patients in dire need of good care. However, being able to understand which tools to bring to the bedside, figuring out the challenges, and creating a solution which works for all involved is a daunting project. After all, in theory, theory and practice are the same. In practice, not so much.
A marked challenge faced by tools coming out into the field is being able to recreate the setup in which the tool was designed to be used. A successful example of surmounting this is the ultraportable chest X-ray machine. The concept is simple. Miniaturise an X-ray machine so that it can be carried out into the field and be used for X-raying a patient at their doorstep. Simple and easy. The advantages are obvious. If the patient is unable to come to the X-ray machine, the X-ray machine can go to the patient.
Unfortunately, X-ray machines are, by law, required to be installed in AERB1-approved rooms because of the emitted radiation. The solution? Create an X-ray machine which emits less radiation. Ultraportable machines emit a lot less radiation than conventional X-ray machines.
Once you have an X-ray machine out in the field, the next challenge is interpreting the X-ray. Typically, once an X-ray is printed, it is taken to a doctor, typically a radiologist, for their opinion. The radiologist's interpretation of an X-ray is then looked at by another specialist or a general physician.
As it turns out, good radiologists are not cheap! Radiologists are some of the best earning doctors in the country, and probably around the world. The solution? Get an AI to interpret the results. X-rays performed in community settings will, ideally, skew towards being normal and not show disease progression. If the AI flags an X-ray it can either be reviewed further or the patient can be asked to go to a facility for a full check-up.
Cough Sound AI
Another area in which people have been trying to get technology to the patient has been in the realm of cough sound AI. The concept, again, is fairly simple. If a patient comes to you with a cough, use a trained AI to match the sound of the patient's cough with a probable disease.
There has been remarkable progress in creating these AI models. Reported sensitivities and specificities of around 90% are not uncommon. Unfortunately, field trials of these AIs have not yet been as successful as those of ultraportable chest X-rays.
Why is it that these models are unable to recreate their success in the lab out in the field? Is there anything special about hospitals which makes them hostile to cough sound AI? Turns out there is. The real world is a noisy place. Getting someone to cough in a way which the AI can capture, remove the noise, and give an accurate response turns out to be a major problem.
This would probably not come as a surprise to a seasoned hospital worker. Patients are remarkably difficult to coax into doing the most straightforward of things. Getting them to cough in a way that an AI trained in the lab is able to accurately capture the sound is probably very difficult.
Unfortunately, it's not just the patient who is at fault here. The levels of background noise varies from region to region. Regardless of whether the area of operation is a health facility, a community-based setup, or a door-to-door setup, the programmer who created the AI would not have thought of every situation in which that AI would end up.
This is a place where advanced noise reduction algorithms would need to be created. While these have been created for transcribing, say, phone calls or audio messages, the challenge of accurately removing background noise from a patient's cough is infinitely greater. Not only is the sound of coughs uniform (you or I probably have very different styles of coughing), the sound of the same person coughing during a bout of pneumonia or active tuberculosis may be very different. That is, after all, what the entire idea of a cough sound AI is predicated upon.
Radiologists in the building - or not?
A cough sound AI demonstrates the failure of AI to perform out of the lab. However, there are times when the competence of AI echoes in its absence.
X-rays are a source of information about patients of all stripes, and are often read by all kinds of doctors, not just radiologists. Some of those doctors are still learning to read X-rays. I am talking, of course, of doctors early on in their careers, not those who have had some experience. It is usually these junior doctors who read the largest volume of X-rays in a hospital.
The rate at which AI is progressing may make it a better steward of X-rays than most non-radiologists. While it is too early to call AI a replacement for radiologists, it may serve as a useful aid to harried junior residents concentrating on ensuring proper diagnosis and treatment for their patients. While purists may bristle at letting AI or computers handle something clearly within a doctor's domain, it may end up being better in the long run for doctors to concentrate on tasks where machines are absolutely unable to match up to them.
Atomic Energy Regulatory Board, India