Inspirational work by Stanford researchers using Google’s TensorFlow to detect malignant skin lesions.
Fascinating news today that the app Cardiogram (iOS, WatchOS) has used machine learning algorithms to differentiate between regular cardiac rhythms and irregular rhythms such as atrial fibrillation (AF) from data collected by the Apple Watch. The study in which this finding was presented states the app picked up 97% of AF episodes. While real-world usage may give different results (e.g. if the Apple Watch wasn’t worn correctly), and consumers should be cautioned about false negatives and the risk in trusting any device absolutely – this is still very exciting news.
AF is a key concern in heart disease. The heart is made of four chambers, two atria and two ventricles. Blood normally flows from the atria to the ventricles and from there to the lungs and the rest of the body. For the atria to transfer the blood over to the ventricles, they need to pump well. In AF, the atria do not pump in a coordinated manner and, instead, “fibrillate”. This causes two key problems. First, the ventricles aren’t filled with blood well enough. This means that when the ventricles contract to send blood to the rest of the body, they are pumping out a lesser amount of blood. One of the organs that gets this reduced amount of blood is the brain – and this is why syncope (fainting) or pre-syncope (feeling as if you’re about to faint) can be associated with AF. The other, potentially more serious, issue with AF is that because the blood isn’t moving along from the atria well enough, it increases the likelihood of clots forming in the atria. These clots can be pumped by the heart into circulation and go on to block blood vessels, such as in the brain causing a stroke. There are other reasons why AF is something to take seriously, but this is the gist of it.
For those who are already in known AF, the cornerstone of management is rate control and anticoagulation medication. It would be interesting to see if the Cardiogram app not only recognizes when the patient has gone into AF, but also accurately notes the actual heart rate during the AF event. Atrial fibrillation where the rate is within accepted range, is a completely different beast to “fast AF” which can cause life-threatening hemodynamic compromise (critically inadequate amount of blood pumped to the body). In cases where rhythm control is also a treatment aim (e.g. when the patient experiences troublesome symptoms despite rate control), I wonder if the Cardiogram app could help give an idea of how often the rhythm is irregular. It would also be very helpful if AF detection by the app automatically fired off an alert to caregivers, particularly in the elderly who may not always be able to signal for help immediately.
Another use I can think of for AF detection is when a patient with known AF presents with syncope. In the absence of witnesses to the syncopal event, history of the presenting complaint can be tricky to elicit. This can make it difficult to determine if the syncope was due to AF (poor pumping by the heart), or a different reason. If the Apple Watch showed an AF event coinciding with the syncopal event (perhaps by measuring a sudden drop in altitude?), that could be pretty handy information for the clinicians. I can think of so many great ways to use the sensors on Apple Watch to help with serious medical conditions, but I’ll leave that for a separate blog post.
I feel Apple is well suited to push a wearable into serious healthcare territory, beyond mere step tracking. Building trust will be key but I think Apple could just pull it off.
James Whittaker of Microsoft gives a stimulating talk on how thinking about “what’s next?” is critical to sustained success. He posits that a brand new thing comes and revolutionizes everything every 10 years, and that technology matures over the preceding 10 years. It does make sense when you think about PC -> mobile -> cloud. He then talks about what he thinks the next big thing will be: machine learning + the concept of microservices in a purchase economy. In other words, the app store is dead! Just a brilliant talk!
Great talk by Prof. Andrew Ng of Stanford/Baidu on how AI will permeate almost every conceivable industry (except hairdressing!). I’ve been interested in AI for a long time and it’s so great to see it finally coming into “the eternal spring”.
- Machine learning used to plateau despite increasing amount of training data – this is no longer the case. Now the more (good) data you have, the better your AI performs. This needs not only AI/ML expertise but also hardware expertise to handle next-level computation with an increasing amount of training data.
- AI progresses fastest when it’s attempting to do something a human can do; after reaching the human-level of accuracy, progress tends to slow down.
- If a typical human can do a task with less than a second of thought, then AI can automate it now or in the near future.
- While programming methodologies such as Agile have had the time to mature into their present forms, there remains a need for an effective method for communication between project managers and engineers in AI projects.
- Particular domains where AI is very likely to take off are:
- speech recognition;
- computer vision, e.g. facial recognition; and
Mo Gawdat is a remarkable thinker and the Chief Business Officer at Google’s [X], an elite team of engineers that comprise Google’s futuristic “dream factory.” Applying his superior skills of logic and problem solving to the issue of happiness, he proposes an algorithm based on an understanding of how the brain takes in and processes joy and sadness. Then he solves for happy.
Amazing talk by Kathy Sierra – author of numerous authoritative texts on Java – on “how to get way better, way faster”.
- No one has infinite cognitive reserves – even though we are often treated as if we do.
- Cognitive reserve can be thought of as a single tank that can be depleted in various ways, leaving less for subsequent activities. Tiny leaks add up.
- Help thy fellow developer by being mindful of draining their cognitive resources unnecessarily.
- Rethink the common adage “practice makes perfect” as “practice makes permanent” instead. It makes whatever you practiced permanent – even if they were bad habits.
- Key message of this talk: to go from novice to expert, what’s needed is a very high quantity of high quality exposures.
- this needs to be applied to sufficiently small chunks of the overall thing being learned
- a chunk is too big if you cannot master it in three 45-90 mins sessions, aiming for 200-300 exposures.