It’s incredible how advanced medical technology has become. Many illnesses that were previously fatal are now treated with ease. Even something as serious as heart failure can be survived, and subsequently treated with an implanted cardiac resynchronization therapy (CRT) device. You’ve probably heard them referred to as ‘pacemakers’ but CRT implants have become a lot more complex over the last few years. Pacemakers send out low energy pulses to restore normal heart rhythm if the patient is experiencing irregularities (arrhythmia). CRT implants can do the same thing, but are advanced enough that they can constantly monitor the heart rhythm and send low or high energy pulses to more effectively maintain a regular rhythm. CRT implants can also include an inbuilt defibrillator, in case of cardiac arrest.

There is a problem however, in that a significant proportion of heart patients do not respond (or respond poorly) to cardiac resynchronization therapy. Given that these devices cost at least $30,000 each, it would benefit the healthcare system tremendously if we could find out beforehand who would, and who would not, benefit from CRT. The patients would also benefit by avoiding an uncomfortable procedure, and being able to explore alternative treatment earlier.

Kevin Howe took up the challenge as part of his summer of research project. Existing research in the area focused largely on applying a set presentation or morphology to predict patient responsiveness to treatment. This was effective in predicting outcomes but required very specific markers to be selected beforehand. As such, existing studies were limited to the small number of markers that were already suspected to correlate with certain outcomes. This limited their ability to effectively utilise large volumes of patient metadata. Kevin’s plan was to incorporate machine learning to analyse much larger datasets, thereby providing more accurate predictions.

For his research, Kevin used the anonymised health records of 61 CRT patients from Waitemata DHB. From these health records, Kevin used a range of different pre-implant and general information for use as potential predictive markers. He then used the post-implant information to mark patients as ‘responsive’ or ‘non-responsive’ to the CRT. The data was analysed using Microsoft’s Azure machine learning platform. Machine learning is a complicated process, but Kevin yielded positive results. After running the experiment 24 times, he was confident that it could be more accurate at predicting the responsiveness to CRT than existing methods. There were minor issues present due to the small sample size, but Kevin felt that these would be lessened with a larger dataset.

As a proof of concept, Kevin’s research was a success. He believes it is possible to use machine learning to create a clinically useable screening tool. With enough data, and with more custom machine learning models, it could become possible to predict all sorts of medical outcomes for individual patients. Knowing how effective certain treatments will be for each patient means clinicians will be able to fully implement precision medicine, providing the best possible care for everyone who walks through their doors.

Kevin Howe is among a group of students who took part in the summer of research programme funded by Precision Driven Health. This month we are featuring a blog series examining these projects. While at an elementary stage and considered to be a ‘proof of concept’, these projects offer fresh insights into what the world of healthcare will look like when precision medicine is fully implemented.