Like so many aspects of our lives, health and medicine has benefited enormously from the adoption of a wide range of technologies that support medical professionals to detect, diagnose and treat health problems. Artificial intelligence and machine learning in particular are revolutionising healthcare; helping to take some of the pressure off of hardworking clinicians and their oversaturated departments, minimising the chances of human error and working efficiently and quickly to the benefit of patient care.
A team of deep learning senior research engineers at Ciklum explored how machine learning can help clinicians to prevent complications related to incorrectly placed tubes and lines in patients.
Competition: Using deep learning algorithms to detect misplaced lines
Between December 2020 and March 2021, Royal Australian and New Zealand College of Radiologists (RANZCR) hosted a competition on Kaggle, an online community of data scientists and machine learning practitioners.
RANZCR tasked teams to develop deep learning algorithms which could automatically detect intravenous lines and catheters in patients which had not been placed correctly. If successful, the application of machine learning in this particular area of medicine could speed up the inspection process of lines, and alert clinicians swiftly and efficiently to those which may be in the wrong position.
The challenge: Eliminate human error
RANZCR is one of many medical organisations working across the world to tackle the problem of incorrectly positioned lines and catheters. If not caught early, these simple errors can result in risky complications, and in some cases, death.
The problem of misplaced lines has been compounded by Covid-19, as growing numbers of patients in hospitals and ICUs across the world find themselves in need of these tubes for life-saving care. Currently, medical professionals must manually check X-rays to determine whether a line in a patient is in the correct place – resulting in numerous delays and potential human errors, particularly as radiologists are often overwhelmed and busy with many other scans to report on.
The process: Creating a machine learning model from scratch
Medical image processing involves incredibly detailed analysis of medical images – such as X-ray, MRI, and CT scans, which are generally in grayscale.
The Ciklum team trained and tested their machine learning model using a dataset, provided by RANZCR, of 40,000 chest X-ray images showing various tube placements.
Each image either fell into the category of normal, borderline, or abnormal. Those tubes in the normal category would not need repositioning, whilst those in borderline or abnormal categories would need confirming or readjusting by a clinician.
Computer vision technology, combined with deep learning, enabled the model to identify and classify the images into these three categories. The benefit of using AI in image processing is that a machine can surpass and rival people’s own visual abilities, eliminating the chances of human error and working faster than a person would.
The result: 97.255% in detecting misplaced lines
The team worked on the project for three months, pre-processing the images before running through the ML model. At the end of the competition, the 1500+ teams were required to submit their best solution. Each solution was tested against a hidden data set, and the overall performance of each team’s approach was evaluated using the mean column-wise area under the curve metric.
The solution provided by Ciklum’s team was able to reach 97.255% of the metric – which was just 0.3% short of the winning team’s result.
The team were delighted to be awarded Silver in RANCZR’s coding competition, and as a result of taking part, were able to develop their own brand new approaches to solving these kinds of problems in the domain of computer vision and medicine.
Other applications of artificial intelligence in healthcare
AI products can help mitigate the thousands of avoidable mistakes in patient care which happen every year, and can also spot tiny, subtle patterns that could indicate a health problem.
Imaging is at the very heart of modern medicine and a vital part of deciding a course of treatment, and so computer vision, does have a fundamental role to play in supporting clinicians. Ciklum’s ML solution created for the RANZCR competition could be used in a variety of clinical settings in which medical images need to be processed, and could be applied to a range of tasks which requires the classification of images containing fine and subtle details – including detecting cancerous cells and other abnormalities.
Machine learning algorithms have also been used to detect dangerous prescription and medication errors, in which a patient is prescribed the incorrect drug or dosage; and in 2020 Cancer Research UK identified that an AI programme gave fewer incorrect diagnoses of breast cancer when assessing mammograms than radiologists. In this way, AI could not only be utilised 24 hours a day to inspect mammograms, but could also result in fewer false positives and fewer false negatives, than when the same inspections are carried out by humans.
ML and computer vision are not the only branches of AI with successful applications in healthcare – in 2016, a healthcare startup Neopenda worked with Ciklum’s R&D engineering team to develop a simple, wearable IoT device that measures, tracks and displays vital signs for critically-ill newborn babies.
Ciklum’s unique experience in solving problems using a deep learning approach can be applied to a whole range of healthcare projects. Discover our capabilities in AI here.