Towards AI safely used for our health. Here is how!
Artificial intelligence (AI), the broad discipline of creating intelligent machines, has been very well established in the everyday technology and adopted in popular culture more than any other technological achievement. Daily, we interact with intelligent systems while sending an email, unlocking our phones, using social media, receiving recommendations for shopping, travel, entertainment etc. AI is using technology like machine learning (ML)*, natural language processing (NLP)*, computer vision etc., in order to carry out tasks which normally require human intelligence, for example problem solving or learning. But how much can we trust AI for applications that are critical like healthcare, aviation or security? To illustrate the importance of this question, imagine a computer vision system providing false information on skin cancer detection, or a self-driving car unable to detect an obstacle in front of it. Such critical applications require absolutely rapid identification, timely responses, reliability and security assurance, topics that are currently addressed by researchers.
EURECOM’s professor Dr. Maria A. Zuluaga, expert in AI and Data Science will guide us through, on what is AI and how it can be of use for critical applications like healthcare.
Q. What is artificial intelligence and machine learning, terms used so often and sometimes interchangeably?
MAZ: Defining Artificial Intelligence is not easy, taking into account that the field is vast and more of a multidisciplinary entity. In a sense, AI is an umbrella discipline that covers anything related to making machines smart, i.e. a car, a TV or a software application. Along with AI we often come across another field, Machine Learning (ML), a subset of AI, referring to systems that can learn by themselves . Contrary to classic programming, which is designed to carry out specific tasks, predetermined a priori, ML algorithms are trained on a given set of data, in order to learn how to give answers to complex questions. There are several types of ML algorithms, such as supervised*, unsupervised*, reinforcement* and deep learning* as indicative examples.
*See Glossary below for AI beginner-friendly explanations
Q. What are the challenges you face in particular for AI on critical applications like healthcare and what is your approach to address them?
MAZ. My research focuses on the successful use of ML techniques in critical domains facing high-risk decisions, such as healthcare. Critical applications are mostly challenged by three limiting factors: data complexity, low error tolerance and reliability. I address these challenges through the development of robust ML algorithms relying on the human-in-the-loop learning principle. In other words, the goal is not to develop AI software systems that will replace physicians, but to give them tools that will help them make more informed, relevant and fairer decisions, while keeping them in the loop. One major difficulty in medicine is that data, are by definition heterogeneous, complex and difficult to use. This is why traditional ML techniques are not particularly suited. In critical applications, quality and quantity of input data are important to obtain correct results. There is a need for methods based on interactivity and on evaluating the uncertainty of the results. Especially, we are able to reduce the dimensionality of the data needed, in order for ML algorithms to be trained efficiently and still provide robust results. This is a very useful contribution particularly for medical imaging and diagnosis, where available data for training algorithms do not exist in high quantity and require very specialised annotation processes.
Q. Could you give us examples of real-life critical applications of your research projects?
MAZ. Currently, in collaboration with University College London (UCL), we are developing interactive ML techniques, in order to ensure the timely medical diagnosis of critical situations for multiple sclerosis (MS) patients. Brain lesions detection for MS patients is a quite well known process. The challenge now is to be able to automatically detect, when such lesions touch or cross blood vessels in the brain, a situation that requires immediate medical attention. The cerebral blood vessel tree is so complex and really hard to annotate purely by using an algorithm, so humans are efficiently feeding the algorithm for high level of precision in annotations.
Another application we are currently pursuing, goes to the direction of personalised medicine. The idea is to use wearable devices, such as watches, t-shirts, etc. and analyse the collected physiological data of patients, in order to provide health recommendations. For example, in collaboration with clinicians in Monaco, we run a project to provide support to patients with chronic diseases, i.e. cardiac diseases, intersecting their data, with information gathered from smart city devices. Based on data for traffic, pollution etc. the algorithm will recommend to them personalised tips for their daily activities, i.e. how safe it is to go outside due to pollution, how long they should walk, whether to avoid physical strain in a humid weather etc.
Furthermore, note that the methods we develop are generic and transversal, and can be relevant to all critical application domains sharing the same difficulties. For example, we are working on a project with Orange (see recent publication [1]), to create an automated system to detect anomalies while monitoring signals on their IT system, allowing damage prevention and timely reaction to possible network issues. To reliably detect such anomalies, we use a method called “auto-encoders”, a type of neural network* which is able to encode a dataset and decode it in a lower dimension, keeping the most valuable information for the reconstruction. What is considered normality in data is much easier to model and because of the larger amount of normal data available, we have a small reconstruction error. On the contrary, anomalies when decoded exhibit a significantly higher estimation error leading to their successful detection.
Q. What do you think of the future of AI and its technologies?
MAZ. There has been an exploding hype recently around AI and ML techniques and we are currently reaching a plateau for the field’s milestones. I think it is time to reflect now a bit deeper and look into what we can really achieve in the future through AI. Also, there has been a massive number of publications on the field and some unfortunate examples of mistakes found in studies. Therefore, it is highly important to increase our caution in reviewing and keep a critical mind in order to keep the essential and move forward based on more solid grounds, towards both theoretical and technical advancements on AI.
by Dora Matzakou for EURECOM
[1] J. Audibert, P. Michiardi, F. Guyard, S. Marti. “USAD: UnSupervised Anomaly Detection on Multivariate Time Series”. In: Proceedings of the 26th ACM SIGKDD Conference on Knowledge Discovery and Data Mining (KDD’20), August 23–27 2020, Virtual Event
