Romain is an Applied Scientist at Onfido, specializing in facial biometric fraud attack detection.
Artificial intelligence (AI) has already started to change many aspects of our lives. We can ask questions out loud and get immediate answers from the likes of Alexa, Google Home and Siri. The keyboard on our mobile phones predicts the next words we want to type as if it knew our train of thoughts. Ask Google Photos about a birthday party two years ago and it will find the images of that joyful moment. There are plenty of examples, but the truth is that they are only the tip of the iceberg.
Artificial intelligence has been applied to a variety of applications and industries. Some examples include: network security, farming, autonomous driving, identity verification (what we do here at Onfido), healthcare and even drug design. And as more industries adopt AI, it will likely see an increased use in society as well.
The potential of AI has even been recognized by the United Nations (UN). Every year in May, the UN organizes AI for Good, a 4-day-long global summit. This year it focused on education, health, inclusivity, human dignity and agriculture. The impact of AI in nearly every sphere of life is constantly growing and the potential for it to be used for the greater good is immense. In this blog post, you will learn about the different aspects of humanitarian crises and their respective complexity. You will also learn about how AI researchers are currently attacking these problems and what solutions they have created. Our aim is to demonstrate that AI can be and is already a technology that can dramatically help us in mitigating crises.
The role of AI in detecting outbreaks
What if we could identify crises more quickly? The deadly example of Coronavirus, which started in China but quickly affected the rest of the world, is a crucial wake-up call. We’ve all seen the dramatic increase in the number of new cases everyday, and it highlights the exponential nature of infectious diseases.
Once viruses have become prevalent among populations, time becomes a valuable resource. For every day not spent preparing for the outbreak, leads to thousands of additional deaths down the line. That is the terrible nature of exponential growth. If we want to mitigate the impact of an outbreak, it’s crucial that we develop technologies to detect them as early as possible.
The reality is that it can take up to several days before people and governments are fully aware of the danger of disease outbreaks. But what solutions can we use to detect outbreaks earlier? Among the possible avenues is social media analysis. Facebook took advantage of this to propose their Crisis Response platform. The idea is that if a high number of people are talking about an event that is unexpected and concerning (such as the Ebola outbreak), it is likely something risky is happening.
And the quicker we are able to detect when a ‘risky’ scenario is happening, the better the response can be. More than 8,000 messages are exchanged on Twitter every second. It is physically impossible for a human to monitor them all. Algorithms and automation are a potential solution.
What properties do algorithms need to have when detecting crises?
But conceiving and running these algorithms is not a straightforward task. Among all the desired properties we want the algorithm to have, two are essential, but difficult to produce.
The algorithm should incorporate language understanding, so it can link two messages that express the same idea, but are written differently. For instance, “I went to the doctor this morning” and “After breakfast I saw Dr. John at the hospital” are two completely different sentences, yet they express roughly the same message.
If we consider the kind of messages that could be exchanged at the beginning of an outbreak, both “my football teammate was feeling unusually tired today. We lost the game…” or “I haven’t eaten more than a banana a day for the last 5 days. Best diet ever??” express potential symptoms of Ebola (respectively weakness and reduced appetite). But neither explicitly mention Ebola.
The second property we would want the algorithm to have is that it should only get triggered for rare events. In essence that means the algorithm should raise an alert only when it has enough evidence that an alert is worth raising. But what is a rare event in the first place? 10,000 messages sent about the same topic over 10 seconds, like when a football team just scored; or 1,000 messages over two hours when there is a growing uncertainty over a disease outbreak?
Using these algorithms in real-world examples
As of today, this kind of algorithm already exists. They are what we commonly call artificial intelligence (AI). In 2015 Columbia University Medical Centre study on the Ebola outbreak analyzed tweets during a 5-day period preceding the official announcement of confirmed Ebola cases in Nigeria. They observed that tweets mentioning Ebola sent from Nigerian cities grew in numbers that were 6 times more than usual. In effect, they managed to identify a humanitarian crisis before the government itself sounded the alarm.
More recently, the achievement of an artificial intelligence company, BlueDot, shed light on what AI is capable of when it comes to real-time outbreak detection. The company used multiple sources of information like flight schedules and public health records and fed them to their algorithm. Which in turn analyzed and cross-referenced them. In December 2019, they became the first to spread the word about an outbreak in Wuhan. For the most AI-savvy readers, we recommend listening to the episode of the This Week in AI podcast where the founder of BlueDot is interviewed.
Detecting outbreaks is a tremendous challenge. As early as 2015, The World Health Organization Ebola Interim Assessment Panel was already pointing out in its report that "better information was needed" and "innovations in data collection should be introduced, including geospatial mapping [...] and platforms for self-monitoring and reporting". No doubt AI will play an important role in shaping a reliable and sustainable solution for outbreak monitoring in the future.
Using AI to improve triage and decision making
Detecting the outbreak of a humanitarian crisis is only one of the many applications that AI could have. Providing assistance in critical situations is another. In 2016, a team from Harvard University and the MIT embedded AI into a mobile app, to provide doctors with an Ebola prognosis prediction based on factors such as temperature and heart rate. This application has the potential to help identify infected people before it’s too late.
To better coordinate the help effort, the Qatar Computing Research Institute proposed in 2014 a system called AIDR that analyses tweets in real-time. The system identifies tweets that contain valuable information (for instance: tweets reporting casualties, infrastructure damage and donations offered), increasing the speed that help can be planned and provided. This system has been tested on tweets related to the 2013 Pakistan earthquake. After only 6 hours of tuning, the AI was then able to detect in real-time 80% of the relevant tweets while misjudging only 20% of them.
Similarly in 2011, UNICEF launched a social messaging tool called U-Report where people are invited to send questions on prevention strategies, counseling or health-related issues. To face the growing demand and the ever-increasing number of messages, they partnered with the Qatar Computing Research Institute in 2016 to apply a similar system to AIDR in the case of HIV/AIDS in Zambia. This means frequently asked questions are answered automatically, while more delicate questions are highlighted by AIDR so a doctor can focus on providing a better-tailored answer.
Image analysis can also assess the impact of a disaster. In a 2018 study by the Skolkovo Institute of Science and technology, researchers used artificial intelligence to analyze satellite images of the region in California impacted by wildfires. Their algorithm compares the images before and after the disaster and is able to provide a damage assessment on a per-building basis. This sort of information can be crucial when coordinating relief efforts.
Looking to the future
The world is constantly changing
The biggest strength of modern artificial intelligence algorithms relies on their capacity of incorporating a multitude of parameters and data points (far too many for a human to fully grasp) to make predictions about the future. In other words, algorithms are able to find correlations between a set of certain events like tweets frequency and hospital admission numbers, and an outcome, e.g. an outbreak. The algorithm learns this correlation. So much so that if the very same set of events happens today, the algorithm will consider that an outbreak is imminent and will raise the alert.
But the world is constantly changing. Social media platforms are more and more focused on images and videos rather than text. Driven by internet adoption in developing countries, the web welcomes new users every year. In these conditions, the AI research community is still facing considerable challenges: how to make relevant today the correlations that were learned in the past?
Whilst this is very much an open research question, we suggest that self-supervised learning is one key area where improvements could emerge.
Self-supervised learning has received tremendous attention over the last few years because it aims at solving an inherent drawback of modern AI algorithms: their appetite for annotated data points. Annotation comes into play when a human being actually pairs some data points with an outcome that we want the AI algorithm to learn. Typically: given the conditions that preceded the Wuhan outbreak, a human being will add the annotation: “outbreak”. Now the AI algorithm will be encouraged to draw correlations between the conditions in Wuhan before the outbreak and the actual outbreak.
Modern AI algorithms typically require from tens of thousands to millions of such annotations in order to reach a satisfactory level of accuracy. Thankfully for humanity, we don’t have records of millions of Covid-19-like or Ebola-like outbreaks. But for AI researchers, this makes it challenging for algorithms to learn reliable correlations.
The promise of self-supervised learning is to harness the AI algorithm with the power of exploring by itself the correlations in the data without the need for annotations. New research papers on the subject are written every week. Recently, researchers managed for the first time to create an algorithm capable of categorizing notions such as “dog”, “cat”, “car” and a thousand other categories without the need for single annotation. While at first this might seem far from the application of humanitarian crises, one must appreciate the implication of this kind of algorithmic improvements.
Having an algorithm that can draw correlations without any annotation opens up a wealth of possibilities in terms of the events it could detect. With this, even the most common incidents (think of car crashes, food poisoning in a restaurant) could be analyzed and used in order to make higher-level predictions. But even more importantly, it means that the algorithm has a better chance at analyzing all the patterns that arise within an outbreak. Predicting the emergence of a new disease is one thing, but we could also be trying to predict a host of other critical information: the geographical areas that will be most affected, which food or household items will be lacking, what is the virality of the disease, what will be the economic impact, etc.
In a way, the algorithm would make the most out of every situation and of past outbreaks in particular.
AI certainly has a lot of potential to predict and support humanitarian crises and natural disasters. As we have shown, we already have the tools in place to detect outbreaks quicker than ever before. We can leverage classification and prediction systems to assist the on-call teams by routing information and helping on the decision process. Whilst we should look to the recent success of BlueDot who managed to predict the outbreak of Covid-19, we believe AI has even more potential to offer in the years to come. And although we covered self-supervised learning in more depth there is no doubt that all areas of AI research (on-device computation, self-supervised learning, domain adaptation, multimodal learning, etc.) will be vital in order to make this a reality.
You can read more about our approach to artificial intelligence in our Center of Applied AI pages.