According to the WHO, there are approximately 1.1 billion smokers in the world. Russia ranks fifth in the number of smokers - over 45 million people. To combat the sad statistics, scientists have proposed a way to combat smoking based on artificial intelligence.
About 400,000 Russians die each year from smoking-related diseases. And while the state is taking measures to limit tobacco consumption at the legislative level, researchers are developing effective methods based on artificial intelligence (AI) technologies. Andrey Polyakov, a researcher at Philips Research Lab Rus, spoke about how neural networks and machine learning can help in the fight against smoking.
What can be said in general about the study: how did the idea originate, why should artificial intelligence help people quit smoking?
- One of the most effective smoking cessation strategies is medical advice. During consultations, the specialist provides psychological support to the person quitting smoking, not to let him break down. But face-to-face consultations are quite an expensive pleasure for the healthcare system, and patients do not always have the opportunity to visit a doctor often due to the remoteness of specialized clinics.
Employees of the Russian and Dutch laboratories Philips Research thought about solving these problems. Scientists have set themselves the goal of scaling consultation sessions to a wide audience of smokers who have smartphones with internet access. The results of the study were presented in the summer of 2018 in Stockholm at the IJCAI-2018 conference. The idea is to automate a therapeutic intervention and provide remote assistance to a person to quit smoking using the capabilities of artificial intelligence.
We are talking about a conversational agent on a smartphone that is able to select and apply one of the patient-supporting strategies. He can recognize the emotional coloring of the patient's speech or text messages, respond appropriately to it and help the person get rid of the bad habit.
What principles of AI are the basis of the method?
- These principles are based on the modeling of smoking cessation methodology using cognitive-behavioral therapy and motivational interviewing, which are usually conducted by a doctor at the reception. Naturally, in a live conversation, a person can understand the mood and state of the interlocutor thanks to various verbal and non-verbal signals: these include speech, voice, facial expressions, gestures.
Promotional video:
In our research, we were interested in the language in which we communicate in instant messengers and social networks. In order for artificial intelligence to replace a psychotherapist, it needs to be able to recognize a person's spoken and written speech, its emotional coloring, as well as maintain a conversation and respond to changes in the patient's condition.
How does artificial intelligence learn to analyze speech?
- Deep learning methods, in particular recurrent neural networks, combined with the availability of computing tools and accumulated data, have made a breakthrough in many areas of artificial intelligence, including speech recognition and processing. With the help of these technologies, several high-tech companies have been able to create voice assistants with whom you can communicate and set tasks: Siri from Apple, Google Assistant from Google, Alice from Yandex.
Although recurrent neural networks are a popular text recognition tool, they require a large amount of labeled data that is difficult to collect. In addition, the communication process is an example of AI learning in a non-stationary environment, since our speech changes greatly both over time and under the influence of national characteristics of different cultures.
These factors require local configuration and maintenance of the classifier (in our case, a deep learning recurrent neural network) already at the level of an individual user. One of the popular approaches to continuous improvement of a classifier is active learning. The main idea of these methods is to mark up only a part of the received data that is of interest for further application.
Typically, today's active AI learning methods work well for traditional tasks. In doing so, they can lead to technology instability, which is common in deep learning neural network architectures.
Our method is a new algorithm for active learning of neural networks based on the following principles: semi-supervised learning, recurrent neural networks, and deep learning and natural language processing.
The mechanism of work is as follows: the algorithm is given a text message, as it happens when communicating in instant messengers. The task of the algorithm is to recognize its emotional coloring in relation to the topic of smoking. It can be positive (“I personally quit, I don’t smoke, I am cheerful and full of energy”), negative (“I smoke again”) or neutral (“Moscow is the capital of Russia”).
Twitter posts processed by neural networks during research / Philips Research Press Service.
Depending on the emotional coloring, the algorithm applies appropriate behavioral strategies: change the topic of the conversation in the case of a positive coloring, support the conversation with a negative coloring, and react neutrally in the case of a neutral message.
How was the study of the effectiveness of this method carried out, what were its results?
- The purpose of our study was to develop a new method for searching and selecting data of particular interest. To show what kind of data we are interested in, consider the following example. Imagine a jury taking a case in court and deciding by a majority whether a person is guilty or not. In this case, the jury can always turn to the magician Merlin, who knows for sure whether the defendant is guilty. But he demands payment for his services.
The jury wants to do their job conscientiously, but at the same time has a limited budget and cannot contact Merlin for every case. A case is considered uninteresting if the jury votes almost unanimously for guilt or innocence, this is a simple case. But if the votes of the jury are divided, then this is of interest.
In this case, the jury turns to the magician, receives an answer and, when considering the next similar cases, will make more coordinated decisions, which in the future makes similar cases simple. Moving on to the terminology of the algorithm, a jury means a classifier (neural network), a jury means a committee of classifiers, a court case means a tweet message, and Merlin means an expert marking up messages.
Thus, several neural networks, based on the accumulated experience, decide what emotional coloring a particular tweet carries. For example, if they almost unanimously give a tweet a positive emotional connotation, then it is classified as positive. If the neural networks "get confused in the readings", then the tweet is marked as interesting.
Further, all interesting cases are collected, which are ranked according to the degree of confidence in the predictions of the classifiers, after which these cases are sent to the expert for marking. Further, the specialist conducts additional training of neural networks based on the analyzed cases.
What did you manage to create in the end?
- As a result of the research, a new Query by Embedded Commettee (QBEC) active learning algorithm was created, which differs from the existing ones in terms of accuracy and speed. During the experiment, we applied a new algorithm to classify short text messages from Twitter using recurrent neural networks.
First, a training database for AI was collected and manually tagged from more than 2,300 English-language Twitter posts published from October 2017 to January 2018. The October messages were linked to the Stoptober European smoking cessation campaign. As part of this campaign, people quit smoking and post tweets for a month in which they share their impressions of quitting cigarettes.
The December messages were written by people who were going to quit smoking by New Year. Additionally, a test base was collected and manually marked up. The applied text classification system was based on modern architectures of deep learning recurrent neural networks. She was trained on the tweet training base.
The accuracy of the classifier that was learned with its help was very low and barely exceeded 50%. Then we conducted another experiment, in which we consistently applied the active learning mechanism: every day the classifier received a new portion of targeted messages (about 3000 daily) and gave 30 of the most interesting cases for marking.
These messages were manually tagged and added to the training database, which was used to build the next classifier model. The study showed that this method of teaching artificial intelligence allowed for a qualitative improvement in the algorithm. Computational experiments and theoretical calculations demonstrate a much higher speed of the QBEC algorithm.
This circumstance makes it possible to run the QBEC active learning algorithm even on a user device, such as a smartphone. This means that we have a chance to create an effective voice assistant that can take over the function of a doctor and help people who are trying to quit smoking.
What predictions can be made based on these results, how effective will AI be in helping people quit smoking in the future?
- The research results show that artificial intelligence is able to recognize the emotions of the patient from the text of the message, while active learning algorithms can continuously improve the accuracy of data classification. Our challenge today is to ensure that in the future, the percentage of people who quit smoking with the help of AI technology will not be lower than the percentage of people who quit smoking through face-to-face consultations.
The introduction of AI into medicine can reduce the financial burden on the healthcare system and reach many more patients who want to quit cigarettes and lead a healthy lifestyle.
It can be assumed that in the future this approach will be applied, among other things, to help patients with alcohol or drug addiction. Also, doctors will be able to more often turn to the capabilities of AI in identifying mental disorders.
For example, recently, scientists from the University of Pennsylvania developed a neural network that analyzes user posts on Facebook and determines whether people are depressed. The diagnosis of this disease is not always unambiguous, therefore, the accuracy of the algorithm during the study in 70% of cases was comparable to the results of medical screenings.
Such examples prove that the possibilities of using artificial intelligence in medicine are endless and can help doctors solve many social problems.