This Harvard scientist is developing a tool that predicts suicide years before it happens

Medical data are fed into the computer for analysis. The computer detects patterns. A certain accumulation of symptoms could indicate an increased risk for a particular disease. In this way artificial intelligence is learning how to assist physicians to diagnose and identify problems, and make forecasts about the future. Nothing surprising so far about that. But what if an analysis of one’s medical data could reveal the inner workings of the mind? Identify early signs of psychosis, or discern that someone might have a tendency to put an end to their life – even before the person had thought of it?

That’s exactly the topic studied by Ben Reis, director of the Predictive Medicine Group at Harvard Medical School. Prof. Reis, who was born in Israel but moved to Boston as a child with his Israeli parents, is developing tools that can detect a high risk for suicide three years before the act itself, tools that indicate physical abuse and – his latest project – a system for the early diagnosis of psychosis.

“Psychosis is a category that describes a whole range of conditions that indicate the loss of contact with reality,” Reis says in articulate, fluent Hebrew, despite his many years in the United States. “The definition is very broad. It encompasses schizophrenia, bipolar disorder and many other illnesses. Psychosis usually appears for the first time during adolescence, or in the early 20s, but in most cases, people get to a psychiatric clinic only at a later stage in life. That means that precious time is wasted, because the earlier the treatment of psychosis begins, the better.”

Let’s talk about what you have been doing. You took medical big data, which also included data about people who suffered from mental illness, and you tried to understand whether certain patterns could identify people suffering from psychosis.

“Yes. We asked ourselves whether a computer could detect psychosis within a few months of its eruption, instead of a few years.”

And what did you discover?

“That yes, the computer can identify an initial appearance of psychosis within three months. We are currently developing the models, but it can already be said that a tool like this could very much change the field of mental health.”

• 'We think of Israel as a country where it’s impossible to disappear. It's not true'
• They wanted to cure heart disease. They discovered a way to grow a different kind of muscle
• 'They became millionaires, while I raised their children for them'

Explain.

“Today, from the moment psychosis appears, it takes the medical system two to three years before it is able to make the diagnosis. It’s only after the diagnosis that the physicians and the family can reconstruct events of the past and understand how they are part of the psychosis. Which is a problem, because although there are effective methods for treating psychosis, they are most effectual during the first months. In many cases, the patient can be restored to the level of health and independence that preceded the eruption, but that’s a lot harder if two or three years have passed.”

According to the computer, what are the risk factors for psychosis, and the attendant physical characteristics? Blood pressure, for example?

“With the methods of the past, researchers identified several discrete risk factors. The big data era makes possible models that are far more complex. Each model factors in about 30,000 different parameters that might be found in a medical file. Each parameter can slightly reduce or slightly raise the risk. What this means is that there is no short list of discrete risk factors that we can point to.”

In contrast to the system for diagnosing psychosis, the tool for detecting a suicidal tendency extracted a number of risk factors from the data. Says Reis: “It turns out that the external pressures that usually lead to a suicidal condition can be detected in [an individual’s] medical file in multiple ways, which the computer identified as patterns. For example, stomach problems in the form of an ulcer, or problems of the immune system, which is weakened by stress.”

The conversation with Haaretz was preceded by an online public conversation under the auspices of the Van Leer Jerusalem Institute. There, Reis presented a graph showing the principal mortality factors in the West over the past 100 years. The graph shows a clear trend of a decline in mortality in nearly every disease, including tuberculosis, pneumonia and influenza. The only mortality factor on Reis’ graph that shows a rise is suicide.

“How is it that modern medicine succeeds in minimizing the other factors of death and fails when it comes to suicide?” Reis asked, and suggested an answer: “People feel shame at having suicidal thoughts, and refrain from getting help.”

Reis relates that during a talk he gave in the United States about a predictive tool he’s developing, one that is intended to identify people who have been subjected to domestic abuse, a psychiatrist who specializes in suicide suggested that a designated tool should be developed for that subject. Reis took the suggestion to heart and received a development grant from the Tommy Fuss Fund, which was established by a couple whose son committed suicide. “The Fuss family supports our research because it wants cases like that of their son not to recur, for there to be another layer of protection. Maybe the next time a doctor sits with a patient who is thinking suicidal thoughts, signals from the computer will indicate a high risk for suicide and the doctor will ask the patient if they want professional help.”

Medical databases include the patient’s family status: single, married, separated, divorced or widowed. The segmentation created by the software revealed that the highest suicide rate exists among those who are separated – people in an intermediate state, shortly after the presumed crisis, with change lying ahead. This might not be only a statistical correlation, but a situation that leads to an increase in the suicide rate; however, that’s a subject for a separate study.

Reis notes that the system he and his team developed could identify – after the fact, of course – between a third and half of the cases that would end in acts of suicide an average of three years before they occurred, and did so solely by reviewing medical files. A pilot study will soon be undertaken in six hospitals across the United States. Afterward, if all the required authorizations are received, the researchers will explore ways to make the program available for clinical use.

As with other predictive tools, ethical questions arise. How should the risk discovered by the computer be presented to physicians? How open should the physician be with the patient about the computer’s findings, and what kind of precautions can be taken against self-fulfilling prophecies? For the sake of scientific accuracy, should other parameters, such as socioeconomic status and skin color, also be factored in, or would that be racist and cause unnecessary biases.

Reis’ interlocutor in the recent online conversation under the auspices of the Van Leer Institute was Oren Harman, a writer and a historian of science. “A great responsibility devolves on people like Ben to act morally,” Prof. Harman said. “My impression from Ben’s group is that they are being careful and are aware of the ethical issues. They’re not a group of geeks who are writing code with their heads plunged into the console, but people who are capable of understanding the social implications of the science they are practicing.”

Harman added that in his view, “The great task of everyone who is engaged in machine learning is for the public to learn and know what happens in those black boxes of data and computerized learning, because it’s already affecting many areas of our life, as well as decisions that are made about us. Our computer literacy needs to step up a notch.”

What was Grandma’s cause of death?

Reis’ research group at Harvard is engaged in a variety of fields of inquiry, all of which he subsumes under the category of “predictive medicine.” “We have projects in the realm of public health that are based on numbers and general information,” he explains. “There are projects relating to personal health, based on individual information, and projects in which we are trying to come up with medical predictions through observation of the morbidity within the family.”

What does that mean, in practice?

“You are undoubtedly familiar with the situation in which the physician asks you about the family’s medical history. Is there cancer in the family? You don’t know, so you call your mother and ask her what Grandma died of. You thought it was cancer, but she tells you it was actually diabetes. That information, which comes from the patient, is known to be very partial and often mistaken. We are asking whether it is possible to create a medical protocol that will be able to know the truth about Grandma – scientifically, of course – and which will help us diagnose what her granddaughter has.”

To know what happened in the past, because history often repeats itself in genetics.

“Yes. In the past, family doctors were exactly that: they knew the grandparents, helped deliver the grandchild and watched everyone grow. They knew the medical history of every member of the family and managed a clinic of true family medicine. Nowadays, the family doctor is more of a code word for a person who administers initial treatment. But if we provide family doctors with a tool through which they will genuinely know the family and be aware of the context of the illness they are observing at that moment, the medical aid they provide will be more accurate.”

Reis, who is also a faculty member at the Boston Children’s Hospital Computational Health Informatics Program, entered this field after completing medical studies and obtaining a Ph.D. in musicology (about which more below). “I realized that the clinical track in medicine wasn’t for me, and I thought it would be interesting to study something having to do with medicine utilizing my other background, in predictions. I had a hammer – namely, a tool for constructing forecasts – so everything looked like a nail, namely a sequence to be predicted.”

Reis decided to focus on a medical phenomenon that is difficult to diagnose, but whose early discovery could save lives. “I looked for an area in medicine where an objective computer, relying on data, could make a contribution. My supervisor told me that there is a dangerous, sad and tragic medical situation that exactly fits that definition: domestic abuse. In fact, 80 percent of the cases of abuse in the family are never reported. Society relies on teachers and physicians to detect it along the way, while they are doing other things. The problem is that this is successful in identifying only 20 percent of the cases.”

Can you explain a little about the development of the system?

“The computer was given a list of around 500,000 anonymous patients, in which all identifying details were hidden. We even changed the dates of the patients’ visits, so that there would be no way to trace the subjects. We ask the computer to perform a controlled study of their data, though we already know that 10,000 of them had been subjected to abuse. After the computer analyzes the data, we ask it if it has observed any variables more frequently among those who had suffered abuse, compared to those who did not. It’s told to search for recurring sequences. At the end of the analysis, the computer finds that there are several things that are more frequent in that group, as compared to the rest of the population. Now comes the real test. The computer is fed new data about 500,000 other people and is asked, on the basis of the data it collected from the previous group, who in the new group underwent abuse.”

Because the computer is analyzing data about real people, we can know precisely when it is accurate. In other words, at what stage of the abuse the symptoms will coalesce into a precise prediction. Reis and his team found that the computer is two years ahead of the authorities in detecting abuse.

Reis’ group is in contact with physicians in order to understand from them how this tool can be of use. After all, the software can conclude that a person underwent abuse at an intermediate or even high probability, according to a sequence. The computer can be overly “suspicious” and warn about cases that are not clear-cut, or it can be too restrained and miss things. Together with the physicians, the team is trying to strike the right balance.

Another question relates to how a warning is worded. “That sounds marginal,” Reis says, “because what difference does it make what message the computer transmits to the physicians? But it’s the most important part of the study, because the system is intended to strengthen and empower doctors in their treatment of patients. We want to help them succeed. We are trying to get the computer to project modesty and support, and not have a pretense that it is replacing the physicians.”

In practice that means the computer will signal physicians about the suspicion, will suggest devoting a few more minutes to the patient and asking a few questions, and if needed, also propose bringing in a social worker. “There is one goal: to prevent abuse victims from falling between the cracks. What the software does is to say whether this person is in a higher risk group than is typical. In certain cases, that information could say lives.”

Within a few years, Reis estimates, this tool will be available to be integrated as an application in the computers of health systems. “With regard to the health of individuals, there are a great many hurdles to overcome before the system can start to work. It’s frustrating for anyone who wants to help people, but it’s important for things like this to go through all the tests.”

Vaccine research

One of the striking products of the research Reis has been involved in will be familiar to many in Israel, namely, a study conducted by the Clalit HMO’s research institute in collaboration with Reis’ research group at Harvard about coronavirus vaccines. Together they investigated the vaccines’ efficiency by means of algorithms that examine medical files. They also examined their safety and a number of other, more focused issues – such as how the vaccines and the disease affect pregnant women and children.

According to Reis, “Because Israel was one of the first countries in the world that conducted a large-scale vaccination project, it found itself in the forefront globally in terms of the ability to examine the effectiveness and safety of the vaccines in the field, after their approval following clinical trials.”

The first article to be published based on the joint study showed the vaccines to be more than 90 percent effective in preventing morbidity and mortality. A second article, published a few months later in the New England Journal of Medicine, dealt with the volatile issue of side effects: “We chose 25 side effects in order to compile the broadest possible list, one that included serious phenomena such as stroke, myocarditis, kidney problems and so on. We looked at data of people who were vaccinated and those who were not, and we looked for epidemiological matches in them.

“For example,” he continues, “if there was a diabetic male of a certain age who lives in Bnei Brak and who was vaccinated, we went and found the data of another man of the same age who also lives in Bnei Brak and is diabetic, but not vaccinated. Of course, it’s all done anonymously by the computer. The computer identified epidemiological matches and pointed to a few hundred thousand matches of this sort. The reason we do the matching is to ensure that there is no single cause responsible for the observed differences, such as age or background illnesses.”

Reis explains that “the first part of the study examined how many cases of stroke occurred among both the vaccinated and the unvaccinated, comparing the epidemiological ‘twins’ [whose profiles varied only in the vaccination question]. We discovered negligible differences. In regard to myocarditis, which was much talked about, we found that statistically, there was an increase in the frequency of cases in two individuals out of 100,000 who were vaccinated compared to the non-vaccinated. A parallel study conducted by Clalit [health maintenance organization] about these cases found that they were mostly mild and ended safely after four days.”

And then you conducted a similar examination of epidemiological twins on a new group of nonvaccinated people, one “twin” was sick with COVID and the other “twin” didn’t get the virus?

“Yes. Here we discovered serious and quite scary numbers. The data speaks for itself. There is a rise of hundreds of cases per 100,000 persons in side effects caused by the virus. More heart ailments, more kidney problems. By the way, in regard to myocarditis, which came up in connection with the vaccine and because of which people were to exercise caution before being vaccinated – we found that among non-vaccinated individuals who were infected by the coronavirus, 11 people per 100,000 fell ill with it, and in a way that was serious and life-threatening.”

Numbers that gave the vaccine more of a tailwind.

“When you look at the data coldly and drily, without dramas or preconceptions, but just let the findings speak for themselves, you see something clear, which can be summed up with the following advice: If you’re scared of side effects, then what you should find more threatening is the virus, not the vaccine.”

Because the computer is objective; the algorithm has no opinion about vaccines and diseases.

“Exactly. If the experiment is constructed correctly, the truth emerges, whether you like it or not.”

Computer infants

Two loves propel Reis’ professional life: medicine and music. As an undergraduate he focused on both, and on the way to becoming a medical researcher, he took an academic “break” from the former in order to obtain a doctorate in musicology. He spent three years at Cambridge University researching musical prediction, seeking to understand how the human brain learns the musical language of its culture.

Because it’s difficult to do research about the way that infants become acquainted with music and store it in their memory, Reis created musical infants of a virtual nature. They consist of software that “remembers” all the musical information to which it is exposed. In this way, he “raised” a computer-infant that was exposed to Chinese folk music along with another that was exposed to German folk music, and so on. In the next stage of the research, different songs were played to the computer-infants, some exclusive to the musical culture to which they had been exposed and some from a foreign musical culture. The computers were asked to predict the coming notes in each song that was played.

“We had a collection of thousands of folk tunes. Based on the computers’ responses, we created a map of musical connections between places,” Reis related in the conversation with Oren Harman. “We discovered, for example, that the folk music of Germany and that of England are closer to each other than the folk music of France and England, or that the musical cultures of India and China are similar, whereas that of Japan is separate and anomalous, different from the music of its neighbors in the East.”

In my video conversation with Reis a few days later, I ask him whether he discovered in his research universal music, of a sort that all the computer-infants were able to predict [the notes to follow in a musical sequence] in one way or another. He smiled broadly and said, “I spent the first year of my doctoral studies reading everything I could exactly on that question: Does a universal music exist, from which all the other styles emerged? A similar question was asked 150 years ago: Are there musical laws that don’t change, that transcend styles? It’s a research field known as ethnomusicology or anthropological musicology, in which an attempt is made to explore as many cultures as possible and to sample their original folk music and then to look for rules that are common to all of them.

Initially, Reis says, “All kinds of Western-centered theories sprang up, according to which they all have seven tones on their scale, like Western music. But gradually more and more styles and laws were discovered. In the end, the only universal thing found in music – and it’s impressive that it’s so narrow – is the octave. That means that almost every person in the world hears first ‘do’ as equal to the last ‘do’ in the octave, only higher.”

Reis elaborates: “The musical experience is based on a physical sensation of frequencies. If you heard the note ‘do’ at a frequency of 440 Hz, and then at a frequency of 880 Hz, your brain will feel that it’s the same thing, only double. The fact that an octave is heard as the closing of a circle everywhere in the world – in other words, ‘We have now stopped climbing and have returned to the beginning, only higher’ – constitutes the tonal framework of all the musical styles we know. What differs between musical styles is only the question of how the space is divided among the octaves. In the West, an octave is divided into seven tones – ‘do,’ ‘re,’ mi, and so on), and in China into five – if you play only the black keys on a piano, it will divide the octave into five [tones], and will sound like Chinese music.”

Reis notes that, “Every culture that has ever been discovered has had music, and within that music there is always something that is the equivalent of an octave. But every attempt to find other general laws has failed. I find that exciting, because it shows that there is no end to musical creativity and that every exposure to different musical cultures and styles is highly enriching. It’s wonderful that there is so much freedom of creation and no universal laws, even if it made my doctorate harder.”