East Coast Swing

This story opens at an aggressively mediocre Mexican restaurant near Madison Square Garden^[1] in 2012. I was pretending to eat dinner with an aspiring entrepreneur who was pitching a data science startup idea she’d been scheming up^[2]. Discussion over and food sufficiently shifted around the plate to imply eating, she stood to leave, but stopped and said, “You know, I’d drop it all if you could come up with a treatment for manic depression.”

The goal of many of my research stories is to convey why I believe so strongly in the good AI can do in the world, while remaining honest that our collective decisions about it should never be easy. I’ve shared personal stories of using AI for diabetes, autism, education, and even global intelligence. There are so many other projects I want to share–removing bias from hiring, inventing cognitive armor against Alzheimer’s, cyberstalking 100,000 entrepreneurs–but a look at the incredible power of dumb luck^[3] will have to do for now.

I never learned how manic depression had touched the life of that entrepreneur so profoundly, but her 17 words had lit a fire. I spent nearly the entire night looking at research on bipolar disorder (the more modern term for manic depression). I’m a neuroscientist by training and certainly know the basics, but I spend my time making information theoretic models and exploring neural computation. Exploring deeply into this domain of neuropsychiatry was a shock. Globally in 2013, nearly 50 million people experienced the extreme manic and depressive mood swings of bipolar disorder, accounting for 10 million years lived with disability. Nearly 3% of US adults have struggled with bipolar, particularly in yearly adulthood, with more than 4 in 5 experiencing serious symptoms. Although many are able to manage their episodes either with medication or social support, untreated manic episodes can lead to self-destructive behaviors, ruining marriages and careers. Severe cases can trigger institutionalization and dramatically heighten rates of suicide^[4]. Research on the economic burden of bipolar reveals it to be one of the costliest mental health disorders in the US and globally, and these studies can never capture the full drag on quality of life and human capacity. While drug treatments such as lithium exist, they only work for less than half of sufferers^[5], and the side-effects are notoriously so unpleasant that many stop taking them.

My research in theoretical neuroscience did not provide much insight into the clinical realities of bipolar or the everyday challenges for sufferers. Still, I was immediately intrigued by the idea of leveraging machine learning to predict manic or depressive episodes based on simple, everyday data. In fact, as I was reading the research literature into the wee hours of the night, I found a complex story which hints that aggressive, continuous (and unrealistically intrusive) diagnostic testing could offer predictive signals for upcoming episodes, even predicting first-time episodes in those with a family history of bipolar. Perhaps proxies of these signals could be surfaced in everyday, more naturalistic data, like changes in movement patterns or language in text messages. Given the continuing mysteries and frustrations of chronic drug treatments, it seemed like there was so much more to learn.

The reality, though, is that life is busy. That night would just have been a side note in my list of far too many projects^[6] if not for another startup pitch I heard the very next day.

The company was called Emozia. They were a young group of recent university grads in New York City with the enormous ambition of adding a layer of emotional computing to our digital lives. They were building a platform for mobile phones that took seemingly unrelated, passive data–gyroscope, ambient light, accelerometer, GPS, weather, power use–to infer the current emotional state of users. The system didn’t read your texts or listen to your calls^[7], and it didn’t require any intrusive tests or games. You just keep your phone with you, and Emozia would pick up trends throughout the day, for example, mapping emotional states as you walk across your city.

I was intrigued and agreed to join their board as the Chief Science Advisor, but I had one “ask”^[8]: Let me use your data and technology to help people with bipolar disorder. It is impossible to drag everyone into a lab every week for a battery of exploratory tests, probing for changes in sleep and cognitive function. So, let's find out what we can do with just their phone. It was such an amazing coincidence–one night someone places a profound and engaging problem in my mind, and the next day the seemingly perfect platform to explore the problem falls right into my lap. It would have been eerily perfect if only there was someone crazy enough to share their personal data with me^[9]. That person walked into my life the very same day.

Out of respect for their privacy, I'll skip sharing any further personal details, but when they heard about my desire to leverage Emozia for bipolar, they immediately volunteered^[10]. We tracked them continuously through multiple manic episodes. Our volunteer was one of the many for whom drugs were ineffective. Upon recognizing that they were in the midst of an episode (typically after multiple sleepless nights), they would simply lock themselves in their bedroom under the care of loved ones. That was life for someone with bipolar.

When we began analyzing the data, there were no subtle clues or weak indicators. Rather, the pattern couldn’t have been stronger if we’d plotted a line graph that spelled out, “It started here, stupid!”^[11]. The indicators of a manic episode were visible across the phone sensors, but were strongest in the motion data. In the weeks leading up to subjective onset (when our volunteer first recognized they were in a manic episode), the way they moved through space changed dramatically at all scales, from ballistic movements in personal space (gyroscope and accelerometer) to travel patterns across their city (GPS). Once the system had learned baseline movement patterns, changes in those patterns produced clear signals of oncoming manic episodes weeks in advance of subjective report. Integrating in weaker signals from the other sensors (ambient noise was a clear correlate, for example), our daily machine learning analysis could provide 2-4 weeks of buffer time. Individual signal samples from the Emozia platform may have been very messy and would never have satisfied clinical lab standards, but when integrated together continuously over time, a rich story emerged. And it’s a story we can read at a massive scale across thousands and millions of people.

There are so many huge unanswered questions here:

• Can knowing a manic episode is imminent actually make a difference?
• Does weeks or even a month of warning time provide any benefit?
• How can we intervene in ways that respect the uncertainty of the model?
• Might users reduce medication and its notoriously hated side-effects most of the time, only upping dosage during high risk periods?
• Can a check-in with a doctor, time off from work, or identifying stressors and triggers actually help reduce acute incidents, jobs loss, divorce, institutionalization, or suicide?

These were open questions that might now be answered, but they weren’t the most immediate dilemma. During the time of this side project, Emozia was piloting a trial of its market analysis tool. Analyzing the resulting anonymized data from thousands of people, we saw dozens of individuals who showed the same change in movement patterns as our volunteers. We could see people who might well be in the midst of manic episodes, some surely undiagnosed, but our tool was not approved for any medical use and our contracts barred sharing any personal identifiers. I felt like Jimmy Stewart in Rear Window, except that everyone believed me, but we’d all chosen to do nothing.

What do you do when you see people might need help, but you’ve promised to do nothing?

Within just a few months of launching our bipolar side project at Emozia, our pilot app was detecting the evidence of impending manic episodes. It offered hope that technology like ours might someday lessen the burdens of bipolar sufferers. Unexpectedly, however, we were also seeing evidence of manic-like behavior in anonymized data from non-volunteers. My first reaction, as a scientist, was excitement at seeing our models generalizing robustly to so many people. They might not have bipolar, but they were clearly experiencing a deep disruption of their normal life patterns. Almost immediately though, I felt helpless. We could see people who might be struggling but had no way to reach them.

If these patterns were visible to our little startup, they were certainly visible to Apple, Google, Samsung, Huawei, Verizon, BT, and others. I suspect that if these companies advertised the depth of mobility and metadata insights they collect about us, we would all immediately demand it be private and inviolate. But the data offers more than a brisk business in ad targeting and political influence. As I had seen for myself, it also revealed a real-time window into our health and wellbeing.

What should these companies do if they can see that someone is in the midst of a manic episode right now? If someone has never been diagnosed with bipolar or major depression, autism or Alzheimer's, but AI combined with ubiquitous computing allows early detection and monitoring, what do we do with that information? Laws and liability aside, what is our responsibility to each other? Doing nothing might be the correct choice when balancing out the cost of wrongly intruding on people's lives, but it is a choice we should be making together.

For me, it is not an abstract choice. In 2010, as I lay on a cot next to Thumper’s bed in the PICU following his collapse and subsequent diagnosis with type-1 diabetes, my fogged mind dwelt on the months that had led up to this moment. Beyond the fear and confusion of any parent, I experienced a more rarified form of self-castigation—wasn’t this preventable? One particular symptom repeatedly surfaced, at the pediatricians’ office, during the ER admissions interview, and with our new endocrinologist: bedwetting.

If you recall from “Jitterbug”, as glucose builds up in the bloodstream, the body tries to get rid of it in any way it can. Excessive urination is one of the classic symptoms of all types of diabetes^[12], as it’s one of the most effective ways for the body to clear the harmful excess. Sufferers drink and drink and then whizz away a frappuccino’s worth of sugar.

Do you also recall my mentioning that Thumper had likely been suffering from type-1 diabetes for months? The onset is slow as the immune system progressively wears down the pancreas’s production of insulin-secreting beta cells. During this period the symptoms are subtle, but build momentum as cells die and sugar levels rise. Long before the vomiting, delirium, and acidic blood, there is thirst and the need to pee. For little kids, like a 3-year-old Thumper, it means bedwetting.

I can recall so clearly my frustration and even anger during the Summer of 2010. My wife and I had potty trained Thumper from a few months old, and by 3, he was a potty champion^[13]. Then, seemingly out of nowhere, came a nightly torrent of bedwetting. I regularly change his sheets 3 times a night, all while he’s guzzling water right before bed. We tried special bedding and underwear. I tried understanding and love. I tried shouting and fury. My child’s own body was destroying itself, and I shouted at him for doing the one thing that kept him alive. None of it worked, of course, the anger least of all, but I still hate myself 10 years later.

There is one thing I tried that should have worked: I googled “bedwetting”. These searches led me to articles about infections, ads for urine triggered alarms, and training techniques for kids. Obviously, none of this helped. In retrospect, I surely must have read articles about type-1 diabetes, but I cannot recall a single one. But when I say that my searches should have worked, I’m not referring to the articles that Google returned; I’m talking about the search itself, conducted again and again and again. Night after night. Google knew Thumper had diabetes.

Let’s be super clear, no one at Google actually “knew” anything of the sort. I’m not suggesting the company maliciously withheld knowledge of my son’s illness. Google doesn’t have a secret database filled with all of our undiagnosed illnesses^[14]. But I am saying something almost as profound—Google, Facebook, and other tech companies could deploy algorithms to help reveal these hidden diseases, but choose not to.

And they are right… or at least their choices are less wrong than other options.

Imagine Google holds the following hypothesis: This user’s child has type-1 diabetes. Given a boringly meandering search history through ‘80s pop stars, exotic locales, sci-fi book reviews, and Python functions, the likelihood my child has diabetes is very low. Fewer than 20,000 kids were diagnosed in 2010. So, the background rate is <1%.

When I ran my first search on “bedwetting”, though, that probability changed. It’s no longer the probability that a family selected at random has a diabetic child, it’s the probability that a family searching about bedwetting has a diabetic child. That first search doesn’t change the likelihood too much, but then I conduct a second and third. I return weekly running searches on bedwetting in kids. Now it is the probability that our family has a diabetic child given 6 weeks of near nightly searches. That number is much, much higher than the background rate. Imagine that the likelihood is now 10%. That means the likelihood that Thumper has diabetes is 10 to 100 times higher than the background rate. It would be morally wrong to withhold that information.

But it would also be wrong to share it. While the probability that our family has diabetes has increased a hundredfold given our search history, there is still a 90% probability that we don’t. If Google put a big banner warning at the top of my search results, “See a doctor immediately! Your child might have diabetes,” they would be wrong 9 times out of 10. I wouldn’t be surprised if 2 or more of those families sued Google for the false alarms. Multiple lawsuits for each family helped is not a wonderful incentive for Google or anyone else to take the risk. So given the choice between damned if you do or they never know if you don’t, they choose don’t.

What a lazy choice. I’m profoundly dissatisfied with the assumption that there are no options between intrusion or nothing, blaring alarms or willful ignorance. We too often envision these choices as saving all the little babies versus jackbooted thugs kicking in doors^[15]. If we can see clear evidence that someone is in a major depressive episode with a dramatically elevated risk of suicide, we surely have other options than sending terrifying, unsolicited message to their loved ones. I’ve had many conversations with executives involved in those decisions and a frequent position, at least in public, is “who are we to decide for our users.” But making no choice is still a choice with moral consequences^[16]. It turns out the reason I don’t want Big Brother making these decisions for me isn’t because they are too controlling, but because they choose to do nothing. And so we slouch on with diagnosis by tragedy.

Let’s take this decision back and decide collectively how knowable risks about diabetes or bipolar might be used in a way that prevents emergency room visits or lost jobs without triggering false alarms (or abuse). One alternative I’ve explored in my work is designing “curious” AIs who actively seek more information before taking action. This is a bit like a scientist testing a hypothesis. The bipolar alert system we design had volunteers identify one or more trusted confidants—spouse, doctor, children—who could be notified in case of an elevated manic risk prediction. Rather than, “You might be having a manic episode,” the system instead sends a message asking diagnostic questions of the confidant or prompts for self-reflection by the user. The system can scale the urgency and explicitness of its messages in scale with its idiosyncratic confidence in this specific user’s risk.

In the case of my Google searches for “diabetes”, scaling the degree of intervention with AI confidence can mean changing the search results to reflect increasing risk with repeat searches. As the odds of diabetes increases with the number of bedwetting searches, so might the prevalence of targeted results in the returned links. What if articles about diabetes had been injected into my news feed in the months leading up to those terrible days in the hospital? The AI’s policy (the actions it takes) can be proportional to the expected gain (the trade off between benefits and costs). Reserve the strongest, most intrusive interventions for the most dire situations. Even then, we will make mistakes, but at least we will be honestly and explicitly balancing the cost of false alarms against the tragedy of preventable mortality.

Trying to give a window of action to those with bipolar has been another personal project that I piloted and just gave away. Much like diabetes and autism, I’ve simply tried to show that the predictions are possible. As I’ve shared the results of my collaboration with Emozia with others, I’ve been thrilled to learn that this field of research has made many advances that are being translated into potential tools for sufferers today. I still hear from people struggling with bipolar and looking for help. At Socos Labs, we’re considering relaunching this project with updated models and tools. It’s time to move beyond another pilot with volunteered data. The next step is to change people’s lives.