Using Neural Networks to Address Mental Health Crises

Columbia University grad student collaborates w/ NYU Center for Data Science’s Sam Bowman + Koko to invent an online crisis detector

Although the happy faces of friends and family flash up on our social media feeds every day, remember that behind every sepia-toned smile there may lurk a pool of anxiety, fear, or pain.

Despite what our Instagram feeds suggest, our way of being in the world is not filtered by this color or that color, but by our multiple joys and burdens.

Maybe we’re insecure about our weight. Maybe we have been abused in the past, or are being abused right now. Maybe we no longer know how to feel, or why we want to continue living.

During times like these, we need support — and web applications like Koko, an anonymous emotional support network with a messaging service, are already doing their best to comfort those in distress.

But how can we use these online platforms to quickly identify someone who requires immediate support, like emergency services or hospitalization?

Turning to neural networks might be the answer.

Columbia University graduate student Rohan Kshirsagar collaborated with NYU Center for Data Science professor Sam Bowman and Koko’s co-founder Robert Morris to invent a powerful online crisis detector that outperforms existing solutions.

A typical crisis detector simply reads through the particular problems that users express online, and then determines what support is best for each case.

Many existing detectors, however, suffer from suboptimal accuracy rates primarily because they have been trained on restrictive datasets. For example, several detectors are trained on Twitter datasets, but only on tweets referring to a specific issue like suicide, thereby leaving out other crucial crises like abuse or sexual violence.

Some researchers have even had to filter out millions of tweets down into a tiny set of 2000 to cut costs and increase efficiency.

This new crisis detector, however, is outstanding for two reasons.

  1. Their detector was trained on Koko’s large dataset, which contains a diverse set of mental health topics, and requires no filtering as it is already strictly an emotional support network to begin with.
  2. They use an attentive recurrent neural network and a novel explanation algorithm to both detect and explain crisis.

The detector is split into two components.

The first component, which determines whether or not to flag a post, reads all of the posted content. “It also seeds the explanation generation with words from the input that helped to support its prediction,” Kshirsagar adds.

The second part, however, is responsible for rationalizing the crisis prediction.

“It uses a novel algorithm that takes the seeds as input,” explains Kshirsagar, “and generates a coherent concise explanation that rationalizes the neural prediction.”

We often blame social networks for increasing our mental health problems. But, with the right changes, they could also become the answer to those problems. After all, their detector’s accuracy and efficiency could make the difference between life and death.

By Cherrie Kwok