Overview: Using graphs to improve accessibility for mental healthcare

Try our live demo!

We built Mental Health Hero as a graph-based approach to solve one of the largest public health epidemics we're facing today: the prevalence of untreated mental health issues such as anxiety and depression. Only 40% of adults that suffer from mental illness ever seek treatment for it, and our society is suffering as a result [1].

We spoke with with 10 psychologists to prepare for this hackathon and learned that the underlying issue in increasing access to the appropriate mental health treatment is about getting a holistic view of a (potential) patient from a variety of different touchpoints.

Patient 360

Our solution: We used Tigergraph to build a "Patient 360" view of a potential mental healthcare patient using multiple modalities like social media sentiment, therapy session adherence, and written entries. Then, we leveraged similarity algorithms to recommend them the appropriate care based on the outcomes of other patients that had similar behaviors.



Submission Criteria: IMPACT

Reach (# of people helped) Mental Health Hero has the potential to help hundreds of millions of people around the world. The CDC estimates that 400 million people globally are not receiving essential treatment for their mental health disorders [0]. Added to that, only 40% of adults with mental illness are estimated to receive treatment for it, and the delay between symptom onset and treatment is a staggering 11 YEARS; among 10-to-34 year olds in the United States, suicide is the second leading cause of death [1].

Value (social/economic benefit) From an economic perspective, the cumulative output loss associated with mental disorders is over $16 TRILLION worldwide [2], which is higher than that of cancer and diabetes. From a societal perspective, people suffering from depression have trouble holding steady employment and have increased difficulty maintaining their social relationships.

Depth (addressing the root issues) The root issue we found when talking to psychologists is that we don’t have a way to create a holistic view of patients across their entire care journey. This makes it incredibly difficult to identify the patients that need the most help, and even harder to recommend the appropriate treatment plan (e.g. medication, behavioral counseling).

Mental Health Hero tackles this ROOT ISSUE by using graph databases to create a HOLISTIC view of someone from multiple modalities like social media sentiment, therapy session adherence, and written entries. By doing this, we can identify who needs help and what their ideal treatment could be by comparing them with similar patients.

Solving the similarity problem would help MULTIPLE stakeholders in this space: therapists can adjust their care for their patients, insurance companies can see who needs preventative help from their customer base, and governments can create public health programs to categorize their patient population and provide them the appropriate care.



Submission Criteria: INNOVATIVENESS

Novelty (new way to frame complex problems with graph) Our solution recognizes that graph-based approaches from other use cases such as fraud management and product recommendations, can be applied to this mental health space in a completely unique way.

Similarity Scoring

The real problem we’re solving is: how can we categorize potential patients using a holistic view of their BEHAVIORS in addition to traditional approaches like user-provided data through psychological questionnaires like the PHQ?

We framed this complex categorization problem using two graph approaches:

  1. A person has multiple incoming streams of data such as social media posts, therapy visits, diagnoses and treatments. Graph databases are perfectly made to capture this “Customer 360” view of a patient so we can use all those factors to determine their best course of treatment.

  2. We also saw the similarity algorithms of graphs as the ideal way to take that “customer 360” view from above and use that to “embed” a patient in a way which makes it easier to compare them to others. By doing this, we can recommend treatments to patients the same way companies use graphs to recommend products to similar customers.

Creativity (resourcefulness in overcoming challenges) We had to solve two major challenges while building Mental Health Hero:

  • Getting the EXACT data we wanted was impossible: mental health data isn’t easy to come by, let alone the variety of sources we’d need to create a holistic view of one person.
    • Our solution: We took existing disparate datasets for Twitter sentiment analysis, depression scale scores, and we created a semi-synthesized dataset using the little data we could find as inspiration to show what's possible with our application. We expect healthcare systems to have the actual data available to deploy Mental Health Hero at scale.
  • There wasn't an "out of the box" graph similarity algorithm we could use to completely solve our use case based on the approach we wanted to take.
    • Our solution: After exploring multiple approaches to the problem like graph embedding and Jaccard similarity measures, we settled on building a tailored version of the Jaccard similarity measure for our use case that accounted for the multiple vertext types and relationships we needed to traverse for our specific approach, and the fact that multiple vertices coming from one patient could end at the same target vertex (e.g. a patient can have 3 tweets with a specific sentiment, and we'd want to capture the fact that they had this sentiment 3 times vs. just once).



Submission Criteria: AMBITIOUSNESS

Schema and functional scope Given our direction of a “Customer 360” approach angle to the mental healthcare space, we had a complex schema with 25K nodes across 12 vertex types, and 96K+ edges.

We designed Mental Health Hero to be flexible enough to add new data to the holistic patient view based on recent mental health research: things like accelerometer data and prescription information through EHRs.

Functionally, we used multiple features of Tigergraph:

  • We took the Customer 360 model to create a holistic view of a patient
  • We tried numerous similarity measures to help find related patients, including Fast RP embedding. We ended up using a modified version of the Jaccard similarity measure
  • For all the above, we heavily dug into GSQL and we’ve become quite proficient over the course of the hackathon!

We also used the REST++ interface of Tigergraph Cloud to build the frontend demo of our underlying graph architectur, which you can try out here!



Submission Criteria: APPLICABILITY

Adoption (ease of putting solution into real-world use) We designed our solution in a way that it can be deployed with minimal effort by healthcare organizations like hospitals and public health centers.

The complexity in our solution is about getting the right data from patients; we don’t have that data ourselves and had to make a synthetic dataset, but insurance companies and therapists have access to this data – of course they’d have to get patient approval and comply with HIPAA protocols, which is a challenge that we’ve seen other digital health solutions overcome.

Breadth (number and size of industries that could adopt this) Our approach to taking a holistic view of a person to recommend appropriate approaches is broadly applicable to dozens of industries and domains through the Customer 360 approach.We’ve seen tech and product industries adopt these kinds of approach quickly, and they are the inspiration for us adapting this to the mental health space.

In terms of market size, the global mental health market is estimate to grow to over $500 BILLION within a few years, and our solution could be adopted by multiple stakeholders in that market: psychologists, digital health companies, health insurers, and ultimately governments who want to use our solution to tackle this on a public health level. Given the $16+ trillion in lost economic output due to not providing the right care to patients [4], we believe the impact our solution can have is unbounded!




What's next for Mental Health Hero

Through working on this hackathon and speaking with professionals in the mental health space, we've been very inspired to take this project into the real world and find a way to get it into a production use case.

What we're planning on building next:

  • We want to improve the front-end user experience for therapists and public health professionals that could use Mental Health Hero in the work.
  • We just scratched the surface of how we want to tackle this problem with graph-based approaches. We're eager to use Tigergraph's ML workbench to experiment with Graph Neural Networks to better embed patients to find them the relevant care.
  • We're reaching out to healthcare facilities like hospitals to try running a real-world pilot with actual patient data and structure an experiment to measure the efficacy of our "recommendation-based approach"

We also want to give a HUGE SHOUTOUT to Jon Herke, Dan Barcus, and Parker Erickson from the Tigergraph team for tirelessly helping us learn how graph technology works and helping us debug our issues as we were building out our solution! 🙏🏽

References [0] https://futuresrecoveryhealthcare.com/blog/barriers-to-mental-health-treatment/

[1] https://towardsdatascience.com/data-science-in-mental-health-ccd09ba2148a

[2] https://www.ncbi.nlm.nih.gov/pmc/articles/PMC5007565

+ 3 more
Share this project:

Updates